ChatGPT Updates: Search, Ads and Commerce Tracker

OpenAI released OpenAI Signals v3.0 on 26th Oct, with a post on ChatGPT at work. The post covers "165 work activities across nearly 100 countries"; the topic shares below are in the CSV files.

The README describes "classifications of 1,000,000 consumer ChatGPT messages each month from July 2024 to September 2026", from accounts aged 18 and over. Enterprise messages, Codex and users who opted out of training are excluded.

"Seeking information" was 16.7% of messages in Jul 2024 and 20.2% in Sep 2026. The series peaked at 20.7% in Aug 2026.

In the United States the topic went from 18.0% to 20.2% over the same period. The Writing topic fell from 34.1% to 21.4% worldwide.

OpenAI's topic groups three labels: "specific_info", "purchasable_products" and "cooking_and_recipes". The first means "specific information typically found on websites, including information about well-known individuals, current events, historical events, and other facts and knowledge".

Version 3.0 restated every earlier month, raising the sample "from covering 300,000 messages per month to 1,000,000". Figures from earlier versions do not join these.

The dataset reports no clicks, links or citations, and nothing on whether a message ran ChatGPT search. It does not break out product questions on their own.

Shares are of messages, not users or sessions. The labels come from "LLM-based classifiers" with privacy noise added.

PPC Land reported the shares on 27th Oct and calls the topic the closest the data comes to search-engine use. That reading is PPC Land's, not OpenAI's.

OpenAI introduced GPT-6.1 Terra on 26th Oct. It calls it "a fast, efficient model that combines detailed visual reasoning and precise computer use alongside clearer, more natural writing".

The availability line: "Starting today, GPT-6.1 Terra is available to all Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex." The release notes add Business plans; we print both.

Neither document names regular Chat. So the model behind a search-style question in ChatGPT is unchanged.

The release notes say "some Plus and Pro users may see GPT-6.1 Terra at Medium instead of GPT-6.1 Sol at Light in the default Power presets in Work". Terra "does not support UAE inference residency".

API prices are "50% lower than those of GPT-5.6 Terra and GPT-6.1 Sol": $1 per million input tokens, $0.05 cached and $5 per million output. GPT-6.1 Sol costs $2 and $10.

One claim bears on answer quality. On error-inducing prompts, the error rate at low reasoning effort "fell from 53.6% with GPT-5.6 Terra to 16.7%", and to 4.3% at maximum effort.

The system card addendum repeats the broken search tool test. Terra "fails to acknowledge this limitation in 1.83% of cases, compared to 2.08% for GPT-6.1 Sol".

The post says the model "meets the Critical threshold in cybersecurity under our Preparedness Framework". GPT-5.6 Terra, the model it replaces, never had an entry here.

Nothing in the post, the notes or the card names a change to search, citations, the crawler or ads. The Chat line is still the one to watch.

OpenAI's help page on creating ads for ChatGPT now opens: "Starting Monday, October 26, 2026, we're introducing the multi-image ad format in Ads Manager as an open beta."

The 16th Oct advertiser email had promised "soon, a new multi-image visual format", as our earlier entry records. This is that format.

Advertisers "choose Single image or Multiple images under Ad format". Both "are available for standard campaigns with Reach, Clicks, or Conversions objectives".

A multi-image ad takes four to six square images at a 1:1 aspect ratio. "For multi-image ads, Learn more is currently the only call-to-action (CTA) option. Every image is clickable and links to that same URL."

The page is undated beyond its first line. An archived copy from 27th Jul, then titled "Create Ads for ChatGPT", has no ad format choice and no multi-image text.

OpenAI has published no post about the format. The page names no markets, no advertiser and nothing about where in a conversation the ad appears.

The creative enhancements help page changed the same night. It now says "Starting Friday, October 30, 2026", the date OpenAI's email gave, after a week of saying 29th and then 31st Oct.

Similarweb posted the two figures on X on 24th Oct 2026 (21:10 UTC). "93.7% of primary ChatGPT visitors also visited Google." "14.1% of primary Google visitors also visited ChatGPT."

The chart is titled "Audience Overlap: Google.com & ChatGPT.com" and labelled "All Traffic | Worldwide". It runs month by month from Jan 2023 to Sep 2026.

Only September's two figures are printed. The Google-side bar grows through the series while the ChatGPT-side bar stays high throughout.

Asked whether the figures were deduplicated visitors or sessions, Similarweb replied: "It's unique monthly visitors who've visited both sites."

So each percentage starts from a different base. ChatGPT.com's visitors are a small group next to Google.com's, which is why the two shares differ so much.

A visitor counts in the overlap after one visit to the other site in the month. The post says nothing about how often either group used either site.

It covers the two websites only, not the ChatGPT or Google apps. The post does not define "primary", and it gives no visitor totals.

Similarweb sells the traffic estimates behind the chart, and the figures come from its panel. Search Engine Watch covered the post on 25th Oct and made the same points about the denominators.

OpenAI announced the expansion on 21st Oct 2026. "Today, we're expanding ChatGPT Ads to 26 additional countries, extending availability to 89 countries worldwide."

The post counts advertisers for the first time. "Since we launched ads just over eight months ago, over 100,000 advertisers have run campaigns on ChatGPT."

On 23rd Sep the figure was "tens of thousands".

Fourteen of the new countries are in Latin America, "including Argentina, Chile, Colombia, and Peru". The post places them after the Mexico and Brazil launches.

The rest are "across Europe, the Middle East, Africa, and Asia". The post names Nigeria, South Africa, Serbia, Ukraine, Cambodia, Bangladesh and Pakistan.

OpenAI's Ads Manager availability page listed 89 countries when we read it on 21st Oct. Every one is marked Available.

An archived copy of 12th Oct had 63 rows. Exactly 26 were added and none removed; the full list is in the market table below.

Puerto Rico is one of the 26, so OpenAI's count treats it as a country.

Access runs "through technology partners or self-service access through Ads Manager". The tier rule is unchanged: "Ads will appear only on the Free and Go plans."

The post repeats that "advertising does not influence the answers ChatGPT provides". It repeats the 1.2 billion weekly users figure too.

Formats come next: "our next phase of growth will include the development of additional formats, objectives, buying options, and measurement capabilities".

The post gives no start date per country and no revenue figure. It does not say whether the US-only health and finance rules change.

A companion engineering post the same day explains how ChatGPT chooses an ad.

Ritendra Datta and Karthik Gopalratnam of OpenAI published an engineering post on 21st Oct 2026. It is OpenAI's first detailed account of how ChatGPT selects an ad.

The stack relies on "a custom, post-trained LLM that uses a tool call architecture to interpret a conversation's context and underlying intent, and then effectively select, rank, and deliver ads". "We call this LLM the orchestrator."

Ad sessions are stateful: "information is remembered across turns of a conversation and across multiple conversations".

Selection runs "separately from ChatGPT's answer generation, in its own service, to keep ads from influencing answers and maintain ChatGPT's neutral voice". The two services "don't share a content generation model".

The orchestrator "turns conversations into brief text summaries, which are then turned into semantic embeddings" for retrieval. It "doesn't directly decide which ad will be served, nor does it set a price"; an auction tool does.

Personalisation is a user choice. "If the user chooses to personalize ads, the orchestrator also considers available nonsensitive information about the user's past conversations".

The index for product feed ads "is designed to be aware of positive or negative product reviews". Separate "integrity models" decide when a conversation may carry an ad at all.

Timing decides whether an ad shows. "We don't delay chat responses to wait for ads to load", and "If it arrives too late, the ad is not shown for that turn."

Each ad ships with "a model-generated preamble, which is a short, neutral introduction explaining how the ad relates to the conversation". The ad displays "below ChatGPT's response, clearly labeled as sponsored".

"Returning no ad is an ordinary, successful outcome." Feedback comes from "clicks and hides from users", surveys and advertisers.

The post says "ChatGPT Ads reached $1 billion in annualized revenue run rate less than 200 days after launch". The 31st Aug disclosure gave the same run rate.

Ahead are "more dynamic ad formats", with the example of "giving users the ability to talk directly to businesses through ads".

The post names no market or date and announces no feature. It does not say how the preamble is labelled beyond "sponsored", or whether creative enhancements share a model with it.

arXiv 2610.30365 comes from New York University Abu Dhabi and is to appear at NDSS 2027. Michele Guerra, Christina Pöpper, Azza Abouzied and colleagues submitted it on 21st Oct, and we read it in full.

A data void is a search space where "relevant and reliable content is scarce". The authors built 56 of them across five topic areas and published one attacker-controlled page per void.

They then asked three products about each topic in three prompt tones: skeptical, excited and neutral. ChatGPT-5.2 ran "through its UI" on "two newly created free accounts".

Gemini 3 Flash Preview and Perplexity's Sonar Pro ran through their APIs. The authors say the API output matched the UI for those two, "but not for ChatGPT".

An exploit "succeeds if the LLM cites, uses, or lists the attacker-controlled content". The 504 responses come from one snapshot taken "25 days after launch".

"Across all conditions, attacker-content appears in LLM responses 31% of the time." Pages hosted with low visibility "never succeeded"; without them the rate is 42.3%.

Hosting mattered more than the page. On long-standing domains the success rate was 78.3% for Gemini, 51.7% for Sonar Pro and 38.4% for ChatGPT.

On new GitHub Pages domains submitted through Search Console the rates were 14.3%, 30.2% and 42.9%. Nearly every success was a citation; the models used or listed the content in under 3% of conditions.

The authors "find no evidence of large effects of prompt tone or model choice on success rates". Smaller effects "cannot be ruled out".

On optimisation for AI answers, page-level features "are overshadowed here by where the page is launched and the attacker strategy", the paper says.

What the paper does not say: when the 56-void test ran. It never states the dates, and one reference is marked as accessed on 30th Apr 2026.

It tests no Google Search surface, so nothing here applies to AI Overviews or AI Mode. Gemini ran through the API, not the app.

It tests GPT-5.2 on free accounts, not GPT-6. It reports no click or traffic figure, and no AI-visibility vendor is among the authors.

An appendix reports a separate finding from Sep and Oct 2025. ChatGPT's web client exposed its search queries, retrieved URLs and citation lists in the conversation log.

OpenAI "acknowledged these as unintended internal exposure", the authors write. Even so, they could not tell whether ChatGPT uses "a proprietary search backend, a third-party search API (e.g., Google or Bing), or a hybrid".

OpenAI added an "Update on October 20, 2026" block to its election safeguards post of 27th May. The RSS feed kept the May date, so nothing flagged the change.

"Our Democracy Works integration is now live in the U.S. across languages, bringing information about registration, deadlines, voting locations and other election requirements directly into ChatGPT."

Democracy Works "compiles this information with state and local election officials and monitors changes throughout the election period".

"Responses also include links to official election sources so people can check the details for themselves."

Results come from the AP. That integration "brings live vote counts into ChatGPT for elections in the U.S. and Brazil, after polls close".

The May post had promised both "this fall". It also said OpenAI would "continue to refine the way web search surfaces helpful information with source links".

On ads, the update repeats the ban: "We continue to prohibit political ads from running on our platform."

Not said: the ChatGPT tiers covered, how the answer looks, or when the data went live. The phrase "across languages" is not a language list.

For publishers, a voting question answered from partner data need not reach a web source. Google's election tools in AI Mode and AI Overviews use the same Democracy Works data.

So does Anthropic's promise for Claude.ai, which had not been reported live by 2nd Nov. Democracy Works announced the OpenAI partnership on its own site on 27th May.

