Google AI Overviews: Tracker, Timeline and Statistics

Mehrzad Khosravi and Hema Yoganarasimhan of the University of Washington have published a revised version of "Impact of AI Search Summaries on Website Traffic: Evidence from Google AI Overviews and Wikipedia".

The design is the interesting part. AI Overviews became the default US search experience in May 2024 but was not default in Europe until early 2025.

So the authors treat English Wikipedia as exposed and German and French Wikipedia as controls. They compare search referrals to the same articles across editions.

They use Wikimedia's public clickstream data, specifically the other-search measure of monthly human referrals from external search engines. The window runs five pre-treatment months from December 2023 and eight post-treatment months from May to December 2024.

The samples are large: 499,927 matched articles in the English-German comparison and 530,873 in the English-French one.

English Wikipedia's monthly search referrals fell 5.45% relative to German and 4.82% relative to French. Both are significant at p<0.001.

The authors translate that into roughly 100.27 million fewer search-originated visits per month. That is about 1.20 billion a year if the monthly effect held.

Read the estimate for what it is. It is not the same quantity as the click-loss figures elsewhere on this page.

It measures the effect of AI Overviews becoming the default in a country, averaged across every article and every query. It does not measure the effect of an overview appearing on a particular search.

The authors are explicit: "We interpret these estimates as reduced-form effects of moving the U.S. search environment into a default-AIO regime, rather than as effects conditional on an individual AIO display."

They also name the reasons the number is diluted. Only part of English Wikipedia's traffic is American, and the clickstream data mix Google with other search engines.

And Google showed an overview on only some queries, while changing its triggering rules throughout.

The authors compare their result with Agarwal and Sen's randomized experiment themselves. They call their own effects "considerably smaller" and put part of the gap down to the publisher.

Wikipedia is not a generic publisher. It sits at the top of results for exactly the informational queries an overview answers.

Read the version history before quoting any number from this paper.

The first version was posted in February 2026. It compared English against Hindi, Indonesian, Japanese and Portuguese editions across 161,382 article-language pairs.

That version reported a decline of about 15% in daily traffic.

The current version changed the control group, the unit and the sample. The headline number came down to about 5%.

That 15% is what search engines and summaries still return when asked about this paper. If you meet it in a deck or a roundup, it is from a version the authors have replaced.

Two limits belong with the numbers. The measurement window closed in December 2024, so this is an early-AI-Overviews estimate rather than a current one.

The outcome counts only visits arriving from external search. The authors call that a first-order effect.

They note that internal article-to-article navigation is roughly 40% of Wikipedia's traffic. Losing an entry visit may cost the browsing that would have followed.

The clickstream data cannot measure that, so the total effect could be larger.

We have kept this out of the statistics table deliberately. Every figure there is one a site owner can hold against their own numbers.

This one is an estimate for a single publisher, under a design that measures country-level availability. A row would invite exactly the comparison the authors warn against.

For publishers who do sell advertising, the paper offers an illustration rather than a finding.

It applies page-revenue benchmarks of $8.99 to $30.81 per thousand views to that traffic loss. That gives roughly $0.901M to $3.090M a month.

Wikipedia itself carries no advertising. So that figure describes what a comparable ad-supported publisher would forgo.

The consolidated hearing set on 29th Jul was held on 25th Aug before Judge Amit P. Mehta. It covered Google's motions to dismiss both Chegg and Penske Media.

Google's position at argument was the one its briefs have carried since January. AI Overviews is a "product improvement" rather than a separate product, and publishers who dislike it can decline to be indexed.

Publishers have long allowed crawling in exchange for search traffic. Its attorneys described that arrangement as a "vague historical course of dealing".

By the accounts available, the court pushed back on that framing.

Mehta said the situation "all seems really unfair", and noted that publishers have no control over how Google uses their content. He described the claimed product improvement as being built "on the backs of the publishers".

He also said that product improvements "are not immune from antitrust scrutiny", which is the legal point the rest of the case turns on.

The market question drew the other exchange of note. Mehta observed that a market is plainly forming for generative AI inputs.

Penske argues that OpenAI and Perplexity pay for content. Google, it says, uses its search monopoly to get the same inputs for nothing, raising rivals' costs.

Be careful about how much this settles. A judge testing an argument at a hearing is not a judge deciding one.

Mehta has not ruled on either motion. The same judge dismissed the Helena World Chronicle case in full in March.

His reasoning there was that AI Overviews is an integrated feature of Search rather than a product tied to it. Google is asking him to apply that reasoning again.

State the sourcing plainly, because we did not read a transcript and no transcript is public.

Jason Kint, chief executive of the publisher trade association Digital Content Next, attended and posted a thread from the room. Search Engine Journal quoted it at length on 26th Aug.

MLex, reporting independently, headlined its own account with the same "backs of publishers" line.

The dockets caught up on 27th Aug. The court's own record now confirms both the hearing and its outcome.

Chegg's docket carries the minute entry: "Motion Hearing held on 8/25/2026 re 19 MOTION to Dismiss the Amended Complaint filed by GOOGLE LLC, ALPHABET, INC. Matter taken under advisement."

"Taken under advisement" is the formal version of what this entry already said. The judge heard argument and will rule later, in writing.

That written opinion is the next docket event that matters on either case.

Google announced on 20th Aug a "new interactive "Preferred Sources" button that publishers can embed on their pages".

Mrinalini Loew, General Manager of the Google Search Ecosystem, wrote that clicking it "adds the site as a Preferred Source on Google and immediately directs them back to where they left off on the publisher's page", so readers "more easily find their favorite publications in Top Stories, AI Overviews, and AI Mode".

Google's count of sources readers have picked is now "more than 600,000 unique sources". That is up from the 345,000+ it gave in May and 200,000+ in April.

This counts for more than the average personalization announcement because of what the control already does. Since May a reader's picks have been badged inside AI Overviews and AI Mode answers.

Google's stated effect at the time was that "people are twice as likely to click through to a Preferred Source".

Until now the only way to gain one was for a reader to act unprompted. They had to find the star icon in Top Stories or the settings page at google.com/preferences/source.

The button moves the ask onto the publisher's own page, at the moment someone is already reading.

The mechanics are in Search Central documentation, which Google's own documentation-updates log dates to the same day.

The standard set-up is two lines. The first is a script tag loading https://news.google.com/swg/js/v1/publisher.js, preferably in the head.

The second is <div google-add-preferred-source-btn></div>, placed wherever the button should sit.

It renders a localized, Google-styled button, and data-theme takes light or dark while data-lang overrides the reader's browser language.

There is an advanced path for binding the flow to your own UI. It uses an ES module import of publisher.mjs, or a callback queue with preferred-sources-control="manual".

A deeplink fallback exists for sites that cannot run the script: https://www.google.com/preferences/source?q=example.com. It works in a newsletter or a social post as well as on a page.

Google also publishes translated button graphics.

Two eligibility details decide whether this is available to you at all. Only domain-level and subdomain-level sites can appear in the source preferences tool.

So www.example.com and code.example.com qualify and www.example.com/blog does not. That rules out anyone whose publication lives in a subdirectory of a larger site.

And Google tells you to check that your site already appears in the tool before implementing anything. The documentation is also explicit that the button is not a ranking mechanism in itself: "It's not required to do them in order to appear as a preferred source."

Be clear about the limits before building a campaign around this. A preferred source is one reader's setting, applied to that reader's results.

So it accumulates one person at a time, rather than moving a site's standing in general.

Google's documentation describes the badge in Top Stories. For AI Overviews and AI Mode it says content "can be highlighted" for users who selected it.

That is weaker than a guarantee of appearance.

Nothing in the documentation describes any reporting. There is no stated way to see how many readers have added you.

So the only measurable signal is whatever your own analytics show after you ship the button.

Availability is broad. Top Stories gets it globally in every language.

AI Mode and AI Overviews get it in every language and locale where they run.

Google published five education features on 19th Aug, written by Awaneesh Verma, a Distinguished Product Manager on Search.

Most of the coverage will read it as a study-tools story. The availability sentences inside it change what an AI Overview can be.

On interactive visuals, Google's own example is an AI Overview: search "ph scale" and "quickly get an interactive visual in an AI Overview".

The rollout sentence is explicit that this is not only an AI Mode feature: "Generative UI capabilities have launched globally in English in AI Mode and have started rolling out as well in AI Overviews."

Generative UI means Search building a custom tool, chart or simulation for the question instead of describing it. It had been an AI Mode feature since Gemini 3 in November 2025.

Two more features are already live in AI Overviews or promised to them. Practice quizzes are "now globally available in English free of charge, in both AI Overviews and AI Mode", covering any subject plus the ACT, AP, ENEM, GRE, JEE, LSAT, MCAT, NEET and SAT.

Custom file creation builds a document, slide deck or spreadsheet from your uploaded material. It has shipped in AI Mode and "will come soon as well to AI Overviews".

And from the coming weeks, photographing your working with Lens returns an AI Overview that "provides helpful explanations, pinpoints where you may have made mistakes, and coaches you when you're stumped", globally in English.

The publisher question these raise is different from the one this page usually tracks. Every click study here measures what happens when a summary of cited pages sits above the results.

A generated simulation is not a summary of anything. There is no set of pages it stands in for, and nothing to cite.

Google says the quizzes carry "explanations for each response and links to learn more", so that format keeps links. It says nothing either way about the interactive visuals.

Google says the test-prep content is licensed from named partners: The Princeton Review, Careers360, PhysicsWallah and Akira Enem. That is authority bought rather than crawled.

No amount of optimizing gets a publisher into that slot.

Nothing here changes a number in the statistics section. It is a shift in what the container holds.

The panels that measure prevalence and click loss will take months to answer one question. Does a generated answer behave differently from a summarized one?

The AI Mode hub carries the rest of the announcement, including the notebooks integration and the availability table rows.

Five researchers posted a preregistered field experiment to arXiv on 18th Aug. Stephanie T. Wang is at the University of Pennsylvania.

Jeffrey Gleason, Yakov Bart and Christo Wilson are at Northeastern.

Danaé Metaxa is the fifth author.

It is the second randomized test of AI Overviews to reach us, after Agarwal and Sen's in July. It is the first to test AI Overviews and AI Mode in the same design.

US Google users were recruited in mid-March 2026 and screened for Chrome and Google as their primaries. They were given a browser extension.

After three days on unmodified Google, each was randomly assigned to one of three arms for seven days.

The arms were AI Overviews hidden and AI Mode redirected away, unmodified Google, or every query forced into AI Mode.

Of 1,444 enrolled, 1,100 searched at least once and 956 finished the exit survey. In the baseline period, 36% of their searches produced an AI Overview.

Exposure to the no-AI condition raised click-through by 8.8 percentage points. The 95% CI runs 2.3 to 15.3, at p=0.008.

It also cut time spent per session by 0.59 minutes.

The fraction of users clicking through to news sites rose. The authors call that suggestive rather than settled, because it does not survive their alternative measures.

On Reddit, Wikipedia and ad clicks they found nothing. They warn they were not powered to detect small effects there.

Do not read the 8.8 points as a second opinion on Agarwal and Sen's 39.8%. The two measure different quantities.

This study reports a click-through rate in percentage points. Theirs reports outbound clicks per search falling from 0.62 to 0.37 when an overview appears.

Different metric, different sample, different months. What the two agree on is the sign and the mechanism, which is the part a publisher can plan around.

Know one flaw in this arm before quoting it. Compliance collapsed during the study.

On the first day 90% of AI Overviews were successfully hidden. Then the rate fell to zero.

Google had changed the HTML of AI Overviews mid-experiment and broken the extension's logic.