We read both pages on 2nd Nov, the day before the midterm election.

OpenAI emailed advertisers on 20th Oct. Search Engine Roundtable reproduces the email, and OpenAI's help page documents the feature.

The email says: "Starting Friday, October 30, 2026, we'll enable creative enhancements for your new and existing self-serve campaigns."

The help page gives a different day. "Starting Thursday, October 29, 2026, creative enhancements will be enabled automatically for new and existing campaigns." We print both.

The feature uses "your existing ad and landing-page content to adapt the way your message is expressed". Enhancements "can include new or updated headlines and descriptions, image resizing, destination-URL optimization, and translation into other languages".

Variations "may be prepared in advance or created in real time based on the conversation". "When enabled, these ads can run automatically without requiring you to review or approve each version."

There is no preview. The help page's FAQ says "The experience does not provide a preview of every variation."

Advertisers can opt out before the start date or change the setting "at any time in Ads Manager". Existing account and campaign opt-outs are preserved, and campaign settings take priority over the account default.

The help page's example turns "Fresh groceries delivered" into "Groceries for quick weeknight dinners" for a conversation about weeknight dinners.

OpenAI asks advertisers to keep landing pages current on prices and availability and to comply with its ad policies. The page also says "Ads do not influence ChatGPT's answers."

Neither document names markets or says how an enhanced version is labelled. Neither says whether a translated ad can serve where the advertiser did not target, or gives a performance figure.

The help page warns that reporting changes "do not, on their own, establish the effect of creative enhancements". Schwartz's own account showed the setting already toggled on before the start date.

Mahe Chen and Nick Koudas of the University of Toronto posted arXiv 2610.24081 on 16th Oct. They wrote the 10th Oct citations paper too.

The four-page paper names no funder and thanks no company. The study measured the APIs, not the consumer apps.

ChatGPT was "gpt-5.4-mini, Responses API with the web search tool". Gemini was "gemini-3.1-flash-lite with Google Search grounding".

The authors wrote 300 English questions in ten commercial topics. A third asked for something near me, a third depended on the place and a third did not; "No question names a place".

ChatGPT answered with 19 declared locations, 15 US cities plus London, Toronto, Sydney and Dublin. Both engines also answered from five network locations, each question twice, 31,200 answers in all.

"ChatGPT follows the location an application declares, down to the city", the paper says. "It ignores the network."

Gemini "ignores the declared location and follows the country of the network". It changes most of its sources "even for questions that do not depend on place".

With a declared city, ChatGPT wrote the place into 100% of its near-me searches. For place-dependent questions the share was 48%.

Its search backend localized by itself when no query named the place.

For near-me questions in US cities, 79% of ChatGPT's answers cited a source naming the city or state. Without a declared location, it answered 63% of near-me questions by asking for a ZIP code or city.

A near-me request declared in one US city "recommends only 8% of the businesses that the same request recommends in another of our 15 cities". Repeating the request in the same city shared 32%.

Monitoring one city captured 9% of the businesses recommended across the 15 cities, five cities 36% and ten 69%.

The full-size gpt-5.4, run on 60 questions at six locations, "localizes at least as strongly as the small one".

What it does not claim: the APIs and small models "need not behave like the consumer applications". VPN exits were data-centre addresses, and the questions were in English.

No clicks or traffic were measured. The Gemini result concerns API grounding, not the Gemini app, AI Mode or AI Overviews.

"The measurements describe the engines in October 2026." The arXiv listing shows one version, submitted 16th Oct.

For local businesses, visibility in ChatGPT's answers is specific to the user's city. A tool that monitors from one place measures one market.

OpenAI's release note of 16th Oct is headed "Updates to GPT-6 Pro model in ChatGPT". There is no openai.com post.

It lists three changes. "GPT-6 Pro in the chat tab now offers" faster responses and "Improved use of plugins and tools to carry out tasks."

The third is new. "Multiple agents that can work on different parts of a task in parallel", with their progress shown in the conversation.

Users can "open completed tasks to see what each agent did". The note says nothing about how those agents search the web or cite sources.

The update "is rolling out to eligible Pro users in ChatGPT". In managed workspaces it "depends on workspace settings and rollout eligibility".

No version number or new model name appears. Free, Go, Plus and Business are not mentioned, and the note gives no figures.

The note names only "GPT-6 Pro". It does not say how that label relates to the GPT-6 Astra Pro effort covered in the 3rd Sep entry.

The same note launches ChatGPT Slides for Pro, Business and Enterprise users with Space access. That is a productivity feature outside this page's scope.

For site owners nothing changes today. Which model builds search answers is still unstated, as it was on 7th Oct.

Cloudflare and OpenAI published early results of a joint research pilot on 15th Oct. Two of the four authors work at OpenAI.

The companies "launched a research pilot in July 2026". Cloudflare's network "provides signals that can help OpenAI's search crawler discover more useful pages", and OpenAI "independently evaluates and ranks how those pages appear".

The cohort was "~1% of Cloudflare's total network". The post says the pilot "has enabled OpenAI to discover and index hundreds of millions of new web pages".

The signals use "metadata about the web for search and discovery, rather than sharing page content for model training". The post adds: "A signal about a page doesn't grant permission to access it."

Cloudflare warns that "Finding new pages can initially increase crawling as those pages are evaluated and indexed". Site owners manage access through its AI Crawl Control.

The post does not say which sites were in the cohort or whether they opted in. It names no crawler token, no traffic or referral figure and no date for a wider rollout.

The page count is the two companies' own figure and nobody outside them can check it.

For site owners, new pages on Cloudflare domains may reach ChatGPT search sooner. They may also draw more crawling from OpenAI's search crawler.

OpenAI announced the changes on X at 20:02 UTC on 15th Oct. There is no openai.com post, release note or help article yet.

The post says "ChatGPT can now use your tabs in Chrome to make answers more useful and get things done for you".

In Chat, "it can look across your open tabs and gather info". In Work, it "can control your Chrome tabs", with sending messages, filling out forms and downloading files as the examples.

It can also "organize and group" the open tabs.

A second post says the side panel now carries "the full power of ChatGPT", with the user's "entire ChatGPT history". The extension is "Available for all ChatGPT plans".

Engadget adds that OpenAI says the extension also works in Microsoft Edge, Brave and Opera.

The extension is where agentic browsing went when Atlas shut down on 9th Aug. Reading open tabs in Chat is the part that touches answers.

Nothing in the posts describes a change to sources, citations or links. OpenAI names no markets and no date beyond "today".

For site owners the change is the surface, not the feature. An assistant that reads every open tab and fills forms now ships on the free plan inside Chrome.

OpenAI's release notes for 15th Oct say: "We're rolling out GPT-Live-1.1 and GPT-Live-1.1 Quip, a mini version of the model, in ChatGPT Voice". There is no openai.com post.

The note promises "more natural accents across languages". Live also "uses a wider range of short phrases to let you know when it's searching".

New voices cover French, German, Hindi, Japanese, Korean and Mexican Spanish, "plus Bubbie in English". OpenAI adds that "The voices you see may vary by region and account."

The ChatGPT Voice help page splits the models by plan. Free gets "Limited access to GPT-Live-1.1 Quip", Go three hours of Quip, Plus three hours of GPT-Live-1.1.

Pro at $100 a month gets 15 hours and Pro at $200 "Unlimited GPT-Live-1.1", with a footnote on abuse limits. Business plans get three or 15 hours, then "1.25 credits per minute".

The help page says Live "can use web search and memory" and shows "visual results through supported widgets". It "does not support video or screen sharing".

The search angle is thin. Nothing in either page describes a change to sources, citations or links in spoken answers.

What the note does not say: any benchmark, which regions get which voices, or whether GPT-Live-1.1 is in the API.

GPT-Live-1 reached Voice on 8th Jul 2026, with a mini version for Free users, per the same notes. This tracker did not cover that launch.

This is a sighting by named observers, not an OpenAI announcement. OpenAI has published nothing on it; its release notes through 16th Oct do not mention link tracking.

Cedric Huesler of Adobe posted on LinkedIn on 14th Oct. His claim: "GPT-6 opens links with window(.)open(url, “_blank”, “noopener,noreferrer”) and drops ?utm_source=chatgpt(.)com."

The consequence, in his words: "There's no tag and no referrer, so your analytics files the visit under direct."

His check covered "~500 pages where ChatGPT drove 5%+ of visits", comparing 9th to 13th Oct with the week before. ChatGPT visits fell 31,000 a day and direct rose 26,000 a day on the same pages.

Direct rose on two in three pages that lost ChatGPT traffic. He calls the data early: "4 days of data, sampled".

In a comment he adds that shopping links still carry the tag. So the change may not cover every link type.

Search Engine Roundtable wrote it up the same morning. Barry Schwartz says OpenAI "seems to have removed or replaced" the parameters and is "not sure if it was directly intentional".

A second observer, Sébastien Billard, is quoted there. He reports a sharp drop in ChatGPT human traffic since 9th Oct and more OAI-SearchBot crawling since 7th Oct.

The timing matches GPT-6 reaching the Chat tab on 8th Oct. Nothing in that launch mentioned links or tracking.

What is not known: whether the change is deliberate, whether it covers every surface, and whether the tag returns.

Adobe sells an analytics product. The post does not say which tool produced the figures.

For site owners, the server-log signal does not change. The ChatGPT-User agent still identifies fetches, as RESONEO's study noted for thinking-mode citations.

Ahrefs' study of 12th Oct covers 66.6 million free, logged-out GPT-5.6 Luna responses from August and September. Only 14.4% of them "triggered a web search".

When ChatGPT did search, it ran about 1.6 fan-out queries per searching response. Ahrefs' reading is that free sessions lean on what the model already knows.

Of the 15 million fan-out queries, 6.8% used a site: operator, across about 116,000 domains. Among the roughly 160 most-searched domains, 75.9% of those searches went to a brand site.

Apple, IKEA, Nike, Adidas and Samsung top the site: list. Reddit is ninth, and gov.br and gov.uk sit in the top twelve.

"Official" appeared in 13.1% of English fan-out queries in September, up from 10.1% in August. The word "price" went from 6.04% to 8.88%.

Other studies found far more fan-outs per prompt, Nectiv's 7.61 for one. Ahrefs says its broad prompts and free sessions explain the gap.

A separate dataset, Brand Radar's US citations from March to September, shows Reddit falling from "about 1 in every 5" answers to "about 1 in 43". Wikipedia went from 1 in 7 to 1 in 35.

PubMed climbed from 64th to 4th among cited domains and the FDA entered at 8th. WebMD, Healthline and Investopedia fell.

Ahrefs checked the shift on about 25,000 prompts present in both months. It says every charted site moved the same way on those prompts.

What this does not measure: paid or logged-in sessions, which Ahrefs says search more, and clicks. The rank changes are citation counts, not traffic.

Ahrefs sells Brand Radar, the tool the data comes from, and the post ends with its own how-to. A vendor finding that brand sites get searched more is a finding its product tracks.

Researchers at the National University of Singapore and Xi'an Jiaotong University posted arXiv 2610.15382 on 11th Oct. It is an observational study of ChatGPT Web, not a test of any optimisation tactic.

They captured the server-sent event traces behind ChatGPT's answers. Those traces show the rewritten queries, the retrieved pages and the final citations.

The runs cover four settings: GPT-5.2, 5.3 and 5.5 Instant, and GPT-5.4 Thinking. Collection ran from 29th Mar to 29th Jul 2026.

Eighty queries came from the Search Arena dataset, half in Chinese and half in English. The 5.3 and 5.5 Instant settings ran each query 20 times, 1,600 runs apiece.

The Instant settings mostly searched once. Only 5.5% of 5.3 Instant runs and 3.8% of 5.5 Instant runs used more than one retrieval round.

GPT-5.4 Thinking issued 11.82 queries over 3.20 rounds per run on average. Its later rounds used a site: operator in 54.7% of queries, against 25.1% in the first round.

Repeat runs retrieve different pages. For 5.5 Instant, 44.5% of candidate URLs appeared in only one of the 20 runs.

Just 7.7% appeared in at least 14 runs.

Pages retrieved more often across runs were more likely to be cited. The hostname-level AUC was 0.709 for 5.5 Instant, against 0.651 for exact URLs.

Longer snippets helped a little. For 5.5 Instant, 16.9% of cited snippets ran to 500 characters or more, against 1.9% of uncited ones.

Publication recency was no signal. Its AUCs of 0.482 and 0.535 both had confidence intervals that include 0.5.

The paper then asks whether outsiders can see any of this. It compares Google results through SerpAPI with OpenAI's Responses API running the same queries with web search on.

Combined, the two recover 26.2% of the URLs 5.5 Instant cited and 28.3% for 5.3 Instant. At the hostname level the figures rise to 48.7% and 50.3%.

Responses API citations matched ChatGPT Web's cited sources most closely. Its candidate list had the broadest coverage, and Google results were "generally weaker as a standalone proxy".

A model combining the external signals was tested on 5.5 Instant. It reached AUCs of 0.779, 0.763 and 0.742 at URL, hostname and entity level.

The authors call those results "conditional on external candidate coverage".

One warning for anyone running repeat measurements. Two sessions "switched to fallback models after 14 and 16 successful target-model runs", so the paper says to "verify the serving model for each response".

What the paper does not claim: it "does not directly evaluate GEO strategies". Retrieval recurrence "does not imply that a source would be cited if it appeared more often".

It measures no clicks or traffic. Its citation label also counts supporting pages that ChatGPT does not display separately.

The paper does not say which account tier or location the runs used. The authors are at two universities and a Chinese utility's AI lab, and the paper names no vendor.

Search Engine Journal covered the paper on 18th Oct. The arXiv listing shows one version, submitted 11th Oct.

Mahe Chen and Nick Koudas of the University of Toronto posted arXiv 2610.15033 on 10th Oct. The data came from a monitoring platform the authors run.

The paper thanks ktau.ai, a Toronto firm selling placement in AI answers, for help with the data. That interest points one way.

A finding that page tweaks do little, and model memory does more, favours a vendor selling influence on that memory.

The study measured the APIs, not the consumer apps. The engines were GPT-5.4 nano with web search, Gemini 3.1 Flash Lite with Google Search grounding and Perplexity Sonar.

The 651 main questions were in English and written for a US fibre internet provider. Collection ran daily from 24th Jun to 21st Sep 2026.

Across both datasets the paper holds 158,099 answers and 468,037 citations. ChatGPT cited 2.0 pages an answer and Gemini 10.1.

Perplexity's Sonar cited 0.3, and 95.8% of its answers carried no citation at all.

Across websites, every recommended practice looked helpful. An FAQ block went with 18.1% more citations, a summary box 20.9%, an author 14.9% and schema.org markup 7.5%.

Within websites, comparing pages of the same site, "none is significant". The FAQ block's within-website figure was minus 3.4%.

The paper calls this an instance of Simpson's paradox. FAQ blocks sat on 43% of pages at the most-cited quarter of websites and 23% at the least-cited quarter.

What did matter was retrieval. "Every page ChatGPT cites is one its own search retrieved", across 22,315 cited URLs.

Retrieval explained 32% of the within-website variation in ChatGPT's 90-day citation counts, "against 2.8% for all page features".

ChatGPT consulted 26 pages per question and cited 3. It cited 43.8% of the pages it ranked first and 3.5% at ranks 11 to 20.

Those pages are not Google's. Google's top 50 lacked 97% of the pages ChatGPT retrieved, though the websites behind 51% of them did appear there.

Model memory matters too. Some 47% of ChatGPT's citations came from websites "its model names without searching".

The paper's one-line finding: "Pages are thus cited when searched; brands are also named when remembered."

What it does not claim: "The results are associations". The APIs "need not behave like their consumer applications", and Perplexity could not be replayed.

No clicks or traffic were measured, and AI Overviews and AI Mode are not in it. Its closing strategy, placing brand associations "into the prior", is a hypothesis the paper does not test.

The arXiv listing shows one version, submitted 10th Oct.

arXiv 2610.11932 (Qi Liu, Geng Hong and colleagues, Fudan University; accepted at NDSS 2027) was submitted on 8th Oct. We read the paper in full.

It audited ten AI search tools: ChatGPT, Perplexity, Grok and "Google AI" in English, plus six Chinese products including Doubao and DeepSeek. The authors queried each product's built-in search in a browser, in March 2026.

Across "17,211 citation instances over 6,356 unique source domains", the top 20 domains took between 20.5% and 70.8% of each platform's citations. Fifteen of 22 cited publication platforms "had low or medium barriers" to posting.

The authors then published 27 articles on those platforms in April 2026, one topic a made-up concept. Their finding: "8 of 10 platforms cited a fabricated concept within seven days."

After a platform first cited the planted article, Grok cited it in 30 of 32 answers on that topic. ChatGPT did in 7 of 20, Perplexity in 13 of 35 and "Google AI" in 15 of 33.

Those rates count answers after the first citation, which the paper says are "not directly comparable" to whole-window rates. Over the full seven days, 176 of 1,004 responses cited a planted article.

What the paper does not say: which Google product "Google AI" is. Its reference is ai.google, so nothing here applies to AI Overviews or AI Mode.

It names no model tier or login state for ChatGPT.

Its $14 purchase from a GEO vendor produced a citation only in Doubao, which the authors call "not a platform-wide effect estimate". The rates come from one spring 2026 campaign and, the authors say, "may drift with platform updates".

No AI-visibility vendor is among the authors. ChatGPT, Perplexity and "Google AI" "acknowledged receipt" of the authors' disclosure.

Search Engine Journal reported the paper on 19th Oct with the same figures.

OpenAI's post "GPT-6 and Intelligent UI for everyone" went up on 7th Oct. It ends the "not yet available in Chat" line from 22nd Sep.

GPT-6 "starts rolling out globally to ChatGPT Plus, Pro, Business and Enterprise tiers today in the Chat tab". Free and Go follow "starting tomorrow".

Two models sit behind the one name. Paid tiers get GPT-6 Sol and Free and Go get GPT-6 Luna, "both tuned for everyday conversation".

The help page says GPT-6 "replaces Latest" in the model menu. Existing Latest selections "move to GPT-6 at the same effort".

So the paid default moves without a user choosing it. The free default moves a day later.

Intelligent UI is the visible change. Answers can "include graphics, tappable buttons, forms, charts, and interactive experiences" inside the conversation.

ChatGPT can also start answering before it finishes thinking. For questions that need web search, OpenAI says GPT-6 Instant "starts answering 44% sooner, on average, than GPT-5.6 Instant".

That figure is OpenAI's own internal evaluation. The post gives no sample or method.

OpenAI also claims GPT-6 "makes better decisions about when to look something up". No citation or referral change is described.

Web users can turn off "Layout and visuals" in Settings. The help page says some visual elements "may still appear".

Pro effort stays on GPT-6 Astra and does not get Intelligent UI. The Work and Codex models are untouched, and GPT-5.6 Sol remains selectable.

For site owners the open question is the same as before. OpenAI has not said which model builds search answers, and this post does not either.

OpenAI's post "Building advertising for the way people use AI" went up on 5th Oct. It covers a new format, measurement and brand safety.

The audience figure comes first. The post says "ChatGPT reaches 1.2 billion people each week".

That is a ChatGPT number, not the "collective" figure from DevDay. The tile above now reads 1.2B.

The new format is visual. OpenAI says images can show "product inspiration, product usage, or the experiences they make possible".

It will be tested "during image generation in ChatGPT". Ads are "clearly labeled" and stay "separate from the image being created".

The test starts "later this month in the US with an initial group of advertisers". No other market is named.

OpenAI repeats its line on answers: "advertising does not influence the answers ChatGPT provides".

Measurement widens. Hightouch, Tealium and LiveRamp integrations send conversion data to ChatGPT Ads.

Attribution partners now include AppsFlyer, Triple Whale, Adjust, DV Rockerbox, Northbeam, Branch, Singular, Kochava, Airbridge and Tenjin. Fospha, Measured and INCRMNTAL cover wider measurement.

Incrementality is "still in its early stages", OpenAI says. Haus, Measured and WorkMagic will run geo-based experiments.

The post carries three partner figures. DV Rockerbox put WeightWatchers' attributed cost per acquisition "15.3% lower than its blended paid-search benchmark".

WorkMagic reported that 67% of Dose's incremental purchases came from new customers. Triple Whale put 93% of Portland Leather's visitors from ChatGPT Ads as new.

Those are vendor-reported figures for single advertisers, quoted by the seller of the ads. They are not a measurement of the channel.

On brand safety, "Negative Phrases are now available to qualifying advertisers". DoubleVerify and Integral Ad Science get "brand suitability evaluation pilots".

The pilots run "in a controlled testing environment, without accessing private user conversations", the post says.

For site owners the format matters less than the surface. Image generation is not a search answer, so these ads sit beside no citations.

The ad policies page was unchanged when we read it the same day. It still carries the 10th Sep stamp.

OpenAI's post "Our approach to EU text provenance rules" is dated 5th Oct. It answers the EU AI Act's requirement that generated text be "identifiable in a machine-readable way".

Three steps. API customers worldwide can opt in to watermarking "for select models" from 5th Oct, and it stays "off by default".

"Over the coming weeks", an invisible watermark goes onto "eligible ChatGPT and Codex text output in the European Union", across all plans. It is not a global default.

Detector access opens by application, "initially limited to approved researchers and expert organizations". Site owners and publishers get no tool.

The method is called textGrain. It "adds an invisible statistical signal to the model's word choices", and OpenAI plans to open-source it.

OpenAI prints the limits. At a 1% false-positive target the detector caught about 80% of 200-token passages.

For 400-token passages it caught about 95%. Both figures are for psychology content.

Mathematics scored "substantially lower". Replacing 10% of words with synonyms cut detection from about 92% to 66%, and 25% cut it to 17%.

The post lists what a watermark does not do. It "does not measure human contribution", "does not establish ownership or responsibility" and "does not identify the user".

It also says the absence of a watermark "does not prove human authorship". Short, edited or translated text may escape detection.

For publishers the practical point is the gap. EU-generated ChatGPT text will carry a signal, but nobody outside the approved group can read it.

Google has said nothing about reading it. Its published AI-content rule is about scale and value, not provenance.

Images and audio are different. OpenAI's verify tool and Content Provenance API for those stay public.

The ChatGPT release notes for 1st Oct are headed "New ways to shop in ChatGPT". Two features ship.

A "Try on" button "will appear on product listings for clothes and accessories in ChatGPT". The user takes or uploads a selfie, and ChatGPT Images generates the try-on.

The reference photo is kept for later try-ons. It can be changed or deleted under Settings, then Personalization, then Reference photos, as Search Engine Roundtable notes.