Overall only 51.1% were hidden, so the paper reports local average treatment effects. It says plainly that this rests on assuming the assignment affected behaviour only through overview exposure.

If Google's interface change did anything else to those users, that assumption weakens.

The finding with no counterpart in any panel study is the null. Removing AI Overviews produced no detectable change in trust, usefulness, satisfaction, agency or perceived relevance.

The authors read that as the benefits of AI Overviews not showing up in the measures users report.

The AI Mode arm is the opposite story, and a much larger effect in both directions. It is covered on the AI Mode hub.

There, a forced AI Mode cut click-through 18.8 points, and cut trust and satisfaction with it.

Google gives site owners no clean signal for AI Overviews traffic, so Alex Galinos built one. He ran it for nine months on a transportation-industry brand he manages, publishing the results on Search Engine Land.

The method is cheap enough for anyone to copy. When a user clicks a cited link in an AI Overview, Google sometimes appends a #:~:text= fragment to the URL.

A GA4 custom dimension can fire on any session that arrives carrying one. From September 2025 to June 2026 that captured 51,200 events across 1,661 cited snippets.

The reporting finding is the one to act on. Across more than 50,000 events, 22.4% of AI Overview traffic was filed as Direct, not Organic Search.

That is 11,468 sessions filed to the wrong place.

AI Overviews exist only inside Google Search, so all of it should have been organic. The monthly rate ranged from 16.8% in April 2026 to 29.3% in May 2026.

If your site earns meaningful AI Overview traffic, your organic channel is understated by some version of that gap. Nothing in the interface tells you so.

The share numbers from the same dataset swing enough to be a warning against planning on them. AI Overviews accounted for 7.53% of the site's organic sessions over the whole period.

The share peaked at 16% to 17% in February and March 2026. It had fallen to roughly 2% to 4% by the time of writing.

Citation volume was also heavily concentrated. The single best-performing snippet drove 2,276 events, against an average of 31 across all 1,661.

Two caveats are the author's own and both matter. The #:~:text= fragment is not exclusive to AI Overviews, since featured snippets and People Also Ask use it too.

He checked the site's featured-snippet exposure in Ahrefs at fewer than 30. That argues overviews dominate the sample rather than proving it.

And the GA4 dimension is event-scoped, so the 7.53% share compares events against sessions. He describes that as directional rather than exact.

The limit we would add is the sample. This is one brand in one industry with one analytics set-up.

So the 22.4% is a measured fact about that site, not a rate to apply to yours.

What generalizes is the mechanism. A Google-appended text fragment can be surfaced in your own analytics today.

And channel attribution for AI Overview clicks is not something to assume is correct.

That sits alongside what Search Console does and does not tell you about AI impressions. It also sits alongside the logging error still running in its AI reports.

Site owners watching their AI numbers in Search Console saw them fall off a cliff in mid-August. Google has confirmed the cause, and it is the reporting rather than the surface.

Its data anomalies page carries the note: "A logging error caused a decrease in impressions on the Generative AI performance report in Search for data from August 13 - August 17, 2026. This issue affects data logging only."

That note first described the error as "starting on August 13, 2026" and "ongoing".

Google has since closed the window at five days and dropped the word. There was no announcement and no revision date on the page.

John Mueller added on Bluesky that Google is "working on resolving it". It is "just a logging issue and not representative of visibility changes in Search", and an annotation will follow in Search Console "in the near future".

The practical consequence is simple. A dip in that report dated between the 13th and the 17th is not lost AI visibility.

Any week-on-week or month-on-month comparison spanning those five days is measuring the bug.

Google has not said whether the missing impressions will be backfilled.

Keep the scope narrow. The anomaly note names the Generative AI report in Search only.

Site owners also reported a clicks decline in the ordinary Search performance report from around 12th Aug. Search Engine Roundtable collected those reports, and Google has confirmed nothing about them.

This is the third anomaly note this year to touch generative-AI reporting. The 24th Jun error took a day of Discover data and of generative AI in Discover with it.

A fourth Discover error did the same on 13th Aug, the day this one began.

All three land on the only first-party AI number a publisher gets. That is an impression count, credited only once a link scrolls into view.

A dating note: Google's help pages carry no revision dates. So the closed window above is dated to our reading of the page on 24th Aug 2026.

The illustrated overviews a recipe publisher objected to last week were a test, and the test has stopped.

Robby Stein, VP of Product for Google Search, told Search Engine Roundtable: "This was a small experiment we're no longer running."

Barry Schwartz updated his 14th Aug report on Monday the 17th to record the confirmation. He also re-dated the story.

That answers the scope question this hub left open, and it answers it narrowly. Google has not said how many searches showed generated images, in which markets, or why it stopped.

What it does not retire is the announcement behind it.

Google said in July it was bringing image generation into AI Overviews. Nothing in Stein's sentence withdraws that plan.

This rollout is off. The feature is not disclaimed.

The publisher's argument outlives the test, which is the reason to keep both entries on the record. A generated picture inside a block that names sources can be read as a cited site's work.

That is the question to put to the feature if it returns.

Google said in July it was bringing image generation into AI Overviews, rolling out over the coming weeks. The first reported sighting of generated images inside a live overview came on 13th Aug.

It came from a publisher who found one illustrating their own recipe.

Inspired Taste, a recipe site running since 2010, posted screenshots of an AI Overview for their own brand. It used what they call "AI step cartoons".

Their article and video already demonstrate the steps the cartoons depict. That includes how to cut the squash safely, which they note can be dangerous.

Their objection has two parts, and the second is the one with consequences beyond recipes. The first is substitution: the overview "is directly trying to replace creators who buy groceries for testing, photography, and filming recipe videos".

The second is attribution. "There is also risk that a searcher might think that one of the websites cited in the AIO did the illustrations."

A generated image sits inside a block that names publishers as sources. Nothing marks the picture as Google's rather than theirs.

That is a new way for a cited site to be tied to work it did not produce.

They also put the quality argument back on Google: "Google would deindex any publisher for producing sloppy AI content like these AIOs at extreme scale", so giving such content "position zero above human creators doing the hard work is shameful".

The posts were addressed publicly to two Google Search executives. They credit the recipe changes Google shipped earlier in the year, before asking it not to continue in this direction.

This was one publisher's sighting, reported by Search Engine Roundtable the following day, rather than an announced rollout. Nobody has measured whether an illustrated overview changes clicks.

Google answered on 17th Aug, four days after this sighting, and stopped the test. It called it a small experiment that it is no longer running.

It has still not said how widely the images appeared or in which markets.

What the episode established is narrower than a rollout and more durable than a test. Generated pictures can reach live overviews.

The first publisher to meet one read it as a threat rather than a feature.

On 11th Aug the Alliance de la presse d'information générale filed with the Autorité de la concurrence. That is France's competition regulator.

The Alliance represents close to 300 French daily and regional titles.

Among them are Le Figaro, Le Monde, Ouest-France, Le Parisien, Les Échos and L'Équipe.

It is the first competition action anywhere aimed at how AI Overviews and AI Mode launched in a market.

The complaint is not about the features existing. It is about the sequence.

Google switched both surfaces on in France on 22nd Jul. The Alliance says that happened before any transparent, good-faith negotiation opened.

Its account of what publishers were offered is the substance of the filing. They were offered an update to the existing licence contract.

The only alternative was a technical withdrawal that costs them visibility.

"Ils ont été mis devant le fait accompli," the announcement says, they were presented with a done deal.

The legal frame is French rather than European. France was the first EU country to transpose the neighbouring-rights directive, in the law of 24th Jul 2019.

That law makes use of press content conditional on prior authorisation and payment.

On top of that sits a competition record. The Autorité sanctioned Google in April 2020 and ordered it to negotiate in good faith.

In June 2022 it fined Google 500 million euros for not doing so.

The month after that it made Google's commitments to good-faith negotiation, transparency and non-discrimination binding until 2027.

The Alliance's argument follows from that. Rolling out a new use of press content without prior authorisation or dedicated payment breaches those binding commitments.

That regulator has already ruled once on this exact coupling. In decision 24-D-03 of 15th Mar 2024 it fined Google 250 million euros.

It found that Google had not told press agencies and publishers that Bard was using their content.

The only way to keep content out of the model, it also found, was to block Google's crawler entirely. That block covered Search, Discover and Google News at once.

The Autorité said that obstructed publishers' ability to negotiate payment.

The UK CMA order broke that coupling in Britain two years later. The French filing says the same problem has returned in a new form.

The figure the Alliance leads with is a regulator's, and it needs reading carefully. Arcom, France's audiovisual and digital regulator, puts the traffic loss from AI-generated summaries at 33% to 38%.

That estimate covers European markets where these services are already active.

So it is an estimate of what has happened elsewhere in Europe. It is not a measurement of French traffic, which had three weeks of data at the time.

And it is Arcom's number reaching us through the Alliance rather than a published Arcom study.

Some trade coverage has inverted it into a claim that AI summaries account for 33% to 38% of publishers' losses. That is a different and much weaker statement.

What makes the filing more than a protest is the Meta precedent. On 8th Jul 2026 the Autorité issued four interim orders against Meta.

It was acting on a complaint the Alliance lodged in September 2025.

Its president Benoît Cœuré spoke at the press conference for that decision. In the Alliance's account he said: "les contenus protégés qui seraient utilisés par AI Overviews doivent faire l'objet d'une rémunération au titre du droit voisin".

That means protected content used by AI Overviews should be paid for under neighbouring rights.

So the head of the authority hearing this complaint has already said in public what he thinks the answer is.

The Alliance is explicit that it is not asking for the technology to stop.

Its president Marc Feuillée, who is also chief executive of Le Figaro, put it this way: "Les éditeurs ne demandent pas d'arrêter l'innovation. Ils demandent que cette valeur soit partagée et que l'utilisation de leurs productions soit rémunérée, comme la loi l'exige."

Publishers are not asking for innovation to stop. They are asking for the value to be shared and their work paid for as the law requires.

The filing, the announcement says, is meant to restore the conditions for negotiation rather than replace it. Nothing in the announcement asks for interim measures.

Google answered through the press rather than with a post.

It said it had reworked how it pays the French press. That rework accounts for protected content appearing in AI Overviews and AI Mode.

It told Franceinfo that site owners can choose whether or not to appear in these AI features. That is per AFP's report carried by Euronews.

Those two answers are the shape of the whole dispute. Google says the payments already cover the AI surfaces and an opt-out exists.

The publishers say a payment they did not negotiate and an opt-out that costs visibility are not consent.

For anyone outside France, this is the first test of its kind. Can a neighbouring-rights regime force AI answers to be licensed separately from search?

The Autorité does not announce filings and has published nothing, which is normal. So the next visible step will be either a decision on interim measures or nothing at all for months.

On 5th Aug Google filed a notice of supplemental authority in both US suits over AI Overviews. Those are Chegg, Inc. v. Google LLC and Penske Media Corporation v. Google LLC.

A notice like that points the judge at a decision issued since the briefing closed. The filing itself is not readable in the public record.

What it cites became public on 10th Aug, when the publishers filed a joint response in both cases.

The authority is PhantomALERT Inc. v. Apple Inc., No. 25-7017. The DC Circuit decided it on 24th Jul 2026.

It is not an AI case at all. A traffic-app maker retrofitted its app for Covid-19 outbreaks in 2020, and Apple barred the update from the App Store.

The antitrust suit that followed was dismissed.

The appeals court affirmed, because the complaint did not plausibly allege a relevant product market.

Google's argument, as the response describes it, has two parts. The decision "confirm[s]" that a rule-of-reason tying or reciprocal-dealing claim must allege "substantial anticompetitive effects in the tied-good market".