The second feature is saving: "Save any product to Favorites or create a folder to organize your finds in your ChatGPT Library."

Both are "Available in ChatGPT on mobile and web". The notes give no country list and no subscription tier.

Neither is a search change by itself. The listings are the shopping results ChatGPT already shows.

A retailer's product images now feed a try-on step as well as a click.

Roundtable's aside is that Google has offered virtual try-on in Search "countless times already". ChatGPT's version arrives as its shopping ads tooling grows.

The same day's notes also add document scanning through the ChatGPT camera. That is off this hub's beat.

The source is an email OpenAI sent to advertisers, which Barry Schwartz published at Search Engine Roundtable on 30th Sep. We found no public changelog for Ads Manager carrying the same items.

Bulk creation: "You can now create multiple product feed campaigns and ad groups from a spreadsheet in Ads Manager, with ad templates generated automatically." The path is Campaigns, Bulk uploads, Bulk sheet.

Product review status appears in the Products tab. It is there "to track products through review and understand why products have been rejected".

The Insights API gains conversions beyond the campaign goal. Advertisers "can now retrieve attributed conversion events beyond your campaign's selected goal", such as purchases while optimising for another goal.

Account, brand and legal names can now be set separately in Settings. The brand name is what shows with ads; the legal name is for billing and verification.

The email repeats the seven Asian markets of 23rd Sep. OpenAI's availability page still listed 63 self-service countries when we re-read it on 1st Oct.

Product feed campaigns are the shopping format in ChatGPT Ads. Bulk sheets are tooling for advertisers with large catalogues.

OpenAI held DevDay on 29th Sep and published a recap of "more than 20 major announcements". Four matter for site owners and marketers.

First, the audience figure. The recap says developers can launch experiences "to our collective 1.2B weekly users".

OpenAI does not say which products "collective" covers. The last ChatGPT figure was 1 billion weekly users on 31st Jul.

So the tile above stays at 1B+ until OpenAI states a ChatGPT number.

Second, dots. OpenAI says they run on GPT-6 Astra and have "their own cloud computer".

A dot can also use "their own browser, and the apps you've connected". It keeps working on a task between conversations.

Dots connect to "over 4,000 apps" through plugins. They are rolling out to Pro and Business Premium users aged 18 and over.

Pro access excludes the European Economic Area, Switzerland and the UK at launch, the release notes say.

Third, plugins get more room. Plugin extensions give a plugin "a home in the sidebar" and interactive panels beside the conversation.

Discovery changes with it. OpenAI says better "ranking and recommendations" help people find plugins in the directory and in conversations.

It gives no detail on how that ranking works. Users still choose which plugins to use and approve the access each one gets.

Fourth, a new top tier. Pro 500 costs $500 a month and carries 25 times the Plus usage allowance.

The recap also covers GPT-6.1 Sol, which has its own entry. Nothing in it names ChatGPT search, the crawler or citations.

OpenAI introduced GPT-6.1 Sol on 29th Sep. That is a week after GPT-6 Sol, the model it upgrades.

The availability line repeats last week's split. The model is in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users.

The post says it "is not yet available in Chat". So the model behind a search-style question in ChatGPT is unchanged again.

The release notes describe a staged start, "starting with Pro users". The other paid tiers follow.

API prices match GPT-6 Sol: $2 per million input tokens and $10 per million output. Cached input costs $0.10.

One claim bears on answer quality. At low reasoning effort, responses with a factual error fell from 11.4% to 7.7%.

OpenAI measures that on conversations where users had flagged an earlier model's error. It says the prompts "are not representative of typical usage".

A second test covers a broken search tool. GPT-6.1 Sol failed to tell the user in 2.1% of cases, against 4.9% for GPT-6 Sol.

GPT-6 Luna failed that test in 28.7% of cases. Luna is the model Free and Go users get in the desktop app.

Nothing in the post names a change to search, the crawler or citations. The Chat line is still the one to watch.

Lightspeed Commerce, a point-of-sale and e-commerce vendor for independent retailers, commissioned the study from the consultancy Vaer AI. Its interest runs toward showing that independents are hard to find in AI answers.

The release is dated 29th Sep. It covers "nearly half a million AI-generated shopping recommendations and more than 2.4 million links" from more than 20,000 prompts.

Head to head against a smaller retailer, the models chose the national chain 90% to 94% of the time. That figure is for answers from training data, without web search.

With live search on, large retailers still took 46% to 58% of top recommendations, per Search Engine Land's account. Small or local stores appeared about 10% of the time without search.

Specific queries widened the gap. Local retailers were the top recommendation in roughly a third of broad searches.

For specific products that fell to as little as one in ten.

The mechanism is the finding to keep. Adding "independent" to a prompt more than doubled small-store recommendations, while "local" and "near me" made far less difference.

A companion Censuswide survey of 2,000 North American consumers found 56% had used AI for shopping decisions. Half said they would shop locally more if AI made independents easier to find.

The release does not link a full report. It prints ranges across the four surfaces rather than a figure per surface.

It does not test whether chains win on stock, price or familiarity.

Graphite, a growth agency, published "The State of ChatGPT Ads" with Similarweb. The report page is dated 28th Sep.

Both authors sell into this market. Similarweb sells the ChatGPT ad tracker behind the panel data, and Graphite runs paid search for clients.

So both gain when the channel looks worth measuring and buying.

The headline figure is reach. In the US, 43.4% of ad-eligible ChatGPT users had seen at least one ad as of Sep 2026.

The UK leads at 45.7%. Australia is at 43.3%, Canada 42.9%, New Zealand 40.9%, Japan 35.9%, Brazil 34.9%, South Korea 32.9% and Mexico 27.8%.

That is a different measure from the 26% of US desktop chats Similarweb reported for June. One counts users who ever saw an ad, the other counts chats.

The click-through rates sit in the full report, which is behind a form. Search Engine Journal read it and gives the figures.

For Free and Go users, the all-time rate through 16th Sep was 0.73% in the US. Canada was at 0.75%, the UK 0.74%, Australia 0.67% and New Zealand 0.6%.

The four later markets ran higher: Brazil 1.91%, Mexico 1.55%, Japan 1.53% and South Korea 1.21%.

The rate counts clicks per ad-bearing exchange, pooled across all advertisers. An exchange is one user message and ChatGPT's reply.

So it is not comparable with one campaign's click-through rate. The markets also cover different spans, from five to 25 weeks of data.

The method page says Similarweb discloses "percentages and shares only". Panel sizes are not given.

The report covers "all nine live markets". OpenAI now counts over 60 countries, so most ad markets sit outside it.

Advertiser turnover is high. Only Jotform and Resume.io held a worldwide top-ten impression share in all 20 weeks to 14th Sep.

B2B software took 19.9% of US ad impressions between 30th Mar and 4th Aug. Retail and consumer brands came second at 10.4%.

Graphite's US ad set holds more than 95,000 ads from over 500 advertisers. It was collected between 1st Apr and 16th Aug.

The report's own first advice is caution: "Budget It Like a Test, Not a Core Channel".

OpenAI announced on 23rd Sep 2026 that ChatGPT Ads "will begin rolling out" in seven more markets. They are Indonesia, Malaysia, the Philippines, Singapore, Thailand, Vietnam and Taiwan.

The post says the expansion "follows Asia Pacific launches in Australia, New Zealand, Japan, South Korea and India". That is the first time OpenAI has listed India as a market where ads are live.

Our table below had India at "buying open" since 31st Aug, when OpenAI opened self-service there without saying when ads would serve. The post still gives no start date for India.

The headline count moves too. "With today's announcement, ChatGPT Ads are now available in over 60 countries", the post says.

On 31st Aug the figure was "over 40 countries".

OpenAI's Ads Manager availability page listed 63 countries for self-service buying when we read it on 25th Sep. Every one is marked Available.

That is eleven more than the 52 we counted on 5th Sep. The seven new Asian markets are among the additions.

Access in the new markets runs through the OpenAI Ads Solutions team, agency partners and technology partners. The named agencies are dentsu, Havas Media, Omnicom Media, Publicis Groupe and WPP.

Self-service through Ads Manager "is also available for eligible businesses". The tier rule is unchanged: ads show "only to users on the Free and Go plans".

The post quotes Shopee's Peggy Zhu, who ties the move to the Brazil launch and calls it "another step in our deepening partnership with OpenAI". Shopee is the first named partner in these markets.

OpenAI repeats the $1 billion run rate it disclosed at the end of August. It says "tens of thousands of advertisers" now run ads in ChatGPT.

The forward promise is open-ended. "Over the next few months, we will continue to expand into new markets, build new formats, optimization tools, and measurement solutions."

For search marketers the point is reach. Southeast Asia was the largest gap in the ad footprint.

Ads now serve in most markets where ChatGPT's free tier competes with Google for product queries.

OpenAI introduced GPT-6 Sol and Luna on 22nd Sep as the cheaper tiers of the GPT-6 family. Astra, launched on 3rd Sep, "continues to be our best model across the board".

The availability paragraph draws the line. The models "are available in ChatGPT Work and Codex starting today for all Plus, Pro, Business, Enterprise, and Edu users".

"Free and Go users can access GPT-6 Luna in the desktop app. These models are not yet available in Chat."

The release notes for the day repeat it: "These models are separate from the models available in Chat."

Work is the agent mode. OpenAI's help page describes Chat as "fast, conversational assistance" and Work as "an agent designed for longer, multi-step work and finished deliverables".

So the model that answers a search-style question in ChatGPT has not changed with this release. The standing caveat on this page holds.

API prices halve against the GPT-5.6 promotional rates. Sol costs $2 per million input tokens and $10 per million output; Luna $0.10 and $0.50.

OpenAI says Sol "makes about half as many mistakes as its predecessor" on its internal factuality evaluation. That evaluation is built from conversations where users flagged an error.

The rollout is gradual "throughout the day". Nothing in the post names a change to search, the crawler or citations.

What to watch is the Chat line. When the release notes put Sol or Luna in Chat, the answers users see change for every query.

The paper went to arXiv on 16th Sep.

It names the University of Zurich's Social Computing Group among its affiliations. Lucas G. Uberti-Bona Marin is first author of eight.

The team took 117 real product questions from public chat logs and ran each three times on five surfaces. Those were logged-out ChatGPT and Gemini, both APIs, and AI Overviews in Google Search, from Dutch residential IPs.

ChatGPT stated a preference in the first person in 79% of answers that recommended a product. Gemini did so in 7%, AI Overviews in 2%.

ChatGPT called a product "best" in 73% of answers, against 43% for Gemini and 48% for AI Overviews.

The recommendations did not hold still. The paper says the gap between ChatGPT and its own repeats "is as big as the difference between ChatGPT and AI overview".

The "best" framing itself changed between repeats in 27% to 40% of cases.

Sources diverged more than products.

For the same query, the ChatGPT and Gemini interfaces shared 5.4% of displayed domains on average. In 76.7% of comparisons they shared none.

The API is not the product either. ChatGPT's interface and API shared 12.0% of displayed domains on identical queries sent moments apart; Gemini's shared 14.8%.

The authors' conclusion is about auditing. Neither a single answer nor an API run "can be assumed to represent the commercial advice consumers encounter".

What it does not measure: any ad. The runs were logged out, and the paper's text does not name its collection month.

It matters here because ads now run in more than 60 countries. The paper notes ads are most likely to appear on product queries.