And the publishers' market definitions resemble the ones PhantomALERT rejected.

The publishers say the notice "mischaracterizes plaintiffs' allegations and PhantomALERT".

Their first answer is scope: that case involved "single-brand aftermarket[s]", which are not at issue here. Beyond that, they argue, it restates general principles they had already addressed in opposing the motions.

The rest of the response is the plainest public account of what these two complaints actually allege.

On reciprocal dealing, both point to subcompetitive prices in three content markets: republishing, generative AI training and retrieval augmented generation. Both also point to the elimination of choice among market participants.

On tying, Penske says it pleaded foreclosure in a tied market for Online Publishing "in the form of zero-click searches, $2 billion in lost revenue, and declining quantity of quality content".

Both plead per se and rule-of-reason theories. The response notes that PhantomALERT itself accepts that per se claims need no proof of effects in the tied market.

On market definition the response turns the opinion around. The PhantomALERT complaint was "bereft of allegations speaking to either a hypothetical monopolist test, the Brown Shoe factors, or the availability of substitutes".

These plaintiffs say they pleaded all three in detail for every market they allege. The opinion also says such allegations "need not be detailed", though the same sentence requires "more than labels and conclusions".

Market definition was the fight to watch going into argument. It is the ground Google won on in March.

The same judge held then that AI Overviews is not a separate product from Search. Argument on both motions was heard on 25th Aug.

One caution on sourcing. Google's notice is not available through the free public docket, which serves written opinions rather than party filings.

So its argument reaches this page only through the publishers' description of it.

The response and the PhantomALERT opinion were both read in full.

The figures everyone quotes from Pew now have a peer-reviewed paper behind them. Those are a result click on 8% of visits with an AI Overview, against 15% without.

The extended abstract went to the IC2S2 computational social science conference. Athena Chapekis, Anna Lieb, Sono Shah and Aaron Smith of Pew Research Center posted it to arXiv on 5th Aug.

Read it as the same study, not a second one. It uses the same 900 US adults on the Ipsos KnowledgePanel Digital panel, monitored from 1st to 31st Mar 2025.

That means the same 91,121 visits to Google Search pages and 68,879 distinct queries. Our existing entry covers all of it.

What is new is the modelling.

Two logistic mixed-effects models were fitted over 91,121 observations from 647 panelists. They test whether the gaps hold once query attributes and each panelist's own habits are controlled for.

Both do, at p < 0.01.

That addresses the standing objection to every click-loss study. The objection is that overviews turn up on exactly the queries nobody was going to click anyway.

Holding those attributes constant does not make the association go away.

The paper also names the attributes that predict an overview. They are longer queries, queries that begin with a question word, and queries containing both a noun and a verb.

Phrase a search as a full-sentence question and it is likelier to be answered on the page.

Two limits, both the authors' own. They state that observational data means "we are unable to make causal claims", which is what separates this from the randomized experiment published in July.

And the 1% source-click figure counts only the first three cited sources, the desktop maximum shown in April 2025. So clicks to a source behind "Show all" are outside it.

Two months after the Generative AI performance reports arrived, the rules behind an AI Overviews impression deserve a close read. They sit in Google's Search Console help pages rather than in any announcement.

The impression rule is stricter than the one for ordinary results. Google's help page says that for a link inside an AI Overview, "to be counted as an impression, the link must be scrolled or expanded into view".

A classic listing only has to be served on the results page. An AI Overviews link has to reach the reader's screen.

That is a real difference in threshold on a surface where the citations sit below and inside a generated answer.

Position works differently too. "An AI Overview occupies a single position in search results," Google says, "and all links in the AI Overview are assigned that same position."

Average position therefore tells a publisher where the overview sat on the page. It says nothing about where their link sat within it.

The ordering signal that SEOs are used to reading does not exist inside an overview.

Clicks are the straightforward one. Clicking a link to an external page in the AI Overview counts as a click, in the main performance report.

What the Generative AI report carries is narrower. It reports impressions for AI Overviews and AI Mode, with no queries, clicks, click-through rate or average position.

It also excludes Search Labs experiments.

When this was written Google was still "rolling out this report to a subset of website owners". The report reached every property worldwide on 31st Aug 2026.

There is an aggregation trap in it as well.

Chart data is aggregated by property, so in Google's own example, "if two results from the same site appeared in a generative AI search results feature, they count as a single impression". The page table aggregates by URL instead.

Page-level impressions will not always add up to the chart total. The difference is the dimension rather than an error.

Read against the statistics on this page, the shape of the problem is unchanged. Every click figure a publisher has comes from a third-party panel.

The only first-party number Google offers is an impression count. Its rules make it neither a visibility rank nor a traffic measure.

Carolyn Shelby, writing for Search Engine Journal, draws the practical line.

Do not pour generative and ordinary organic impressions into one bucket. And do not compute a blended click-through rate from them.

A note on the date: Google's help pages carry no revision date.

So this entry is dated to the analysis that set the rules out, not to a documentation change.

Two weeks after Google confirmed the Top Stories carousel inside AI Overviews was live, NewzDash put numbers on it.

Take trending-news searches that produced a Top Stories carousel at all. In the US the carousel sat inside the AI Overview 15.5% of the time.

Its usual position is further down the page. In the UK the figure was 17.5%, per Press Gazette's report of the data.

The analysis covered the 30 days to 28th Jul. It spanned more than 17 million Top Stories across more than 40 countries.

Only the US and the UK showed the format at all. That is the first evidence of where it actually runs, since Google had confirmed only the US.

Entertainment queries trigger it most: 35.1% in the US and 31.5% in the UK. World news follows at 31.8% in the US against 18.9% in the UK.

Sports comes next, at 16.5% in the US against 21.1% in the UK.

The important part is the mechanic, not the percentages. Press Gazette reports the embedded carousel replaces the classic Top Stories box below the AI Overview.

It does not appear in addition to it. So on those searches the publisher links move up the page rather than multiply.

NewzDash founder and CEO John Shehata calls it "a double-edged sword for publishers". The upside, in his words: "Instead of placing a large AI-generated answer above the news results, Google is putting publisher headlines, images and links directly into that experience."

He adds that there is no comparative clickthrough data yet, so the gain is expected rather than measured.

The other edge is the opt-out. Shehata expects publishers who exclude their content from generative AI features to lose eligibility for the embedded placement.

That exclusion is the Search Console control. Google has not confirmed how it applies to the carousel.

Opted-out sites may still qualify for standalone Top Stories. But he notes there is no guarantee Google shows that module elsewhere on the same page.

Read together with the rest of this page, that is a narrowing choice for a news publisher.

The control that keeps a site out of AI answers now has a second effect. It appears to keep the site out of the most prominent link placement Google has put inside them.

Blocking Google-Extended, Shehata confirms separately, affects neither.

On 29th Jul Judge Amit P. Mehta ordered a single consolidated hearing on Google's pending motions to dismiss.

One case is Penske Media Corporation v. Google LLC (No. 25-cv-3192). The other is Chegg, Inc. v. Google LLC (No. 25-cv-543).

The date is 25th Aug 2026, at 2pm. The venue is Courtroom 10 of the US District Court for the District of Columbia.

The order grants a motion Google filed in December, which the publishers had opposed.

Read against May, the scheduling is the news.

When the plaintiffs asked for argument then, the court refused: "The court is well aware of the pending motions in these cases. A motion to set oral argument is unnecessary. The court will set argument, if needed, when it is prepared to address the motions."

It is now prepared to address them. These are the two leading US suits over AI Overviews and traffic loss.

Chegg sued first, in February 2025. Penske is the first major news publisher to follow, in September 2025.

A hearing is not a ruling, and both cases are still at the pleading stage. Nothing about the merits has been decided.

What is at stake in August is whether either complaint survives at all.

The reason to mark the date: the same judge dismissed a different publisher antitrust case in March. He held that AI Overviews is not a separate product from Search.

He also held that publishers harmed in news cannot sue over the search market.

Google will press that reasoning here. Mehta is also the judge who decided the Justice Department's search monopoly case.

The March opinion leans on his own findings there throughout.

Similarweb released its 2026 Generative AI Landscape report, and the figure that travelled was prevalence. AI Overviews now appear on 43% of searches, up from 15% a year earlier, per TechCrunch's read of the report.

Similarweb's own summary page puts it at more than 40% of US searches triggering an AI Overview. It adds a 5.4% rise in the length of a Google search since AI Mode arrived.

That is the searcher-side counterpart to Google's claim that AI Mode questions run about three times longer than classic ones. The report's series runs through May 2026.

Take the prevalence number carefully. At 43% it is the highest on this page, far above the keyword-panel trackers in the statistics section.

Semrush sits near 16% across 10 million keywords and Conductor at 25.11% on an enterprise panel. Similarweb's figure is near BrightEdge's 48%, which tracks commercial verticals only.

Similarweb's data comes from a clickstream panel of real users rather than a fixed keyword set. So it is answering a different question.

But the report page publishes no methodology, no panel size and no detection method for this figure. So the distance between 43% and 16% cannot be reconciled from anything Similarweb has released.

Treat it as one more panel, not as a correction to the others.

The AI Mode number in the same report is far smaller than Google's, because it measures something else.

Visits to AI Mode rose from 126 million in June 2025 to 279 million in May 2026. That is more than double in under a year.

Google says AI Mode passed 1 billion monthly users in May 2026. The two cannot be read side by side.

One counts visits inside a panel of tracked web traffic. The other counts monthly active users across the web and the Google app.

So the growth rate is the usable part rather than the level.

The EU AI Act's transparency rules for AI-generated content apply from 2nd Aug 2026. Providers of systems that generate text, images, audio or video must mark those outputs in a machine-readable format.

That is so they can be detected as artificially generated.

A week before that date, Google signed the Commission's Code of Practice on Transparency of AI-generated Content. The post was by Karen Massin, its head of government affairs and public policy for EU institutions.

The Code was published on 10th Jun and had drawn about 190 signatories by late July. Signing it is the recognised way to show compliance with the marking obligation, though the Commission also allows "alternative equivalently adequate means".

Google's post points to SynthID, which it calls its "industry-leading digital watermarking technology", and to its work on the C2PA content-provenance standard. It signs with a caveat attached, warning that more regulatory complexity, while the technical solutions are still evolving, "could contradict Europe's goals for competitiveness and simplification".

The Code has a second half aimed at whoever publishes the output rather than whoever generates it.

Deployers must label AI-generated text that informs the public on matters of public interest. The exception is text that has been through human review and sits under editorial responsibility.

What none of this says is what any of it means for AI Overviews.

Google's post never mentions Search, AI Overviews or AI Mode. The Code does not mention search engines, and neither do the Commission's final Article 50 guidelines, published on 20th Jul.

So the obvious question is one nobody has answered on the record. Is a generated answer sitting above the classic links content that has to be marked?

And what would marking look like on a search results page?

We are logging the regime here, not an application of it to this surface. We will update this entry when a regulator or Google says which it is.

The Commission issued two non-compliance decisions against Google on 23rd Jul. The first fine, 460 million euros, is for Search self-preferencing.

Google gave its own shopping, hotels, transport and sports services more prominent placement than third-party services in Google Search.

The second fine, 430 million euros, is for stopping Play Store developers from steering users to cheaper offers elsewhere.

It is the first DMA fine Google has been given. Google has been ordered to end the conduct.

Continued non-compliance can draw periodic penalty payments of up to 5% of Alphabet's average daily worldwide turnover.

The part that matters on this page is not the fine. Alongside the decision the Commission said it will consider how Google plans to "apply the principles of the decision to AI Overviews and AI Mode, on which dialogue will continue".

That was reported by Tech Policy Press and RTE.

The DMA rule Google just lost on bars a gatekeeper from ranking its own services above rivals on unfair terms. That rule is now in play for the surfaces where Google decides what gets summarized and which links appear.

Nothing is decided. The plan on the table is Google's own, and the Commission has made no finding on either surface.

The dialogue is open.

Google said complying means "having to strip away real-time Search features Europeans love", naming hotel, flight and restaurant pricing. It has indicated it may go to court.

Google launched AI Overviews ("Aperçus IA") and AI Mode ("Mode IA") in France on 22nd Jul. The post was by Sébastien Missoffe, vice president and managing director of Google France.

Both surfaces went out the same day on desktop and mobile web and in the Android and iOS Google app. The new intelligent search box came with them.

Google credits Gemini 3.5 for AI Mode's reasoning and multimodal understanding. Search Live voice and camera input is part of the French launch.

It closes the last gap in Google's AI search map, more than two years after AI Overviews reached the US. And it lands two months inside the 23rd Sep commitment Google gave French publishers in June.

Google's own post never mentions neighboring rights, the reason France waited. It does point site owners at the Search Console control that keeps a site out of Search's generative AI features.

It says that control is available in France. And it repeats that the control is not used as a ranking signal for results outside those features.

French coverage supplies the rest. Google holds neighboring-rights contracts with about 450 French publishers and press agencies.

It told them in June that AI answers would also count toward those payments, per Blog du Modérateur. Next reports that no collective agreement has been announced and negotiations are still running.

That makes France the first market where Google has said AI answers count toward what it pays publishers. It is also the first place where an opt-out and compensation exist side by side.

So it is the test case for every other press market.

The terms did not hold. On 11th Aug the main publisher body took Google to the Autorité de la concurrence.

It says the launch came before any real negotiation.

The choice publishers were given, it says, was a licence-contract update or a withdrawal that costs them visibility.

The threat went on the record in the week of 20th Jul.

USA Today Co chief executive Mike Reed said "it's time to take a stand and say enough is enough". He is ready to block Google if the traffic slide continues.

People Inc chief executive Neil Vogel said "turning them off and blocking them entirely is 100% on the table". Google is down to about a quarter of his company's traffic, from more than half two years ago.

Reuters president Paul Bascobert said the company is "certainly looking at the economic trade-offs between search and AI summaries".

The Economist is not discussing a full exit, its VP of AI Josh Muncke noting that "traffic can be choppy and declining, but traffic is still traffic". The quotes are collected in PPC Land's write-up of reporting by the Wall Street Journal, Adweek and Nieman Lab.

The numbers behind the mood are Semrush figures the Journal cited, and they come in two forms. The first compares US organic Google traffic in June 2025 with June 2026.

On that measure USA Today's national paper is down 18% and Politico down 20%. CNN is down 31% and Business Insider down 31%.

The Journal's corrected chart of trailing twelve-month averages is harsher on several of the same titles. It has the Washington Post down 44%, Business Insider down 43%, Reuters down 35% and USA Today down 28%.

Time is down 27%, the Wall Street Journal itself down 18% and People down 3.2%. Against that, the Guardian is up about 16% and the BBC up 15%.

The word corrected matters here. The Journal revised its Semrush figures on 22nd Jul, the day after publication.

By then a much steeper set had already been reprinted across the search industry.

Business Insider had first been reported at more than 85%. The direction of the trend survived the correction, and for several titles the magnitude did not.

Search Engine Journal adds that Reddit is rethinking its 60 million dollar a year licensing deal with Google. Reed has been reported as prepared to delist within six to twelve months.

What makes it a threat rather than a decision is the plumbing. One Google crawl feeds both the index and the AI answers, so historically leaving the AI features meant leaving Search.

The Search Console control Google added in June is meant to break that link. It keeps a site out of AI features without being used as a ranking signal elsewhere.

When this was written, Google's own help pages still described it as rolling out to "a subset of website owners". So for most publishers the choice was all or nothing.

The control reached every property on 31st Aug 2026.

Google's response to the reporting was that its "AI features highlight links to the web and help creators and publishers grow their audiences", pointing again at the controls site owners have.

The AOP commissioned DJB Strategies to find where the declines are concentrated. The traffic half covers six months across eight major UK publishing groups and 10.8 billion pageviews.

That is about 80% of their total traffic once long-tail URLs were stripped out.

A consumer survey by Ipsos sits alongside it.

Between the first and second quarters of 2025, organic Google Search referrals fell 7.1% across all participants.

AOP managing director Richard Reeves wrote up the findings in Press Gazette. He notes that this was the quarter Google reported 12% Search revenue growth and expanded AI Overviews significantly.

Held at that rate, the study projects Google pageviews to these publishers halving by the third quarter of 2027. That is before any compounding effect from readers getting used to asking an AI instead.

The averages hide most of the story. News publishers lost about 5% a quarter against 18% for the consumer and business-to-business segment.

Reeves attributes that to the tighter restrictions on news in AI Overviews.

Traffic was segmented by topic and content type across 147 domains.

By topic, news rose 2%, sports 37%, travel 22% and health and wellness 7%. Everything else fell, with politics down 7%, finance and business down 23% and technology down 28%.

Content type mattered as much as subject. Guide and how-to referrals fell by more than 20%.

That is the shape of query a reader is most likely to hand to an AI answer instead.

Opinion and commentary rose 4% and interviews 1%, while product reviews were flat at a 1% decline.

Crossing the two produces the widest swings in the study. News interviews gained 63% and health and wellness analysis 46%.

Finance and business how-tos fell 52% and food and drink how-tos 54%.

Reportage, which carried more sessions than every other content type combined, fell 4%.

What the study does not do is isolate AI Overviews. It measures Google referral traffic quarter on quarter.

It does not compare pages that drew an overview against pages that did not. So the decline cannot be pinned on any single cause on this evidence.

That is why these figures are not in the statistics table on this page. Every row there is an AI-Overviews-attributed measurement.

Two limits go with the numbers. The traffic window is 2025, so it predates the France launch and the merger of AI Overviews and AI Mode.

And the participants are premium publishers, which Reeves says makes them unrepresentative of smaller sites.

Those sites lean hardest on programmatic revenue and can absorb a shock least.

A Google spokesperson told Search Engine Land that the Top Stories carousel inside AI Overviews is "fully live in the United States for mobile users".

It appears on searches about developing topics, highlights a reader's preferred sources, and Google says more regions are planned.

This is the productized version of the developing-topics carousel Google announced on May 27. It was first spotted in testing in late June.

It is one of the few spots where an AI Overview hands over ordinary article links instead of a summary. That is why Search Engine Land reads it as something that "can improve clicks to news publishers".

Whether it does is unmeasured. Search Console still reports impressions rather than clicks for AI features, so publishers cannot isolate what the carousel sends them.

Two weeks later NewzDash measured how often the carousel lands inside the overview, and found the UK showing it too.

Nick Fox, Google's senior vice president of Knowledge and Information, wrote on LinkedIn that Google is "now sending billions of clicks to websites every week through AI features in Search alone". That is on top of the "billions of clicks to the web every day" it says Search sends overall.

As Search Engine Journal reported, it is the first time Google has attached a specific number to AI-search referrals. It was offered to rebut the argument that AI answers are choking off traffic to publishers.

The figure cannot be checked. Google gave no baseline, denominator or timeframe.

Search Console still reports only impressions for AI features, not clicks. So no publisher can reconcile the aggregate against its own numbers.

As Forbes noted, Google is counting total clicks across an enormous query base while the independent studies measure click rates. Both can hold at once.

Total clicks can rise even as the click-through rate on any single result falls. Ahrefs put that fall at 58% on the top result when an AI Overview is present.

It extends the counterpoint Liz Reid made in 2025, now with a headline number. It still lacks the query-level data that would settle the dispute.

The Commission issued two sets of binding specification measures to Google on 16th Jul. One opens up Android features to rival AI assistants.

The other spells out how Google must share Search data under Article 6(11) of the Digital Markets Act. That is the reason this sits on an AI Overviews page.

The decision "specifies that AI chatbots offering search functionalities are eligible to receive shared data, and that, subject to anonymisation, Google should share the same data that it collects to optimise its own search services".

The Commission is explicit about why that clause exists: sharing matters "particularly" for AI chatbots with search functions "to avoid that Google Search's advantage extends to Google's AI chatbots too".

Google has owed rivals this data since the DMA took effect. But the Commission says its own offer was stripping out "between 90 and 100% of unique search queries" and "unduly restricting the potential pool of beneficiaries by excluding AI Chatbots that provide search services", so there had been "no meaningful uptake".

The new measures cover anonymised ranking, query, click and view data from free and paid search.

That means queries entered at any access point, plus query language and device type. It also covers the URLs users viewed, what they did with the results and where each result ranked.

Recipients may use it to improve query understanding and indexing. They may also improve the ranking and retrieval systems that AI chatbots rely on for grounding.

They may not train general-purpose AI models on it, use it for advertising or profiling, or systematically replicate Google's results.

Eligibility is narrow. It takes a genuine EU search service with at least 50,000 monthly EU users and two years of operation.

A startup under two years old that has raised more than 50 million euros also qualifies. An independent audit is needed before any data changes hands.

The clock is running. The eligibility form and an applications page are due by the end of August 2026.

Template licences and test samples follow by September.

The finished dataset is due by November, and a final pricing offer by January 2027.

Nothing here changes what AI Overviews or AI Mode do, and the decision names neither surface.

What it changes is the raw material available to everyone building against them.

From next year, a Perplexity or a privacy-focused engine can ask for that query and click stream. Only Google Search collects it at scale.

The data comes on a minimum seven-day delay and for up to five years each.

The thing to watch is what it produces. It could be credible alternatives to Google's AI answers.

Or it could be a compliance process that stalls the way the last offer did.

Ozone is a software platform that tracks ad opportunities and pricing for premium publishers rather than selling the ads itself. It gave Digiday US and UK benchmarking data covering roughly 20 billion impressions.

The cohort includes the Guardian, News UK and Dow Jones' Wall Street Journal.

Ad request volumes are the count of chances a publisher had to show an ad.

Between April and June 2026 they fell 39% to 41% year over year in the UK. In the US the fall was 32% to 37%.

Fewer visits means fewer pages, and fewer pages means less inventory to sell.

Across the first half of 2026, combined programmatic spend in the cohort was down 30.6% year over year. But the two markets diverged sharply: UK spend fell 14.3% while US spend fell 44%.

The difference is yield. UK eCPMs were up about 30% year over year in June, against about 7% in the US.

So British publishers recovered more of the loss by charging more for what was left.

Ozone's Eliza Simonova called scarcity attracting higher payment "a healthy sign even as total spend declines".

"The open web is shrinking", Ozone chief operating officer Danny Spears told Digiday. Gabe Dorosz of the news-media association INMA used the terms the ad market has run on for fifteen years: "AI-driven traffic declines have now killed the idea that there actually is infinite supply".

What this is not is an AI Overviews measurement. Ozone counts ad requests, not overviews.

Both Digiday and PPC Land's write-up attribute the contraction to zero-click AI search broadly. Digiday notes that platforms, "particularly Google", now answer in place rather than sending the visit on.

PPC Land names a second cause outright, quoting Nova Studio's Matt Barash that "discovery has moved" from search toward social feeds.

Nothing in the data separates the two. So read it as the shape of the decline rather than a measurement of what caused it.

Its value here is that it sits downstream of everything else on this page. The click studies in the statistics section measure the traffic that stops arriving.