The behaviour it measures is the organic answer those ads sit beside.

OpenAI announced a new ad format on 16th Sep. Sponsored Agents "let people start a conversation with a business-sponsored agent after clicking an ad in ChatGPT".

The conversation is opt-in and labelled. OpenAI says it "is distinct from ChatGPT's independent answers and separate from the original conversation that the user started in ChatGPT".

Scope is narrow for now. "Sponsored Agents are now being tested with select advertisers in the United States."

For a site owner this is the first ad unit that keeps the shopper inside ChatGPT after the click. The visit to the advertiser's site comes, if at all, at the end of that agent conversation.

The rest of the post is tooling. Advertisers can "create, update, and analyze campaigns directly in ChatGPT with the Ads Manager plugin", using plain-language prompts.

Ads Manager gains suggested copy and imagery drawn from the landing page. An opt-in "text customization" setting rewrites headlines to fit the conversation and translates copy into the user's language.

Two integrations arrive the same day. HubSpot users can connect a ChatGPT Ads account and manage leads there.

US Shopify merchants get a ChatGPT Ads app in the Shopify App Store.

The Shopify app goes international "in markets where ChatGPT Ads are available starting September 23".

Search Engine Roundtable published the advertiser email sent alongside the post. It lists items the post does not carry.

Conversion optimization with impression billing is "now generally available". Attribution windows become selectable at 7, 14 or 30 days for clicks and 0 or 1 day for views.

Platform targeting splits into Android app, Android web, desktop web, iOS app and iOS web. An Event Quality Score rates Pixel and Conversions API setups from 1 to 10.

The email is the only copy of that list we have found, so we cite the write-up as one.

Ofcom published its News Report 2026 on 15th Sep. The survey covered 5,397 UK adults aged 16 and over.

"For the first time we asked about the use of AI applications and services for news, and 21% of respondents told us they had used an AI tool for news."

"The most popular of these was ChatGPT, used by 13% of respondents for news."

Read the question before using the 21%. It listed "search engines, AI Overviews or AI apps" as options for where news was seen in the last month.

So the AI figure can include people who saw news in Google's AI Overviews, not only chatbot users.

Ofcom says use "is driven primarily by younger age groups" and is higher in AB socioeconomic households. It calls AI "an evolving dimension within the news eco-system" that it will keep monitoring.

Among online intermediary brands, Google has the highest reach for news at 60%, "dominated by Google Search". Meta follows at 51%.

Among children aged 12 to 15, ChatGPT and Google Gemini both appear in the top 20 ways of getting news. TikTok leads that list at 42%.

Ofcom is the UK communications regulator, so the figure comes from a regulator's own survey rather than a vendor's. Press Gazette covered the report on 15th Sep.

The ChatGPT release notes for 14th Sep carry the change. "We're retiring automatic switching from Instant to Thinking (reasoning) for ChatGPT Plus and Pro users globally."

The picker stays. "You can still select an available option in the model picker to give ChatGPT more time to think or reason."

The "Higher intelligence" setting also goes from ChatGPT on the web for those plans. One exception is kept: "ChatGPT can still switch automatically for safety purposes."

The note says nothing about the free tier or Business plans. It points readers to the GPT-5.6 and GPT-6 Pro help page.

Why this page cares: thinking mode is where ChatGPT opens pages. RESONEO's study on this page found every page ChatGPT itself opened was fetched in thinking mode.

A Plus user who leaves the picker on Instant now gets no automatic promotion to a reasoning run. How often that changes browsing is not measured anywhere we found.

The ad policies page is stamped "Updated: September 10, 2026". The changelog adds a sixth revision in six months.

"v1.6 (September 2026): Updated to clarify our right to decline ads where they conflict with our advertising principles, business interests, or competitive position."

The sentence that changed sits in the overview. OpenAI "reserves the right to decline, restrict, or remove advertisers and to refuse to display ad content and website links for any reason, including where they conflict with our advertising principles, business interests, or competitive position".

An archived copy from 18th Aug, stamped 10th Aug, read: "OpenAI reserves the right to refuse to display ad content and website links for any reason."

The refusal right now covers advertisers as well as ad content. It names two grounds no earlier copy stated.

No archived copy of the 31st Aug version exists. We cannot show whether version 1.5 had already touched this sentence, though its changelog line mentions only legal services.

A "competitive position" ground matters for anyone selling a rival assistant, search product or model. The page does not define the term or list categories.

The earlier changelog lines are unchanged. Version 1.5, stamped 31st Aug, permitted licensed legal services in the US.

The entry date is the page's own stamp. We read the page the next morning.

OpenAI introduced ChatGPT Images 2.5 on 8th Sep. It promises "sharper details, faster generation, more precise editing, and better tools for creating and sharing".

The post says image generation latency falls "by up to 50% compared with Images 2.0". Edits are meant to change only what was asked and keep the rest.

Three product features arrive with it. Sketch lets a user draw inside ChatGPT as a reference.

Templates start from formats like posters, and a shared image can carry its prompt.

Availability is the widest OpenAI has printed this month. "Images 2.5 is available to all ChatGPT, ChatGPT Work, and Codex users across desktop, mobile, and web."

The release notes add the limits. Templates "are not yet available in Work mode" and "existing image-generation limits are unchanged".

Developers get two API models. GPT-Image-2.5 Flare is "the default choice for most applications"; GPT-Image-2.5 Sunburst trades longer generation for "tighter control across edits".

OpenAI says people create "more than 3 billion images" a week across ChatGPT Images and the API. Outputs keep C2PA metadata and an invisible watermark.

Nothing in the post names search, the crawler or citations. This entry is here because the hub carries every OpenAI model release.

OpenAI introduced GPT-6 Astra on 3rd Sep as "the world's most intelligent and aligned model".

The availability sentence is the one that matters here. Astra "is rolling out today to a limited set of organizations and over the coming days will become available to all ChatGPT Plus, Pro, Business, and Enterprise users".

The ChatGPT release notes for the same day are blunter. "Astra is not yet generally available."

Free and Go users are not named anywhere in the post. Their default stays GPT-5.6 Luna.

Usage is "included within the existing subscription allowances", with credits for sale beyond that. Pro, Business and Enterprise plans also get a higher-effort GPT-6 Astra Pro.

One line is new for a ChatGPT model launch. "Enterprise administrators can enable Astra for their workspace; access is off by default at launch."

The reason is in the safety overview. Astra is OpenAI's first model to reach the "Critical" cybersecurity level under its Preparedness Framework.

OpenAI says extra safety checks "can sometimes slow, pause, or stop legitimate work". A paused task in ChatGPT waits for the user to review it.

OpenAI describes the model as "state-of-the-art on computer use, browsing, software engineering, cybersecurity, science, and professional work". In ChatGPT it can "create, host, and share websites, web apps, and games directly from a prompt".

Nothing in the post touches search. It does not say which model answers ChatGPT search queries.

The crawler and the citation format are unchanged. So the standing caveat on this page holds.

OpenAI has named the search model once, at launch, and Astra does not change that.

API pricing is $10 per million input tokens and $50 per million output tokens. The model is also on Amazon Bedrock.

What to watch is the default. The post does not say whether Astra becomes the model paid users get without choosing it.

If it does, the answers those users see change for every query, search included.

OpenAI has published its first ads revenue figure since March. From its post of 31st Aug: "In less than 200 days after launch, ChatGPT Ads has reached $1 billion in annualized revenue run rate."

That is ten times the 100-million-dollar annualized figure an OpenAI spokesperson gave Reuters six weeks after the February launch. Both are annualized run rates, so the two are comparable.

The operational news is one sentence: "Starting later today, advertisers can purchase ChatGPT ads directly via Ads Manager across India, Europe, the Middle East, and North Africa."

That retires the promise Europe had been carrying. OpenAI announced 31 European countries on 18th Aug.

On 18th Aug, self-service was "later this summer". Day one was managed access through the ads team and six agency groups.

The post also gives the first market count OpenAI has published: "ChatGPT Ads are now available in over 40 countries through the OpenAI Ads Solutions team and our agency and technology partners."

It links that phrase to an Ads Manager availability page, which we read on 1st Sep. It lists 47 countries, every one marked Available for self-service.

Those 47 break down against what this page already tracked. Nine are the markets where ads were live.

Another 31 are the European countries announced on 18th Aug. Seven appear for the first time.

The seven are India, Algeria, Bahrain, Lebanon, Morocco, Qatar and Tunisia, which is what "India, the Middle East, and North Africa" means in country terms. OpenAI has not said when ads began serving to users in any of them.

One caution about that page. Its own access-type table still tells advertisers "in the new European markets" to get started "through the OpenAI Ads Solutions team, agency partners, or technology partners".

That is the 18th Aug arrangement, not the self-service one announced above it.

The page had been updated minutes before we read it, mid-rollout. So we take the country list as the newer statement and the prose above it as the older one.

Three platform figures are new. OpenAI puts its partner network at "more than 50 technology and measurement partners".

It says "CPC and outcome-optimized bidding now account for the majority of campaigns". And the small and medium businesses that arrived with the May Ads Manager launch "now represent a material share of the business".

Two outcome numbers arrive with no advertiser named and no method given. Read them as OpenAI's own examples rather than as measurement.

They are "an ecommerce advertiser achieving a 3x return on ad spend across its campaigns over 28 days", and a technology partner reporting "that more than 80% of ad-driven ChatGPT traffic is from new customers".

One line matters for anyone assessing the targeting. OpenAI writes that the system uses the context of the current conversation, and that "Depending on the country and a user's settings, the ads system may also use context from the user's broader ChatGPT experience".

That is the cross-conversation personalization question, stated by OpenAI itself. Its answer is that the behaviour varies by country.

It is as close as the post comes to acknowledging EU law, which it never names.

The tier rules have not moved. Ads run on Free and Go only, and every paid tier from Plus upward stays ad-free.

Update, 5th Sep: we re-read the availability page. It now lists 52 countries, every one marked Available, and its access-type table still carries the 18th Aug wording.

The five added since 1st Sep are Egypt, Iraq, Jordan, Kuwait and Oman. OpenAI has not announced them anywhere else that we have found.

On 4th Sep Search Engine Roundtable published an OpenAI advertiser email. It lists an Ads Manager plugin for ChatGPT and Codex and custom-audience changes.

The email also lists more conversion-matching fields, carousel reporting and pacing for total-budget campaigns.

OpenAI has not posted that list on its own site. The write-up is the only copy, so we cite it as one.

The ad policies page is stamped "Updated: August 31, 2026". Its changelog adds a fifth revision: "v1.5 (Aug 2026): Updated to reflect that legal services are permitted in the US."

The legal services section now permits ads for "legal advice, representation, or legal services" in the US. The condition is that the advertiser is "licensed to practice law in the jurisdiction where the ad is shown".

Immigration, personal injury, legal claims and document preparation are named as covered services. "Ads for legal services outside of the US are currently prohibited."

The overview paragraph lists the three regulated categories together now. OpenAI "may approve ads from approved advertisers within the financial services, healthcare & medicine, and legal services categories", reviewed "manually on a case-by-case basis".

The changelog also records version 1.4, dated August with no day, which "Clarified our policy stance on housing and job listings."

When we wrote our entry on health and finance advertisers in August, legal services could not advertise at all. That was the one place the placement layer and the content layer pointed in opposite directions, and it has closed.

The entry date is the page's own update stamp. The changelog says only "Aug 2026", and no archived copy exists to narrow it further.