This measures what happens to the business when it does. It is the number underneath the publishers who spent the following week saying they might block Google altogether.

On 14th Jul 2026 the ZAK issued formal orders against AI search services for the first time. The ZAK is the licensing and supervision commission that coordinates Germany's 14 state media authorities.

Two proceedings targeted Google's AI Overviews and Perplexity's chatbot and AI news page. The Hamburg/Schleswig-Holstein and Berlin-Brandenburg media authorities ran them.

The ZAK says it is the first time German media law has been applied to AI search and AI chatbots.

The central finding is the one publishers have been arguing for two years. AI answers are the provider's own content, not a display of someone else's.

So the liability privilege in the EU's Digital Services Act does not cover them.

"AI search engines and chatbots are content providers, and we will consistently apply German media law to them from now on," ZAK chairman Thorsten Schmiege said, per Reuters.

The specific complaint against AI Overviews is placement. The regulator says the generated answer is presented as an essential part of the results, sitting prominently at the top.

That leaves the classic list of links harder to find and, in its words, impermissibly discriminated against.

The chatbot half of the decision reaches further. An AI chatbot may attach third-party material as "sources", "further reading" or whole link lists.

Then, the ZAK says, it is deciding what gets found.

That meets the German definition of a media intermediary and brings transparency and diversity obligations with it.

As The Decoder reports, Perplexity's order also cites a missing German representative and inadequate transparency disclosures.

The reasoning rests on a legal opinion the media authorities published three weeks earlier. Its authors are Prof. Jan Oster and Prof. Christoph Busch.

It covers how integrating AI applications into search engines sits with German media law.

This is a regulator arriving at the conclusion the Regional Court of Munich I reached seven weeks before it. A generated answer is Google speaking.

The difference is reach. The court settled one dispute between named parties.

The ZAK is applying an ongoing supervisory regime to the product itself.

That regime comes with a complaint route for publishers who believe they are being disadvantaged.

Google told Reuters its AI summaries "enhance the search experience in Germany" and "help users discover new content and ask follow-up questions", and said it will appeal.

Perplexity declined to comment beyond saying it complies with GDPR and holds SOC 2 Type II certification. The orders are not final while appeals are open.

Common Sense Media's Youth AI Safety Institute rated Google Search's AI features "Unacceptable Risk", its lowest tier, for children. Across 2,624 test interactions in seven test plans, the two features failed seven of the Institute's eight AI principles.

They also failed all five severe-harm "Red Lines".

The assessment's headline: "Google Search's AI answers are not safe enough to be kids' default answer machine."

AI Overviews, the answer that appears automatically above the results, fared worst. It pointed users to an appropriate resource in only 58% of 214 crisis prompts, where AI Mode managed 77%.

It also missed 29% of explicit statements referencing suicide.

As PBS reported, neither feature can be turned off by parents or schools, unlike Google's standalone Gemini app. That is despite roughly three-quarters of US teens and tweens now using AI answers in search.

Google said its AI features have "strong quality and safety guardrails" and called the testing flawed for using "a narrow set of ambiguous and contrived queries".

In a post marking Google Images' 25th anniversary, Google said it is bringing image generation directly into AI Overviews: "Using our latest Nano Banana model, this update transforms a simple text prompt into a high-quality, custom visual made completely from scratch."

9to5Google identifies the model as Nano Banana 2 Lite.

Google says the feature "will start to roll out over the coming weeks in English, for all regions that currently support image creation in AI Mode", so availability follows the existing AI Mode image-creation footprint rather than the full AI Overviews footprint.

It extends the merged AI search experience. The generative image tools already in AI Mode now sit in the overview above the results.

What happened next, for anyone reading this later: generated illustrations were first spotted in live overviews on 13th Aug. Google stopped that test on the 17th, calling it a small experiment.

This announcement has not been withdrawn.

Further Exploration, the first of the five link improvements Google announced on 6th May, started turning up in live results. SERP analyst Brodie Clark posted examples on X and in his SERP Alerts newsletter on 2nd Jul, describing "an early iteration of this within AI overviews, which features a 'further exploration' section at the end of the answer".

Google's own description of the feature, from the May announcement, is that "you'll start to see suggestions for where to go next at the end of many AI responses", linking to "unique articles or in-depth analyses on different facets of your topic". The sighted version matches that: a block of suggested reading below the generated answer.

The reason it gets its own entry, separate from the announcement, is direction.

Almost every AI Overviews change this page tracks moves in one direction. The answer absorbs more of the work the click used to do.

This one adds outbound links, and it sits at the end of the response. That is where an unsatisfied reader is most likely to still be looking.

Treat it as a sighting rather than a launch. Google has not announced a rollout, named a market or given a date.

There is no measurement of how often the section appears or what it does to clicks.

What is established is that the feature has left testing and is reaching real searches.

Saharsh Agarwal of the Indian School of Business and Ananya Sen of Carnegie Mellon's Heinz College ran the study. They describe "The Impact of Google AI Overviews on Publisher Traffic and User Experience" as the first randomized field experiment on AI Overviews.

A custom browser extension recruited through Prolific put 1,065 US desktop Chrome users into groups. Some saw Google normally, some had AI Overviews hidden, and the rest were sent to AI Mode.

Across 396 control and 374 treatment users the study analysed 68,089 unique searches. An AI Overview was present on roughly 41% of queries.

Every other click figure on this page compares queries that happened to have an overview with queries that did not. This one changes the overview and holds the user constant.

Outbound organic clicks rose from 0.37 per search to 0.62 when the overview was hidden. The probability of a search ending with no click fell from 0.73 to 0.54.

The authors put that at a 39.8% reduction in clicks attributable to AI Overviews.

The effect sits almost entirely on informational queries, worth about 0.26 extra clicks per search. Navigational and transactional estimates were inconclusive.

The June revision added the part that matters most for the argument Google has been making. It found no statistically significant difference across bounce rate, quick returns to the results page and time on site.

That compares the clicks sites get with overviews against the extra ones they get without.

Google's public position is that click quality is up even if volume moved. On this panel the recovered clicks were ordinary clicks.

Two caveats belong with the numbers. It is an SSRN working paper that has not been through peer review.

It was first posted in April 2026 and last revised on 17th Jun.

And 1,065 self-selected desktop Chrome users in the US is a small panel. It is tiny next to the multi-million-keyword trackers elsewhere on this page.

What it buys in exchange is causality, which none of the others can claim.

SparkToro's zero-click study ran on Similarweb desktop and mobile clickstream data for US Google searches from January to April 2026. It found searches ending without a click 68.01% of the time, as Search Engine Land reported.

That is 7.56 points above the 60.45% SparkToro measured in 2024. Earlier readings were 64.82% in 2020 and 50.33% in 2019.

Over the two years to 2026 the share of searches producing at least one click fell 9.51 points. That is a 22.9% decline.

The share that led to another Google search rose 7.2 points.

Searchers are not only clicking out less, they are staying inside Google and searching again.

What the study does not do is attribute the shift to AI Overviews. It measures clicks, not overviews.

The AI-answer figures it cites for context come from the same third-party trackers already in the statistics section. Those are overviews on more than 20% of searches, and top-result click-through down nearly 60% where one appears.

SparkToro also cautions that its year-to-year comparisons rest on different providers, panels and methods. So the 2019 to 2026 line is a trend rather than a like-for-like series.

Its own 2019 figure is the useful check on any single-cause story. Half of all searches already ended without a click five years before AI Overviews launched.

The AI Mode figure has no equivalent anywhere else on this page. Just 0.34% of searches in the panel moved into AI Mode across those four months.

SparkToro says that means the tab "played only a limited role" in the click decline.

That does not contradict Google's 1 billion monthly users, which counts people rather than share of searches. But it is the clearest illustration yet of the gap between the two.

There is a very large user base and, on this panel, a very small share of actual search volume.

In the period measured, the click decline was happening on the ordinary results page with an overview on top. It was not happening inside the AI Mode tab.

The UK Competition and Markets Authority ordered Google to let publishers stop "their content being used to power AI features in search, such as AI Overviews" without losing their normal search rankings. It used the strategic market status it gave Google in October 2025.

The CMA calls the opt-out a world first, and it is the first binding regulatory remedy against AI Overviews anywhere. Sarah Cardell, the CMA's chief executive, framed it as bargaining power: "With features like AI Overviews rapidly reshaping online search, it is crucial that content publishers, including news organisations, have appropriate bargaining power over how their content is used."

Three requirements beyond the headline opt-out matter for site owners. Google must also let publishers "opt-out of allowing their content to be used for the 'fine-tuning' of AI models".

That is a separate lever from the AI-features opt-out and from the Google-Extended token that governs Gemini training.

It must attribute publisher content "using clear links" in AI-generated results. It has nine months to implement everything, but the CMA says it "expects important parts of the controls to become available to publishers well before that deadline".

Compliance reports carrying key data and metrics are due every six months for the first year.

The order grew out of the conduct-requirements consultation the CMA opened in January. Google shipped the Search Console side of it the same day the order landed.

On 28th May 2026 the Regional Court of Munich I granted a temporary injunction (case 26 O 869/26). It bars Google from spreading false factual claims about two Munich-based publishers through its AI-generated search overviews.

For certain queries the overview had linked the two companies to scams, subscription traps and dubious business practices. It mixed them up with genuinely questionable firms and drew connections that appeared in none of the sources it cited.

Google must remove the material and stop its systems generating comparable claims about them.

The reasoning is what makes it a landmark. The court found the AI function "generates a coherent, flowing text" that "appears as direct information from Google, not as a mere forwarding of third-party content", as heise reports.

It treated Google as the direct author of a false statement, not as an intermediary passing on someone else's words.

That is the search safe harbour that has protected result pages for two decades. It does not extend to a generated answer.

heise notes the decision goes further than an earlier Frankfurt case. That case declined a cease-and-desist application rather than banning specific claims.

Two caveats. This is summary proceedings, a preliminary injunction rather than a final judgment, and Google retains the right to appeal.

At the time of the reporting Google had not commented.

It also binds Google only as to these two publishers and these claims. What travels is the principle, and it arrived four days before the CMA order.

Within a week, AI Overviews acquired a publisher opt-out in the UK and an author in Germany.

Any answer engine that generates prose about identifiable people or companies now has a European precedent to reckon with. It is the same question the Ontario defamation suit puts to a Canadian court.

Google brought preferred sources into AI Overviews and AI Mode. Websites a searcher has picked in Search personalization settings are now clearly labeled when they appear in AI responses.

Product manager Duncan Osborn wrote that "people are twice as likely to click through to a Preferred Source". More than 345,000 unique sources had already been selected, he added.

9to5Google noted the labels initially appear only when a chosen site naturally surfaces in a response. Google plans to have preferred sources appear more often in AI features over time.

Sources are managed at google.com/preferences/source.

The lever had become usable outside English four weeks earlier, when preferred sources expanded to every language Google Search supports.

The same announcement added a prominent carousel that highlights preferred sources on searches about developing topics. It promised a similar carousel of perspectives from forums, discussions and social media.

And it expanded the "Highly Cited" badge to more web article links.

For publishers this is a rare direct lever inside the AI answers. Readers who pick a site see it labeled, and click it more.

Google announced AI performance insights in Merchant Center. It is a report for retailers on how their products perform when a shopping journey begins in an AI answer.

It covers three surfaces together: AI Overviews in Search, AI Mode and the Gemini app.

The report has four parts. Share of voice benchmarks a brand's visibility "against brands similar to yours on shopping journeys starting from AI Mode, AI Overviews, or the Gemini app".

Shopping funnel performance splits that visibility across the discovery, evaluation and purchase stages.

Product term insights show which product terms are popular and what share of them a brand holds. Product attributes insights score attribute completeness, flagging the structured data a feed is missing.

What makes it stand out on this page is who it is for. The Search Console generative AI reports that shipped a week later measure impressions for site owners.

This measures competitive share for retailers, on the same surfaces.

Neither reports clicks. So the gap between appearing in an AI answer and getting traffic from it stays unmeasured in Google's own tooling.

Two caveats. Google said the report "will be rolled out to the U.S., Canada, Australia, India, and New Zealand in the coming months" and has announced neither a date nor general availability.

So this is a commitment rather than a shipped feature.

And share of voice is Google grading visibility on its own surfaces by its own definition. There is no published methodology behind the benchmark.

PPC Land covered the announcement the same day.

Eric Van Buskirk and his team at ClickStream Solutions analysed anonymized clickstream data provided by Surfer SEO. It covered about 846,000 US Google Search sessions from February and March 2026.

They published the findings in Search Engine Journal.

The study's own methodology page puts the total at about 860,000 sessions rather than 846,000. The write-ups all use the lower figure, so that is the one quoted here.

Cursor position was sampled once a second, up to 60 samples per session. Each sample recorded cursor and viewport coordinates and a timestamp.

Three datasets were used: a balanced set of 74,848 sessions and a representative set of 99,994.

A filtered set of 99,994 was drawn from 238,280. It dropped sessions with fewer than 3 or more than 25 samples.

Comparisons used Mann-Whitney U and Kruskal-Wallis tests.

The headline finding is about search intent. With no overview on the page, 12% of navigational searchers were still active at 21 seconds.

For local searchers it was 32%. That is the long-standing pattern where intent predicts dwell time.

With an overview present, all five intent types cluster between 41.9% and 48.5% at 21 seconds. At three seconds they sit between 91.9% and 93.1%.

The five are informational, local, navigational, transactional and video. Navigational sessions alone go from 12% to 46%.

The cursor measures point the same way. With an overview, cursors were still about 44% of the time against 29% without.

Yet they covered 83% of the viewport against 66%. The study reads that as evaluating rather than scanning.

On navigational queries, cursor spread widens from 8% to 27.5%.

Scrolling stops being one directional. The share of sessions that reverse direction at all rises from 51% to 59%.

Among those that do, the median session spends 47.5% of its scrolling going back up the page. Without an overview that is 27%.

On navigational queries that share goes from 23% to 44%.

This study contains no click or traffic figure at all, and it comes with limits of its own.

"Active" means a session was still producing cursor samples.

The methodology page says that is not evidence anyone was reading intently. Cursor position is treated as an attention signal rather than as gaze.

Surfer's panel is made up of marketing-interested users and may skew towards search-savvy behaviour. Only 1.1% of the sessions were marketing-related queries, though.

All of it is US data.

Read alongside the click studies in the statistics section, it describes what happens before a click. It does not describe the click itself.

Kevin Indig, who published the findings the day before in his own newsletter, draws the practical conclusion for site owners. On a page with an overview your listing is read more than once.

So the title and description compete in a slower, more comparative setting than rank position alone suggests.

At I/O 2026 Google merged AI Overviews and AI Mode into one search experience across desktop and mobile worldwide. The overview is now the quick answer, and follow-ups continue into an AI Mode conversation.

Both launch from what Google called its biggest search box upgrade in more than 25 years.

Gemini 3.5 Flash became the default model for the merged experience globally. Sundar Pichai's keynote put AI Overviews past 2.5 billion monthly active users and AI Mode past 1 billion.

The merger formalizes what the January handoff started. AI Overviews are no longer a feature bolted onto the results but the entry point to Google's AI search.

Google Marketing Live, held alongside I/O, previewed Gemini-generated conversational ad formats for the new experience.

SISTRIX used its search visibility index to estimate how often an AI Overview supplants links to a UK newsbrand section. It gave the results to Press Gazette.

The window is 14th Jan to 24th Apr 2026.

Health is the most exposed vertical by a distance at 72%. Tech follows at 49%, then travel at 39%, lifestyle at 37% and money at 35%.

At the other end, football sits at 5% and puzzles at 7%, with TV and wider sport both at 15%.

Press Gazette attributes part of that to the nature of the results. The rest it puts down to Google not running AI Overviews on breaking news.

Steve Paine of SISTRIX draws the operational conclusion for publishers: "The results here show that some topics carry much higher risk than others. If your business model relies on clicks from search, assess the exposure rate and consider focusing on low-exposure topics and niches."

He is blunter about why health tops the table: "Health is a high-value topic and Google appears to want to take control of that. Other topics such as sports, that might be harder or more expensive for Google to track, have a much lower rate of exposure."

The method has real limits. Nine newsbrands were included: the Guardian, Daily Mail, Telegraph, Times, Independent, Sun, Mirror, Express and CNN.

They were chosen from the most visible UK news sites in search. Then they were filtered to those whose URLs mark content type clearly.

Not every brand appears in every topic, because each breakdown needs a cleanly defined section.

It is also UK data. Press Gazette says it understands the pattern tracks closely with US results.

That is the paper's understanding rather than a published measurement. So treat the numbers as UK figures until someone publishes the American ones.

Press Gazette pairs the exposure rates with another SISTRIX finding. Top-result click-through falls from 27% to 11% when an overview appears.

That figure is the German keyword study already on this page, restated, not a separate UK measurement. So the two should not be read as corroborating each other.

Two publisher figures in the same piece set the scale. Healthline was the fastest-falling major UK news site in March, down 48% year on year on Ipsos iris data.

People Inc, whose Google traffic has dropped 63% in two years, is moving from evergreen content towards reporting new information.

Most measurement on this page finds AI Overviews taking clicks away.

A working paper by Peibo Zhang, Ruomeng Cui and Dennis J. Zhang finds the opposite for one platform. The design behind it is unusually clean.

It is posted to arXiv and not yet peer reviewed.

The identification trick is Google's own content policy. Safe-for-work subreddits are indexed and can be cited in AI Overviews.

Not-safe-for-work subreddits are indexed in organic search but barred from AI Overview summaries.

That gives a treatment group and a control group inside one site. The authors built a balanced panel of public subreddits from the Arctic Shift archive.

They tracked daily comments and daily unique commenters from 1st Jan 2024 to 31st Jul 2025. The 15th Aug 2024 international rollout is the treatment date.

Daily comments in safe-for-work communities rose 12.0% against the control, and daily commenting users rose 12.4%.

Those two rates are nearly identical. So the growth came from more people joining discussions rather than existing commenters posting more.

The gain is not spread evenly. The authors split content into search goods, where a fact can be summarised, and experience goods.

Experience goods are those where the value is opinion, advice and discussion.

Experience-good communities responded 2.3 to 2.8 times more strongly.

Their reading is that a static summary cannot substitute for a discussion thread. So it works as a discovery route into one instead.

The same framework predicts the negative effect other researchers have measured for Wikipedia.

Then AI Mode arrived. The authors extend the panel to December 2025.

They use the 21st Aug 2025 global rollout as a second treatment date.

The experience-good premium falls by 59% for commenting users and turns slightly negative for comments.

The authors argue that this is the interface rather than growing trust in AI answers. A trust story predicts the decline should land on search goods, where the summary already answers the question.

It lands on experience goods instead.

Read carefully what this does and does not measure. It counts comments and commenters inside Reddit, not clicks, referral sessions or publisher traffic.

And it covers one text-discussion platform.

So it is not evidence that AI Overviews send traffic to your site. That is why we have kept it out of the statistics table.

It is evidence that the effect's direction depends on what a site hosts and which interface a searcher uses.

Haofei Xu, Umar Iqbal and Jacob M. Montgomery of Washington University in St. Louis ran the study. It covered 55,393 trending queries across 19 topic categories.

The run covered 13th Mar to 21st Apr 2026 and captured 7,583 AI Overviews.

The paper is a preprint under review.

Take the activation number with its corpus attached. Overall activation was 13.7%, but these are Google Trends queries and 51.4% of them are sport.

So this is a specific slice of search rather than a general prevalence estimate.

Inside that corpus the phrasing split is wide. Queries whose first word is one of fifteen interrogatives triggered an overview 64.7% of the time.

Everything else triggered one 9.5% of the time.

The fifteen words: who, what, where, when, why, how, which, is, are, was, can, do, does, did and has.

The source-selection finding is the one to act on. Averaged across the overviews, 29.8% of cited domains appeared nowhere on the corresponding first page.

At URL level 28.5% came from hosts the first page did not surface for that query.

Citation is also broader than ranking. The top five hostnames account for 20.0% of AI Overview citations against 39.1% of first-page citations.

And 56.3% of the hosts cited across the 40 days were cited exactly once.

The five most-cited hosts were youtube.com at 5.49% and en.wikipedia.org at 4.39%.

Then came facebook.com at 3.68% and instagram.com at 3.65%. Fifth was usatoday.com at 2.80%.

On source quality the study cuts against the common complaint. Cited domains scored higher on the paper's credibility measure than the first-page results beside them.

The scores were 0.732 against 0.645. The off-page citations scored higher still.

Fidelity is where it gets uncomfortable. Decomposing the overviews into 98,020 verified claims, 11.03% were not supported by the sources cited.

Of those, 6.98% were mentioned by no cited source and 2.66% contradicted by one. In 1.39% the cited sources disagreed.

The authors bracket that figure themselves rather than leaving it to a reader. Their pipeline does not crawl social and video sources.

So they put the floor at roughly 5.3% if every such claim turned out to be supported. They call the combined omitted-and-incorrect rate a conservative ceiling.

The publisher economics are measured indirectly but plainly. Of the 61,212 URLs cited, 50.63% carried visible advertising.

The authors call that a lower bound, because they treat uncrawled social and video pages as carrying none.

Google's own inventory comes through untouched. Where sponsored ads and an overview appeared on the same page they occupied separate regions.

There was no sponsored slot inside the overview container. So the overview adds to Google's ad space while substituting for the click to the publisher.

Google announced five link changes for AI Mode and AI Overviews, framed as helping people explore the web.

They are:

• a Further Exploration section with suggestions for where to go next at the end of many AI responses
• more links placed directly next to the text they support
• a hover preview showing the website name or page title before a click
• labels that make links from a searcher's active news subscriptions stand out
• previews of perspectives from public discussions and social media with the creator or community named

Search product vice president Hema Budaraju wrote that in testing people were "significantly more likely to click links that were labeled as their subscriptions".

Coverage framed the changes against the click-loss studies this page tracks. The preferred-sources rollout three weeks later extended the same idea from subscriptions to any site a reader chooses.

Ashley MacIsaac, the Juno-winning Cape Breton fiddler, filed a statement of claim against Google LLC in February. He filed in the Ontario Superior Court of Justice, seeking at least C$1.5 million.

An AI Overview about him stated he had three convictions. They were sexual assault, internet luring involving a child and assault causing bodily harm.

It also said he was listed on the national sex offender registry.

None of it was true. The summary had folded in coverage of a different man in Atlantic Canada who shares his surname.

The damages sought are split into C$500,000 each of general, aggravated and punitive damages.

The harm was immediate and offline. The Sipekne'katik First Nation confronted him with the summary in December 2025 and cancelled one of his scheduled concerts.

It later issued a public apology.

Billboard Canada quotes him saying "I feared for my own safety going on stage because of what I was labelled as".

The claim argues Google "knew, or ought to have known, that the AI overview was imperfect and could return information that was untrue".