The paper is "The Language of the Question Selects the Market", submitted to arXiv on 30th Aug. Its author is Dmitrij Żatuchin of the Estonian Entrepreneurship University of Applied Sciences and Rankfor.AI.

Rankfor.AI is a Tallinn company that sells AI-visibility measurement.

Its conclusion is that visibility has to be measured per language and per market. That is also its author's commercial case, which the paper discloses.

The 246 run records are public on Zenodo.

The design is small and controlled. There are eleven cells, with six identical runs per prompt in fresh browser contexts.

Exit IPs in Berlin, Oslo, Istanbul and Tallinn were verified per arm. The web interface was used logged out.

An API arm ran on gpt-5.6-terra with web search on and off. Everything was collected on 29th and 30th Aug 2026.

The treatment question asked for accounting software for freelancers, a category with established domestic suppliers in every country tested. In the four cells asked in the country's official language, a global supplier appeared in 1 of 24 runs.

In the two English cells on the Istanbul and Tallinn connections, a local supplier appeared in 0 of 12 runs. Berlin and Oslo were softer.

Asked in English, the German connection named Lexware, sevdesk and Accountable in three runs of six each. FreshBooks and Wave appeared in all six of those runs, and the Norwegian connection named Fiken once.

Language and location act on different things. Holding the language at Turkish and moving the exit IP to Berlin returned German suppliers written in Turkish.

Russian asked from Tallinn named an Estonian supplier in four runs of six and a global one in all six. That sits between the Estonian and English cells.

The model knew where the buyer was in every one of those runs. Two of the six English answers from Estonia even noted that Estonian VAT and e-invoicing argue for a local product.

They then named five global ones and no Estonian one.

The stability result constrains everyone who measures this. Across six identical runs the top recommendation changed on four of six prompts.

The rate was the same in the browser and in the API, with search on or off. So a single screenshot of a ChatGPT recommendation is one draw from a distribution.

The negative control disconfirmed the author's own expectation. A project-management question was chosen because it was expected to lack domestic incumbents.

It showed no language effect at all. Yet the category does have domestic suppliers, and none was named in any language.

That points at national regulation rather than national supply as the thing that switches local suppliers on.

The limits are the paper's own. Six runs per cell is a lower bound, and there is one model and one prompt per category.

Time of day was uncontrolled, and the model behind the logged-out interface is undisclosed. One asymmetry is unexplained: Germany and Norway surface a local name in English when Estonia and Turkey never do.

It is a single-author preprint and not peer reviewed.

The practical reading is for anyone selling into a non-English market. English-language measurement of that market may show nothing at all.

A domestic supplier can be absent from the English answer and present in every local-language one from the same city.

OpenAI has announced the biggest expansion of its advertising pilot so far. From its own post of 18th Aug: "Next week, ChatGPT Ads will expand to 31 European countries, including Germany, France, Spain, Italy, Sweden, Norway, Denmark, the Netherlands, and Austria."

Read that as announced rather than live. OpenAI names nine of the 31 countries and gives no country-by-country dates.

There is no list of the other 22, and "next week" is the only timing in the post.

Our ads tracker therefore carries Europe as announced until launches are confirmed market by market. That is how we handled Brazil and Mexico for the three months between their announcement and their launch.

The buying route is the practical catch for a European advertiser. Access at launch is "through the OpenAI Ads Solutions team, agency partners, and technology partners", and OpenAI says "self-service access through Ads Manager will follow later this summer".

US advertisers have had the self-serve console since May, and UK advertisers since June. It is not part of day one in Europe.

Trade reporting has since supplied both the date and the names the post leaves out.

Digiday reported on 19th Aug that OpenAI "will begin serving ads to people using ChatGPT across 31 European markets on Monday (August 24)". From that date, advertisers in the new markets can buy through Publicis, Omnicom, WPP, Havas, Dentsu and MediaPlus.

That answers who the "agency partners" are, and it is OpenAI talking rather than an inference. The same piece quotes Dave Dugan, the company's vice president of ads.

He said "we're especially excited to bring ChatGPT Ads to Europe as we hit the six month milestone for our global ads business".

Digiday also notes the US self-serve tool launched in the spring with a 50,000-dollar minimum spend, since dropped. That is the shape the European self-serve tier is likely to arrive in.

OpenAI's own post was never updated, and no report of serving appeared in the following fortnight. That silence was expected: an expansion sold through the ads team and six agency groups produces no public signal.

The confirmation arrived on 31st Aug in a different post. It put ChatGPT Ads at "over 40 countries" and opened self-service buying across Europe the same day.

Our tracker moved Europe to live at that point.

Everything else about the format holds. Ads run only for people on the Free and Go plans, and Plus, Pro and Enterprise stay ad-free.

That is the same tier split as every existing market.

What the post does not address is the reason Europe was absent in the first place. Brazil and Mexico completed the original list on 11th Aug.

At that point the trade press had a working explanation. The General Data Protection Regulation and the Digital Markets Act made continental Europe harder.

OpenAI's announcement mentions neither. It says nothing about whether ad personalization or the automatic advanced matching default behave differently under EU law.

Nor does it say whether the data flowing to advertisers does.

For anyone planning European campaigns, that is the open question. It will be answered by the ad policy and the Ads Manager rather than by this post.

Two figures from the same announcement are new.

OpenAI puts its own history at "In February, we began testing ads in ChatGPT with a pilot in the United States. Over the past six months, we have expanded to eight additional markets", and says "tens of thousands of marketers have now advertised on ChatGPT", the first order-of-magnitude count it has given for advertisers.

It also lists what the platform has gained since February. Bidding now goes beyond CPM and CPC to support conversion optimization, and there is geo-targeting and custom audiences.

Measurement goes past clicks, through an OpenAI Pixel, a Conversions API and third-party measurement integrations.

This is a sighting rather than an announcement, and the two observations are nine days apart.

On 18th Aug, product-sighting tracker Radu Oncescu posted screenshots on X. He wrote: "OpenAI has now added a new "Shopping" tab for users based in the EU."

His post adds two observations: no in-app checkout, and visual search available.

On 27th Aug, Search Engine Roundtable reported the Shopping tab pinned to the left sidebar. It opens "a landing page and a dedicated search page focused on shopping", and the interface itself labels the tab "new".

OpenAI has published nothing about it. So there is no stated scope and no market list beyond the EU sighting.

There is also no word on whether every account gets it.

The reason it belongs on this page is the arc it completes. When OpenAI deprioritized Instant Checkout in March, its stated redirection was "we focus our efforts on product discovery".

A persistent shopping surface with its own search page is that focus made visible in the interface.

Oncescu's "no in-app checkout" note matches the March direction: merchants keep their own checkout, ChatGPT keeps the discovery.

For merchants the practical reading is simple. Shopping inside ChatGPT now has a dedicated entry point rather than living only inside conversations.

That raises the value of the product feeds OpenAI added to the Ads Manager without announcement in June.

RESONEO captured and dissected 1,200 ChatGPT answers, 88,000 search results and 26,900 distinct pages in July 2026. It then replayed the same prompts six weeks later and published the analysis on Search Engine Land.

Its author, RESONEO co-founder Olivier de Segonzac, discloses the commercial interest in the piece. The finding it organizes everything around is that ChatGPT grounds answers in three layers with separate rules.

There is a discovery index that finds pages, and a cache holding full copies of pages it has already fetched. And there is a small set of pages it opens live.

The layer that decides most outcomes is the cheapest one. Instant mode is the fast mode free users get.

In it, no page is opened at all in 93% of the answers RESONEO captured. The model sees only a page title plus a snippet of roughly 200 characters.

That snippet is anchored on the H1 and whatever visible text sits around it. That can mean a category label, a byline, a publication date or a table of contents.

The study measured one snippet that was entirely table of contents.

It is cut at indexing time and does not change with the question asked. The meta description is ignored entirely.

For the pipelines fed by scraped Google results, the meta description is used about one time in three. So the same tag does different work depending on which layer answers.

More than 90% of ChatGPT users are on the free tier. So this is the layer that governs how often a site shows up in most conversations.

The funnel numbers put the value of a page open in perspective. Across the corpus, 61,332 URLs reached the sources sidebar and 5,032 became the lead source behind a citation.

Only 759 pages were actually opened, all of them in thinking mode.

An opened page ended up cited 74% of the time. A page that was retrieved but never opened was cited 7% of the time.

The cache is where the site-owner consequences are sharpest. It is keyed by URL and shared across all users and tiers, so one person's fetch serves everybody else.

It refreshes on a stale-while-revalidate pattern, and a copy counts as fresh for about 30 minutes. After that, the next user still gets the stale copy immediately while a background fetch updates it.

That makes the recrawl cadence of your pages a function of how often ChatGPT users ask about them. The study reports copies served more than 90 days after the fetch.

Per the analysis, a Cache-Control: no-store header is ignored and so is noindex.

The HTML-to-Markdown conversion strips scripts, iframes and JSON-LD, so structured data does not reach the model by this path. Image alt text survives, and text hidden by CSS is still extracted.

Two hard limits frame every fetch. The robot does not execute JavaScript.

And a page over 4 MB is rejected outright with an HTTP 400 rather than truncated.

There is a measurement gap in this that anyone reporting on ChatGPT exposure should know about. The utm_source=chatgpt.com parameter is appended to the links displayed to users, so it counts outbound clicks.

But RESONEO found the pages ChatGPT opens itself in thinking mode carry no such parameter in the citations shown.

The pages with the 74% citation rate are therefore the ones leaving no UTM trace. That makes the ChatGPT-User user agent in server logs the better signal.

The study also reports an API parameter that returns the crawl date of the stored copy of a URL. That would let a site owner see when a ChatGPT user last caused a page to be read.

None of this is confirmed by OpenAI. The pipeline names come from a field in the raw data stream ChatGPT sends to the browser.

RESONEO's own Chrome extension captured that field.

That field disappeared on 21st Jul. The pipelines are now identified by formatting signatures, with a classifier the authors put at about 98% accuracy.

OpenAI documents none of the names and speaks publicly only of "third-party search providers".

Treat the architecture as a careful outside reconstruction. Some mechanics are checkable on your own server: the fetch behaviour, the 4 MB limit and the ChatGPT-User visits.

Those are the parts you can verify yourself.

The rule most people still quote is the one OpenAI published in February: ads are "not eligible to appear near sensitive or regulated topics like health, mental health or politics". That stopped being the whole picture in April, and OpenAI never went back to update the February post.

The current position sits in OpenAI's ad policies, whose changelog dates the shift to version 1.1 in April 2026: "Medical, legal, and financial advice contexts are no longer categorically blocked from ads by default. Sensitive conversations and other prohibited contexts remain ineligible for ads."

Two separate layers decide whether an ad runs, and only one of them loosened.

Placement governs which conversations can carry an ad at all, and that is where the regulated-advice block was lifted. Content governs who may advertise, and there the categories are still restricted.

In the US, OpenAI says it may approve health advertisers case by case. The categories are consumer medical devices and wearables, dental services, dietary supplements, disease-awareness campaigns and health insurance.

Also listed are hospitals and urgent care, medical testing, minimally invasive cosmetic procedures, and vision products and services.

Financial advertisers can be approved for auto loans, credit cards, credit monitoring, deposit accounts and financial planning. Also listed are insurance, investment services, mortgages, personal loans and payment services.

Proof of licensure may be required, and both categories are "generally prohibited" outside the US.

What did not change is just as important. Political content is still disallowed.

Legal services still could not advertise at all, even though legal-advice conversations became eligible to carry ads. That was the one place the two layers pointed in opposite directions.

Sensitive user contexts, including suicide and self-harm, remain off limits.

Search Engine Journal set the change out on 12th Aug and makes the practical point. OpenAI has revised these policies four times in five months.

So re-check any campaign assumption more than a few weeks old.

OpenAI added a dated update to "Testing ads in ChatGPT", the post it has used as the running record of this rollout since February: "ChatGPT Ads has now launched in the United Kingdom, Mexico, Brazil, Japan, and South Korea. We're continuing to expand to more markets this year."

Three of those five were already live: the UK on 6th Jun, and Japan and South Korea on 22nd Jun. The new arrivals are Brazil and Mexico.

Benji Shomair, OpenAI's VP of Monetization, named them on 7th May. They were pending for three months after that.

PPC Land reported on 7th Aug that OpenAI had told advertisers the two markets would launch "in the coming week". It worked out that completing the list would take the pilot to nine countries.

Those are the United States, Canada, Australia, New Zealand, the United Kingdom, Japan, South Korea, Brazil and Mexico.

Search Engine Roundtable's account of the same advertiser email still had both listed as coming soon.

They are the pilot's first Latin American markets. They are also its first outside the wealthy English-speaking bloc plus Japan and South Korea.

Continental Europe was still absent at this point. PPC Land tied that to the General Data Protection Regulation and the Digital Markets Act applying there directly.

OpenAI then announced 31 European countries a week later, on 18th Aug.

The tier rules are unchanged. Ads run on Free and Go only, and every paid tier from Plus upward stays ad-free.

The market-by-market picture is in the ads tracker below.

This is a sighting rather than an announcement, and it is written as one. OpenAI has published nothing about it and has not been asked on the record.

There is no measurement of its effect yet.

Glenn Gabe posted on X that he had "noticed ChatGPT not providing a visible link to view sources. You have to click the overflow menu and then click view sources."

His reading of the consequence is the reason this is on the page. He says it "can definitely decrease visibility and clicks for citations that show up in the right-side panel".

He said he was seeing it "across both paid and free accounts" that day. The visible link still appeared occasionally, so it is intermittent rather than uniform.

Barry Schwartz covered it at Search Engine Roundtable two days later, with screenshots of both states. He described it as OpenAI "testing or rolling out" a change.

The Sources button used to sit in the footer of the response. Now it sometimes appears inside the three-dot "More actions" menu instead.

Opening it still produces the same right-hand panel of cited sources.

What is at stake is one click, and one click is a lot. Citations are already scarce on this surface: Similarweb put web citations in 6.8% of ChatGPT answers in May 2026.

If the route to them moves behind a menu, fewer readers ever reach a publisher's link. That holds whatever the citation rate does.

Gabe asks whether this relates to ads. Ads ran on the free and Go tiers in nine markets at the time.

That is his speculation and it should be read as such. OpenAI has given no reason for the change, and nobody has shown a connection between the two.

The precedent on this page argues for taking an unannounced interface change seriously. In May 2026 OpenAI altered how prominently it presented links and said nothing about it.

Similarweb later measured referrals rising 157.7% in a week.

Link presentation on this surface has moved traffic by multiples before, in both directions, with no post to mark it.

So the useful response is measurement rather than reaction. Anyone tracking ChatGPT referrals has a dated hypothesis to test against their own analytics.

That is the only way this gets settled, because OpenAI does not document the interface. It has not said whether this is a test, a staged rollout or already reversed.

Atlas launched on 21st Oct 2025, macOS-first with Windows, iOS and Android promised. OpenAI never confirmed those ports shipped.

The shutdown was announced on 9th Jul and took effect on 9th Aug. That is 292 days after launch by our count of the two dates.

OpenAI warned users the browser "may no longer open, browse, or support agentic workflows after that date", per its help-center guidance reported by 9to5Mac. Bookmarks did not transfer automatically.

The features outlive the product. Agentic browsing moves into the ChatGPT desktop app, which gains tabs, logins, downloads and extensions.

It also moves into a ChatGPT extension that works inside Chrome's sidebar.

TechCrunch tied the decision to Applications CEO Fidji Simo telling teams to cut back on "side quests".

The pattern is industry-wide: Microsoft retired Copilot Mode in Edge three months earlier. The AI-browser thesis of 2025 has so far ended in features folded back into existing surfaces.

OpenAI sent ChatGPT Ads advertisers a product update in the week beginning 3rd Aug. It sets out nine changes, and one of them carries a hard date.

Automatic advanced matching uses hashed customer details collected from website forms to attribute more conversions to ads. It is already the default for newly created web pixels.

On 17th Aug it switches on for pixels that already exist.

Opting out is the advertiser's job. The setting sits under Tools, then Conversions, then Data Source, then Edit pixel.

No action is required to accept the change. That is the familiar opt-out-with-a-deadline shape platforms use to move an installed base onto a new default.

The compliance question is not settled by the toggle. Advertisers under consent rules still need a lawful basis to send hashed form data to a third party.

ChatGPT ads did not run in continental Europe when this shipped. UK advertisers already operated under an equivalent consent regime.

Europe has since gone live with self-serve buying. OpenAI has still not said whether this default behaves differently under EU law.

The nearest it has come is a line in the 31st Aug post. It says the ads system may use context from a user's broader ChatGPT experience "Depending on the country and a user's settings".

Two other items in the same update matter for retailers. Conversion-optimised cost-per-click bidding began rolling out on 5th Jun for other campaign types.

It is now in beta for product feed campaigns. It supports click-billed campaigns with fixed or manual bidding and a positive conversion bid cap.

And OpenAI confirmed it has started testing a multi-product carousel for product feed ads. It shows several products from one advertiser in a single unit.

The rest of the list is plumbing with real consequences. There are dynamic URL macros for campaign, ad group, ad and account identifiers, plus Triple Whale and Hightouch measurement integrations.

Pixel validation diagnostics now say why an event was dropped.

OpenAI did not publish any of it. Barry Schwartz posted the contents of the email on 5th Aug and wrote it up two days later.

PPC Land's account is the fullest.

That is this beat's recurring shape. The changes that move money reach advertisers by email and reach everyone else through the trade press.

OpenAI announced that GPT-5.6 Luna replaces GPT-5.5 as the default for Free and Go users. Free text chats become unlimited starting the week of 10th Aug.

OpenAI's internal evaluations claim factual errors are 62% less common with Luna than with GPT-5.5-Instant. For its Sol sibling the claim is 68%.

Those are the vendor's own numbers, so read them as claims rather than benchmarks.

The same day OpenAI published "From asking to doing", a global usage report on how people use ChatGPT.

The strategic reading is simple. The free tier is now ad-supported in seven markets.

Removing the message ceiling maximizes the surface those ads run on.

Chat model defaults are separate from search. OpenAI has never named the model behind ChatGPT search answers since launch, and we do not guess.

OpenAI disclosed the milestone in its post "Building abundant intelligence", and TechCrunch's later coverage describes ChatGPT as having "recently crossed the 1 billion weekly user mark".

The disclosed series, each figure from OpenAI, runs as follows. It was 200 million weekly users in August 2024 and 400 million in February 2025.

Then 700 million in August 2025 and 800 million at DevDay that October. Then 900 million in February 2026 alongside 50 million paying subscribers, and now a billion.

Two reading rules keep the number honest. It counts weekly actives, so it is not comparable with monthly-user figures like Google's 2.5 billion for AI Overviews.

And user counts are not search counts. OpenAI's only search-volume disclosure remains the 1 billion web searches per week it stated in April 2025.

Sam Altman also said in February that India alone accounts for 100 million weekly users, OpenAI's largest student base.

iPullRank published a zero-click analysis built on the Datos clickstream panel. It covers 13.1 billion search events and 10.2 billion sessions from 9.1 million opted-in users.

The window ran from October 2024 to December 2025 with continuous daily coverage. Ranked by the sites people click to from Google Search, ChatGPT came sixth.

Only YouTube, Google's own pages, Reddit, Facebook and Wikipedia ranked above it.

The finding that changes how the ChatGPT-versus-Google story reads is the paid share. About 4.75% of ChatGPT's Google traffic arrives through ads.

That compares with 0.2% for YouTube, zero for Wikipedia and under 2% for Amazon.

iPullRank's reading: "OpenAI isn't routing around Google. It's paying Google for traffic".

The authors state their own limit plainly. They have no cost-per-click data and capture only the paid clicks that landed on ChatGPT, not everything OpenAI bid on.

So this is a share of traffic and not a spend estimate.

The study also scores each big destination on how often Google answers the query itself instead of sending the click. OpenAI has the lowest interception rate of the top destinations at 11.1%, against 30.5% for Wikipedia and 33.4% for IMDB.

The reason is intent. Google can answer a definition or a cast list on the results page.

But it cannot be the thing a person typed "chatgpt" to reach.

Category-wide, AI tools drew about 720,000 navigational Google searches a week inside the panel. DeepSeek pulled navigational searches at close to ChatGPT's click-through reliability.

The authors' conclusion is that AI tools have become a destination on Google rather than a replacement for it.

Know what this does not measure before quoting it as an AI-versus-search figure. It counts clicks leaving Google, not ChatGPT's own traffic, query volume or user numbers.

None of it is attributed to AI Overviews or any other single Google feature.

The study's window-wide zero-click rate is 46.96% of 10.2 billion searches. That is a Google-wide measure, and it rose about 2.6 percentage points across the fifteen months.

It is also an opted-in panel rather than a census. So the levels are directional, and the comparisons between destinations are the useful part.

The 38-page report, covered in detail by PPC Land, is the closest thing this beat has to an annual census.

The citation numbers first. The share of ChatGPT answers containing web citations rose from roughly 1.3% in June 2025 to 6.8% in May 2026.

That is answers, not queries, and it varies hugely by industry: travel 22.6%, retail 13.5%, finance 8.0%, education 4.8%.

Where the clicks land: 58.8% of ChatGPT referral traffic goes to homepages. That is roughly double the share before the May link change.

Yet 65% of cited URLs sit two or three folders deep in a site. ChatGPT cites deep pages and sends people to brands.

Ads appeared in 26% of US desktop ChatGPT chats by June 2026, five months after launch.

And competition is real. ChatGPT's share of visits to standalone generative-AI sites was above 80% in late 2024.

By May 2026 it was just above 50%, even as its US app audience grew 87% year over year.

Panel-based measurement, as always, so treat levels as directional and trends as the signal.

Japan and South Korea became the first non-English markets with live ChatGPT ads, per PPC Land.

The news came once again through a LinkedIn post rather than a press release. The poster was Benji Shomair, OpenAI's VP of Monetization.

The market ledger now reads: US live 9th Feb, then Canada, Australia and New Zealand confirmed 26th Mar. The UK followed on 6th Jun, and Japan and South Korea on 22nd Jun.

Brazil and Mexico were also named in the 7th May announcement. They followed on 11th Aug, completing the list at nine markets.

The UK also gained access to the self-serve Ads Manager beta, which US businesses have had since 5th May.

The full market, tier and exclusion picture lives in the ads tracker below.

Two academics published a set of experiments on brand bias in LLM product recommendations. It is the clearest account we have read of how an unknown brand actually competes inside an assistant's answer.

Xi Chu of Trine University and YuPeng Hou of Texas A&M posted it to arXiv on 16th Jun. It is a preprint and carries no peer-review statement.