It puts the liability question plainly: "If a human spokesperson made these false allegations on Google's behalf, a significant award of punitive damages would be warranted. Google should not have lesser liability because the defamatory statements were published by software that Google created and controls."

Google has not commented on the suit. In December it said AI Overviews are "dynamic and frequently changing" and that it uses error cases to improve its systems.

It is the first high-profile case brought by an individual rather than a publisher or a company. It asks a Canadian court the question the Munich court answered three weeks later in Germany.

When a generated answer defames someone, is that Google speaking?

SISTRIX tracked 82,619 prompts once a week for 17 weeks, in six countries. The surfaces were AI Overviews, AI Mode and ChatGPT Search.

It found AI Overviews by far the most stable of the three.

An average overview cites about 11 domains, and 8 of them are there every week.

That stability is unevenly spread. In 53% of prompts not a single source changed across the whole 17 weeks.

In 28% there were isolated changes. In the remaining 19% the sources churned as fast as AI Mode's, at 46% new domains a week.

The study checked that the frozen half is not a measurement artifact. For 87% of those stable prompts the answer text changed from week to week.

The cited domains stayed identical. So Google writes new sentences every week and draws them from the same eight sites.

For a publisher that is blunter than any click figure in the statistics section. Being cited is durable, and so is not being cited.

The study's own summary: "Once you're in, you stay in. If you're not in, you don't get in either."

The comparison with AI Mode is the other important finding. On the same prompt the two surfaces cite different domains 83% of the time.

That is an overlap of 0.17 on the Jaccard index. AI Mode replaces 56% of its cited domains weekly, against ChatGPT Search's 74%.

One caveat on that overlap. The measurement window closed on 8th Apr 2026.

That is six weeks before Google merged AI Overviews and AI Mode into one experience at the I/O event. So it describes the two as they were before the merger.

Content type predicted durability better than anything else measured. News articles held a place in the stable core only 1.4% of the time.

That is the lowest of any category.

Video was highest at 24%.

The panel is SISTRIX's own AI research index. So this is vendor panel data like the rest of the statistics section, rather than ground truth.

The study says its figures are conservative. Measured per URL instead of per domain, the weekly drift runs about 15% higher.

PPC Land covered it the day it was published.

Google rolled preferred sources out globally in all supported languages on 30th Apr. It had run the feature worldwide in English only since late 2025.

Readers add outlets with the star icon next to Top Stories or at google.com/preferences/source, per 9to5Google.

Google's numbers at the time: readers are twice as likely to click through to a preferred source. More than 200,000 unique sites had been picked, a figure Google put at 345,000+ four weeks later.

It earns a place on this page because of what happened next. On May 27 Google extended the same selections into AI features.

A reader's picks are now labeled inside AI Overviews and AI Mode answers.

That is one of the few direct publisher levers in AI search.

Until this expansion that lever only worked for English-language readers.

Search Engine Journal notes that Google describes the control as working alongside its ranking systems. It up-ranks sites a user has asked for.

Its documentation is explicit that preferred-source picks do not override relevance.

Researchers at the New Jersey Institute of Technology and co-authors published the result at SIGIR 2026. They compared what Google Search, AI Overviews and Gemini 2.5 Flash retrieve for the same query.

All responses were collected on 7th and 8th Dec 2025.

The scale is 14,212 queries, including an 11,500-query benchmark the authors released publicly. Results came through SerpAPI simulating a mobile device in Newark, New Jersey.

A separate desktop sample was run as a check on the result.

Their prevalence figure is 51.5%, measured on the ORCAS subset of representative real-user queries. Across the whole benchmark, which is weighted towards questions and product queries, it is 65.6%.

The source comparison carries a lesson for anyone tracking citations. Jaccard similarity between the engines' retrieved sources runs between 0.11 and 0.18.

So on average only 18% of the sources returned by an AI Overview or the first page appear in both.

That 18% is not the same measurement as the AI Overviews against AI Mode overlap already on this page. This one compares an overview against the first page of the same search.

The Ahrefs and SISTRIX figures compare two AI surfaces with each other.

Now the finding that bears on a robots.txt decision. Sites that disallow Google-Extended were significantly less likely to be retrieved by AI Overviews.

That comes from a linear probability model with query fixed effects. It is despite Google stating that the token does not affect AI Overview access.

The authors flag it as surprising rather than damning, and two limits belong with it. The estimate is relative to the other engines in the comparison.

A source had to be retrieved by at least one of them to enter the dataset at all.

And an association in a single December 2025 snapshot is not a demonstrated cause. Sites that block AI crawlers differ from sites that do not in ways this design cannot separate out.

The paper does not claim to resolve that.

Google's documentation has not changed and we have not restated it here. The token governs Gemini training and grounding, and Google says it "does not impact a site's inclusion in Google Search nor is it used as a ranking signal in Google Search".

Our Gemini crawler explainer sets out what it does control.

Two further results are useful for context. Both generative surfaces were less likely than classic search to retrieve popular, government and education domains.

They were more likely to retrieve Google-owned properties.

AI Overviews were also less consistent than classic search. Results moved across repeat runs of the same query and across small edits to its wording.

That is a reason to be careful reading any single citation check as a stable result.

Seer Interactive's 2026 update draws on 5.47 million queries. It is the first large dataset to show click-through rates on AI Overview queries moving back up.

Organic CTR recovered from 1.3% in December 2025 to 2.4% in February 2026.

The finding to act on is about citation. Brands cited inside an AI Overview earn roughly 120% more organic clicks per impression than uncited brands.

Recovery or not, CTRs remain far below the pre-AIO baseline. Seer's September 2025 analysis had informational-query organic CTR down 61% since June 2024.

So the remaining clicks concentrate on whoever is cited in the answer.

The clearest published split of AI-era publisher damage by content category comes from an ad exchange, not a search tool. Index Exchange analysed 1,200 publishers on its platform through 2025 and published the result on 31st Mar 2026.

Its headline finding: "69% of publishers experienced year-over-year declines in ad opportunities, our closest proxy for monetizable traffic".

Ad opportunities are slot requests, the ad calls a publisher's pages generate. So this measures monetization rather than traffic.

The split is the useful part. Categories built on information-dense, answerable content saw both the largest increases in AI Overviews and the steepest declines.

Health and careers ad requests fell 40 to 50%.

News and politics, which the company says has significantly lower exposure to overviews, fell 7%.

Index Exchange reads that as a content property rather than a favour. Breaking stories, opinion and curated judgment resist compression into one generated answer.

A symptom list or a resume how-to does not.

Marybeth McGaugh, its chief customer officer, wrote the report.

PPC Land, reporting the same analysis, puts the average decline across that 69% at 14%. It says the exposure comparison was drawn against Semrush measurements of AI Overview prevalence.

It also carries the counterweight the report makes itself. Among highly ranked news publishers, spend sometimes rose while audience measures softened.

Scarcity of trusted inventory can lift CPMs.

What this does not establish: nobody's clicks were counted here. It is a category-level correlation between overview exposure and ad calls.

It is not a measured effect on any publisher's traffic. That is why it stays out of the statistics table on this page.

Alongside SISTRIX's per-vertical exposure rates it points the same way for health. Do not pair its 7% with the AOP study's 7.1%.

One is a year of ad requests in one category. The other is a single quarter of Google referrals to eight UK publishing groups.

On 20th Mar 2026 the US District Court for the District of Columbia granted Google's motion to dismiss in full.

The case is Helena World Chronicle, LLC v. Google LLC (No. 23-cv-3677). The ruling covers the amended complaint.

The plaintiffs are two small news publishers, Helena World Chronicle in Arkansas and Emmerich Newspapers in Mississippi. They sued on behalf of themselves and other publishers.

Their five counts ran from monopolization of general search and of online news to an unlawful tie. They also covered the Android, YouTube and DeepMind acquisitions, and Google's 2016 agreement with Apple.

The tying count is the one that matters for this page.

Google argued that Search and AI Overviews are not separate products, which is what a tie requires. The court's answer: "They are correct."

Its reasoning was that the alternative proves too much. "If AI Overviews is a separate product from Google Search, then so is potentially every other type of search result Google delivers."

A results page returns organic links, ads, vertical offerings and AI Overviews. The court could not see demand for one of those as separable from demand for Search itself.

It then handed the publishers a map.

What they likely meant to plead, the court said, is a tie to the Gemini chatbot, not AI Overviews. It noted in a footnote that "Gemini may very well be a separate product. Gemini is not a 'search feature' that appears on the Google SERP."

AI Overviews "today is not a chatbot and is distinct from Gemini", the same footnote adds.

Keep that line in mind now that AI Overviews is the front end of one experience with AI Mode. The product it was distinguished from keeps moving closer.

The rest of the case failed on older antitrust ground.

The publishers had no antitrust standing on the search-market counts. Their injuries were suffered in online news rather than in general search.

Their estimate of a 66% share of online news, inferred from total visits to Google, YouTube and Gemini, was "utterly fanciful" and "at once both wildly overinclusive and underinclusive".

The acquisition counts were time-barred.

"The acquisitions of Android, YouTube, and DeepMind may very well have been anticompetitive. But Plaintiffs are 10 to 20 years too late; those claims are now stale."

What the ruling does not decide is whether AI Overviews costs publishers traffic. It is a pleading decision about who may sue and over what, and the accompanying order is final and appealable.

One detail: these publishers had asked for an opt-out from AI Overviews without demotion on the results page. The UK regulator ordered exactly that in June.

The pattern this page keeps finding is that publisher controls arrive through regulators rather than through courts.

The same judge has Google's motions to dismiss the Chegg and Penske Media complaints in front of him. They were argued at a consolidated hearing on 25th Aug 2026.

The CMA opened the consultation that produced the June order on 28th Jan.

It proposed exactly what publishers had been asking for since 2025: "Publishers will be able to opt out of their content being used to power AI features such as AI Overviews or to train AI models outside of Google search". It added practical steps to attribute publisher content in AI results.

Sarah Cardell, the CMA's chief executive, promised "a fairer deal for content publishers, particularly news organisations, over how their content is used in Google's AI Overviews". Responses were due by 25th Feb.

Google's answer, reported when the responses were published, was a qualified yes. The Register reported that Google said it would eventually let sites opt out of generative AI features in Search.

It noted that "details and timelines, however, were conspicuously absent".

Google made the same argument in public that month. Separating AI features from ordinary Search is a large engineering problem, because one crawl and one index feed both.

This is the step our timeline used to skip. It is useful to remember when Google says a control is coming.

The distance from "we are exploring updates to our controls" in March to a shipped Search Console toggle in June was about ten weeks. It closed only once the CMA made it a binding requirement.

The reason to date the commitment is simple. It lets you measure the gap between a promise to a regulator and a control a publisher can switch on.

Chartbeat segmented its global client network by size and found the damage scales inversely with the publisher. Search referral traffic to small publishers (1,000 to 10,000 daily page views) fell 60% over two years.

Mid-sized publishers (10,000 to 100,000) lost 47%, and large publishers (over 100,000) lost 22%.

Narrowing to a single year, page views from Google Search fell 34% between December 2024 and December 2025. Google Discover fell 15% over the same period.

Axios received the data exclusively. Chartbeat has not published it independently.

The second finding matters as much as the first. ChatGPT referrals grew more than 200% in that window.

But chatbots still account for less than 1% of publisher page-view referrals.

The traffic leaving search is not arriving anywhere else. That is the hole in the argument that publishers are simply being redistributed to AI platforms.

Read it alongside the click studies in the statistics section rather than instead of them. Those measure the click rate on a query, while this measures arrivals at a site.

So it captures everything happening at once: AI Overviews, ranking changes, Discover swings and platform shifts. It cannot attribute the loss to any single cause.

What it does establish is the distribution. The sites least able to absorb the loss are taking nearly three times the hit of the largest ones.

BrightEdge data covers February 2025 to February 2026 across nine industries. It puts AI Overview presence at roughly 48% of tracked queries, up 58% year over year.

Healthcare leads at 88%, followed by education at 83% and B2B technology at 82%.

That reads far higher than broad-panel figures like Semrush's 15.69%, and both are right. BrightEdge tracks commercial and enterprise keywords, where Google triggers AIOs far more often than across the whole query stream.

See the statistics section for how the panels compare.

SISTRIX founder Johannes Beus measured AI Overviews across more than 100 million German keywords. He found them on just over 20% of them.

Where one appears, "the CTR at position 1 drops from 27% to 11%, a loss of almost 60%".

Scaled across the market, SISTRIX puts the cost at 265 million organic clicks a month. Averaged over every keyword, including the ~80% with no overview at all, the loss is 6.6%.

The per-industry spread is the more useful finding. Click loss runs "from 1% to over 24%".

Recipe sites lose 1.07% and transactional destinations like Booking.com and Amazon under 2%. News and media lose 7.37%.

Parenting and baby content lose over 24%, and health portals up to 30.15%.

The pattern is the one every study keeps finding. The exposure is not the query volume but how completely an answer can be summarized.

Google restricted health queries after the Guardian investigation in the same window this data covers.

This is the largest keyword sample on this page and the only one measuring a non-English market. That matters when the German 59% sits next to Ahrefs' 58% and the US randomized experiment's 39.8%.

As Search Engine Journal put it, whether the gap between markets "reflects actual differences between the markets or differences in measurement methodology isn't clear from the available data".

SISTRIX is a commercial SEO analytics provider measuring its own keyword database. So this is panel data like everything else in the statistics section, not ground truth.

Sulina Connal is Google's managing director for news and books partnerships in Europe. She spoke at the FT Strategies News in the Digital Age conference in London on 11th Feb.

Google was "exploring more granular controls" for publishers, she said.

But Connal said building them is hard: "Because AI is integral to how search works, implementing the new controls is a complex engineering, huge engineering project."

The technical claim is real enough. One crawl and one index feed both ordinary ranking and the grounding of AI answers.

But Paul Bannister of Raptive publicly noted that Google's engineering capacity elsewhere makes the difficulty hard to credit.

Google shipped the control four months later, after the CMA required it.

The commercial half landed at the same event. FT chief executive Jon Slade announced that "Google and the FT have now signed a deal licensing FT journalism into a series of AI pilot projects".

Press Gazette reported Google pursuing "more focused deals" with news publishers. That follows December 2025 agreements with the Guardian and the Washington Post that included cash and extended display rights.

Connal said discussions with publishers were "creating change in our product". That is the pattern France later formalised: Google negotiating individually, market by market, rather than accepting a general rule.

The backdrop, in the same coverage, was a Chartbeat finding. Global Google Search traffic to publishers dropped by about a third in 2025.

Google disputed that by suggesting traffic had shifted toward emerging news sources. Our fuller Chartbeat breakdown published a month later.

February is the month the two tracks became visible together. Google was explaining why publishers could not opt out, while paying some of them for the content it was summarising.

The European Publishers Council filed a formal antitrust complaint with the European Commission over AI Overviews and AI Mode. It argues Google repurposes publisher content into AI answers that keep users on Google.

It is the second formal publisher complaint in the EU, following the Independent Publishers Alliance in mid-2025. It lands two months after the Commission opened its own Article 102 investigation into Google's AI search features.

Ahrefs' February 2026 update is based on December 2025 data. It measured position-1 CTR at 0.016 with an AI Overview present and 0.039 without.

That is a 58% reduction.

Its original March 2025 analysis of 300,000 keywords had put the gap at 34.5%. On this dataset, the click cost of an AIO deepened as the feature matured rather than easing.

The same day, Alphabet's Q4 2025 earnings reaffirmed AI Overviews at more than 2 billion monthly users.

The two numbers together are the AIO story in miniature. Reach is enormous and still growing, while click economics worsen for the sites the answers are built from.

Google made Gemini 3 the default model for AI Overviews globally, the biggest model upgrade in the feature's history. That came two months after routing only the hardest queries to it.

On mobile, an AI Overview can now hand off into an AI Mode conversation carrying the query and answer context. Google described it as "one fluid experience... a quick snapshot when you need it, and deeper conversation when you want it".

The handoff was the first real integration between the two surfaces. It previewed the full merger that arrived at I/O in May.

Google removed or restricted AI Overviews on a set of health queries. A Guardian investigation published 2nd Jan had documented misleading and potentially harmful health advice in AIO answers.

Google disputed the investigation's conclusions while still narrowing triggering on the affected queries.

The sequence mirrors May 2024, when the "glue on pizza" failures led to restricted triggering and new safety limits within weeks.

YMYL queries remain the highest-risk surface for AI answers. This is the second time Google has pulled AIO coverage back there under public pressure.

Google expanded AI Overview ads to 11 additional countries, on English-language queries and without an announcement. They include Australia, Canada, India, Indonesia and Singapore.

It is the first time AIO ads have run outside the United States. There they launched on mobile in October 2024 and expanded to desktop in 2025.

AI Overviews stopped being a US-only revenue surface with this change. That is useful context when Google describes AI monetization as early-stage.

Semrush's refreshed AI Overviews study, tracking more than 10 million keywords, is the best long-run prevalence series available. In January 2025, 6.49% of queries showed an AIO.

The share peaked at 24.61% in July, and it settled back to 15.69% by November 2025.

The shape matters as much as the level. Google actively tunes when AI Overviews trigger, so prevalence is not a one-way ratchet and any single-month figure dates quickly.

That volatility is why we report multiple panels side by side in the statistics section.

Ahrefs captured an AI Overview and an AI Mode response for the same query and compared the pair. The sample was 540,000 query pairs for citations and 730,000 for content.

The data is US data from September 2025, from its Brand Radar index. The two surfaces cited the same URL 13.7% of the time.

Narrowing to the top three citations in each response lifted the overlap only to 16.3%. Ahrefs reads that as "87% of the time, these two systems are pulling from completely different sources to answer the same query".

The answers themselves agreed. Response pairs scored 86% semantic similarity on average.

And 89.7% of them scored above 0.8. On that scale 1.0 means identical meaning.

Despina Gavoyannis and Xibeijia Guan put it plainly: "9 out of 10 times, AI Mode and AI Overview agreed on what to say".

One number needs keeping straight. The study also reports a Jaccard figure of 0.16, and that one measures words rather than sources.

Only 16% of the unique words in the two responses overlapped. They opened with the same first sentence 2.51% of the time.

AI Mode is not an AI Overview with more detail bolted on.

Two findings are specific to AI Overviews. They return no citations at all in 11% of responses, against 3% for AI Mode.

And they lean less on encyclopedic sources. Wikipedia supplies 18.1% of AI Overview citations against 28.9% of AI Mode's.

The study states its own limit. It compares single generations.

Ahrefs' earlier work found 45% of AI Overview citations changing between generations of the same answer. So the pool each surface draws on may overlap more than one snapshot shows.

The panel is Ahrefs' own index, so this is vendor data like the rest of the statistics section.

The data is from September 2025. That is eight months before Google merged the two surfaces into one experience at I/O 2026.

So it describes them as separate products.

SISTRIX measured the same divergence on a different panel in early 2026. It found the two citing different domains 83% of the time.

So the split held across two panels, two methods and five months.

The European Commission opened a formal antitrust investigation into Google's use of online content for AI purposes. It falls under Article 102 TFEU, the EU's abuse-of-dominance rules.

The Commission is assessing whether AI Overviews and AI Mode are built on web publishers' content "without appropriate compensation for that, and without the possibility for publishers to refuse without losing access to Google Search".

It is also asking whether Google uses YouTube content for its own AI models while restricting rival developers' access.

Opening proceedings does not prejudge the outcome. But this is the most serious regulatory threat AI Overviews has faced.

It reaches model training as well as the answers themselves.

It followed the Independent Publishers Alliance's June complaint over the same features. The European Publishers Council filed its own in February 2026.

The coalition behind the first AI Overviews complaints in Britain and Brussels began lobbying the US Department of Justice. It is the Independent Publishers Alliance, the tech-justice non-profit Foxglove and the advocacy group Movement for the Open Web.

It wants the American regulator to hear its evidence that AI Overviews has caused "substantial and irreparable harm" to independent publishers' traffic and revenue. That is per Digiday, which saw the letter.

Since its earlier filings the coalition interviewed publishers and collected data on traffic, impressions, click-through rates and revenue. The evidence runs from large publishers with millions of clicks a day down to medium and small ones.

The letter says some of them are suffering traffic drops of up to 90%. That has forced them to close publications or cut staff drastically.

Google gave Digiday the argument it has used throughout: "Our Search AI features are intentionally built to connect people to the web, and we continue to send billions of clicks to websites every day, unlike other companies."

Its spokesperson added that people's preferences and habits are changing. Google is improving Search to meet them "while continuing to drive immense value to the ecosystem".

The same coalition was working three regulators at once. It had already gone to the UK's Competition and Markets Authority and the European Commission.

The Independent Publishers Alliance's EU complaint is the one that put the "cannot opt out without vanishing from Search" argument on the record.

The comparison six months on is the useful part. The British route produced a binding remedy, the CMA's publisher opt-out order of June 2026.

No public response from the Department of Justice has been reported.

The difference is partly procedural. This is lobbying rather than a legal filing.

So unlike Chegg or Penske there is no docket to follow and no obligation on anyone to answer it.

The pressure was transatlantic rather than only American.

The UK's Professional Publishers Association was pressing the CMA with parallel evidence at the same time. Its chief executive Sajeeda Merali told Digiday that "publishers on both sides of the Atlantic are seeing the same patterns: traffic declines linked to AI Overviews, opaque shifts in Google Discover, and no meaningful transparency from Google about how these systems work".

Google launched Gemini 3 and brought it to Search the same day, which it called "the first time we've brought a Gemini model to Search on day one".

Search began routing its "most challenging questions in AI Mode and AI Overviews" to the frontier model. Faster models keep handling simpler queries, with the rollout starting in the US.

Day-one deployment in Search signaled how central AI answers had become to Google's model launches.

The routing tier turned out to be an interim step. Gemini 3 became the default AI Overviews model worldwide in late January 2026.

The UK Competition and Markets Authority confirmed strategic market status for Google in general search and search advertising.

That comes under the digital markets competition regime. It found "substantial and entrenched market power" in a market where Google handles more than 90% of UK searches.

The designation explicitly covers AI Overviews and AI Mode. The standalone Gemini assistant stayed outside it, kept under review as the market develops.

The status is not a finding of wrongdoing. But it lets the CMA impose "proportionate, targeted interventions" on how Google runs search in the UK.

That authority is what the regulator used the following June, when it ordered an AI Overviews opt-out for publishers.

Penske Media, publisher of Rolling Stone, Variety and Billboard, sued Google in federal court in the District of Columbia. The case is 1:25-cv-03192, the first lawsuit by a major US publisher aimed squarely at AI Overviews.

The complaint says AI Overviews now appear on roughly 20% of Google searches that return links to Penske sites. It blames the feature for a one-third drop in affiliate revenue from its late 2024 peak.

The suit is an antitrust claim. Penske argues Google uses its search monopoly to coerce publishers.

Publishers must let their content be republished in AI summaries and used for AI training, it says.

Google said AI Overviews "make Search more helpful and send traffic to more sites" and that it would "defend against these meritless claims".