The design is ten skincare products at a time: one real brand such as CeraVe, and nine fictional names. The fictional names went through generation, web-search deduplication and a recognition screen.

Three models were tested, GPT-4o-mini, Claude Sonnet and Gemini 3 Flash, in English and Chinese.

The first result is the one to sit with. All ten products carried identical ratings, prices, review counts and ingredient text.

The real brand was recommended in 100% of 670 valid trials. Not one fictional brand was ever chosen.

That dominance turns out to be brittle rather than absolute. Give the fictional brand better specifications and the models choose it about 96% of the time.

The threshold is very low. A 0.075-star rating advantage was enough to break the tie, and the paper's abstract rounds this to "less than a +0.1-star" advantage.

Decomposing the variance makes the mechanism plain. Once any distinguishing information exists at all, product parameters explain 82.4% of ranking variance.

List position explains 6.5% and brand identity 1.2%.

The authors call this a conditional monopoly. The barrier facing a new brand is not the incumbent's brand equity.

It is the absence of anything that tells the model the two products differ.

The second experiment is where marketing enters. Five kinds of persuasive language were added to the fictional brand's description, with every specification held equal.

The results split cleanly in two.

Authority language broke the incumbent's hold 73.3% of the time and social proof 50.7%, against a baseline near 4%. Anchoring, scarcity and loss aversion moved almost nothing, at 12.9%, 11.7% and 9.6%.

So the models largely ignore sales talk and treat credentials as evidence. The authority prompts included fabricated citations of the form "peer-reviewed clinical trial, n=120, p<0.01".

The authors put a price on it. Authority language bought the same advantage as a real improvement of +0.17 rating points.

That equals a 15.3% price cut, or 1.9 times the review count.

The models differ, and Claude differs most. Authority worked on Claude 55% of the time against 99% on Gemini.

Stacking authority with social proof cut Claude's rate to 21.2% while raising GPT's to 91.2%.

The third experiment is the one to read twice. It runs the same optimization across more and more of the nine challengers.

A single optimized challenger drops the incumbent's survival rate from 100% to 19.8%. With all nine doing it, the incumbent recovers to 93.8%, because the signal no longer distinguishes anybody.

Per-brand payoff falls from +0.802 for the first mover to +0.007 under universal adoption, in the authors' own payoff proxy. Across 4,745 trials, brands that did not optimize while their competitors did received zero recommendations.

The incentive and the collective result point in opposite directions. Optimizing always beats not optimizing, and everyone optimizing returns the field to roughly where it started.

One probe matters for search specifically, and the authors flag it as directional only. They added a minimal retrieval step, and brand familiarity stopped helping.

The real brand ranked 8.50 out of 10 on embedding similarity, and its survival rate fell to 0.0%.

Authority language raised retrieval similarity as well, so it helped at both the retrieval and the generation stage.

The authors also separate three tiers of this content. They are real certifications and published trials, vague phrasing such as "clinically proven", and invented studies and endorsements.

What they tested is the third, as an upper limit rather than as advice.

What the study does not measure is the important part. No search engine is called anywhere in it.

So there is no AI Overviews, no AI Mode and no ChatGPT search result in any of these numbers.

The retrieval probe is a single embedding model with no reranking and no query rewriting. The authors say they do not claim it generalises to commercial systems.

The main experiments are also one product category, one fixed user persona and three closed models. The tie-breaking result was replicated on USB cables and AA batteries.

But the language and competition results were never tested outside skincare.

For those reasons no figure above is restated as a statistic anywhere on this site.

Read this one with its authorship in view. It is an arXiv preprint by Keisuke Watanabe and Kazuki Nakayashiki of Glasp.

Glasp is the company whose own site the study measures, and the paper is not peer reviewed.

Unlike a vendor case study, it publishes a control group and a design anyone can argue with.

That design is the reason to read it. In January 2026 Glasp applied a bundle of answer-engine optimizations to one identifiable part of its domain.

That part is a corpus of hundreds of thousands of YouTube question-and-answer pages. Every other path was left alone.

The untreated remainder shares the domain, its authority, the analytics configuration and the bot-filtering rules. So it absorbs the same platform growth.

Outcomes come from the Google Analytics 4 Data API and Search Console rather than from third-party estimators.

Start with the number a case study would have led with. Read trough to peak on the daily series, ChatGPT referrals rose 38.2 times.

That runs from the lowest day in early January to a single-day peak on 5th May 2026.

The authors refuse to use it. The peak is a transient spike, and the sustained late-May level is under half of it.

On monthly aggregates the total grew 5.7 times.

Then the control does its work. Treated pages grew 6.1 times between January and May 2026.

Untreated pages on the same domain grew 3.5 times with no optimization at all.

Comparing the two growth rates directly gives 1.75 times on sessions and 2.48 times on engaged sessions.

The headline estimate comes from an interrupted time series on the weekly ratio of treated to control traffic. That covers 47 weeks, split 26 before and 21 after.

The intervention lines up with a discrete level increase of 1.82 times (95% confidence interval 1.31 to 2.54, p = 0.001).

Two details qualify it. The treated corpus was already gaining ChatGPT-referral share before anything was changed, at roughly 11% a month.

That pre-trend is significant on its own.

And the change in slope is not significant, so the intervention moved the level rather than the trajectory.

The paper then tries to break its own result, and that is the part to copy. The authors re-estimated the level break at every candidate date inside the pre-period.

That produced spurious jumps nearly as large as the observed one, giving a permutation p-value of 0.16.

The authors' own verdict is that the effect is "suggestive, not conclusive". Across alternative specifications the estimate stays in a 1.8 to 2.3 times band.

They warn explicitly against reading their end-state figure of 4.85 times as the effect. It folds the pre-existing trend back in.

What was actually done is as important as the size, and it was four things at once.

Duplicate URLs serving the same video were consolidated to one canonical address. AI-bot requests for pages that did not exist were mined out of the 404 logs as demand signals.

Those signals were used to generate new pages.

Titles were rewritten into question form. Lead summaries became standalone two-to-three-sentence answers, prioritized by how often bots asked for each page.

And a rule the authors call SEO Guard locked any page earning meaningful Google clicks out of the rewrite queue. It also unpublished pages with neither organic nor AI interest.

Because the four shipped together, no single tactic gets an effect of its own. And because the 404 mining added pages, the estimate mixes per-page improvement with corpus expansion.

On the question every publisher asks before rewriting anything, the answer is reassuring and narrow. Organic clicks to the treated pages fell about 25% from their second-half-2025 level.

The site as a whole fell about 20%. Impressions stayed near baseline and nothing was deindexed.

The authors also found the opposite of the claim that AI citations lift organic search. Treated organic was flat to declining, not rising.

The limits are stated plainly in the paper and they are why this is an entry and not a benchmark.

It is one domain and one engine, counting sessions whose source contains chatgpt. Every other AI engine together was under 1% of the site's AI-referred traffic, and excluded.

No randomization, an approximate intervention date, and treated and control pages that differ in intent rather than being matched.

A bot-filtering change in mid-March moved session composition. Absolute counts are withheld for confidentiality, so every figure is a ratio or an index.

A result from one site is not a rate anyone can apply to their own. That is why no figure from it is restated as a statistic on this page.

E-commerce analyst Juozas Kaziukenas disclosed on LinkedIn that the Ads Manager had gained a Feeds section, reported by PPC Land.

It takes up to 1 million SKUs per advertiser. New e-commerce partners are asked to submit a 100-product sample first.

The interface showed feed entries dating back to 1st Mar 2026. OpenAI has never published a launch announcement for the feature.

The story underneath is the commerce pivot completing. Product feeds were introduced in September 2025 to power Instant Checkout.

After that was deprioritized in March, the same feed pipes migrated into ad targeting.

For merchants, the practical state as of August 2026 is this. The Agentic Commerce Protocol spec remains live at developers.openai.com/commerce, and feed management happens inside ads.openai.com.

Similarweb's measurement, published 25th May, opens plainly: "On May 7th, 2026, ChatGPT began surfacing more prominent links to brands directly within its answers."

Comparing the week before against the week after, total ChatGPT referrals rose 157.7% and homepage referrals 354.7%. The homepage share of all ChatGPT referrals had been roughly 26 to 32 percent before the change.

It jumped to around 60 percent after, and held there.

Indexed to the end of April, homepage referrals grew nearly sevenfold while total referrals roughly tripled. Engagement quality per visit improved rather than degraded.

OpenAI never announced any of it. That is this beat's defining hazard: the most SEO-relevant ChatGPT changes tend to be discovered by measurement.

A separate OpenAI event landed the same day, the five-market ads expansion announcement, and the two are unrelated.

By the July Similarweb report, homepages were still taking 58.8% of referral traffic, so the change stuck.

OpenAI's post "New ways to buy ChatGPT ads" opened the self-serve beta to all US businesses at ads.openai.com. The spend minimum was removed.

CPC bidding was introduced at a recommended 3 to 5 dollars per click, alongside the default 60-dollar CPM.

Named technology partners: Criteo, Kargo, Adobe, Pacvue and StackAdapt.

Criteo was the first ad tech partner, since March. It said more than 1,000 brands were already live through its API connection.

It put click-through rates at about three times comparable formats.

Those performance figures are Criteo's own.

PPC Land had tracked a quieter rollout in mid-April with a 50,000-dollar minimum. So the May post formalized and loosened an existing pipeline.

Launch pricing in February had been reported at a 60-dollar CPM with a 200,000-dollar commitment. OpenAI itself never published those figures.

Conversion-optimized campaigns followed on 5th Jun for accounts with a conversion event flowing.

OpenAI's own words, quoted by TechCrunch: "We've found that the initial version of Instant Checkout did not offer the level of flexibility that we aspire to provide, so we're allowing merchants to use their own checkout experiences while we focus our efforts on product discovery."

The numbers behind the pivot had surfaced the week before. Products sold directly inside ChatGPT converted three times lower than products routing users to Walmart's own site.

Walmart confirmed it, per CNBC, with Wired first to report it.

Forrester analyst Emily Pfeiffer put the live Shopify merchant count at roughly 30, against the million-plus promised at launch. Etsy called ChatGPT a valuable discovery channel while describing checkout volume as relatively low.

What replaced it is a discovery experience plus merchant-built ChatGPT apps that route users out. The discovery experience finds products from an image or a description plus budget and preference criteria.

Walmart launched an in-app service supporting account linking, loyalty and its own payments. Shopify began connecting storefronts with purchases completed in an in-app browser.

Target, Sephora and Nordstrom backed the new experience.

The feed infrastructure found a second life in ad targeting two months later.

Three weeks after announcing the plan, OpenAI switched ads on in the US. The post was "Testing ads in ChatGPT", covered the same day by TechCrunch and others.

The rules: ads appear below responses, labeled sponsored, for logged-in adults on Free and Go. Plus, Pro, Business, Enterprise and Education stay ad-free.

No ads run against health, mental-health or politics topics, and under-18 accounts are excluded.

That last rule describes February. OpenAI loosened the regulated-advice block in April 2026.

Health and finance ads are now possible in the US case by case.

Two design choices stand out. Free users can opt out of ads in exchange for fewer daily messages.

That prices the ads in messages rather than money.

And advertisers receive only aggregate views and clicks, not user data.

Ads do not influence the answers ChatGPT gives you, and we keep your conversations with ChatGPT private from advertisers.

OpenAI, "Testing ads in ChatGPT", 9th Feb 2026

Launch pricing was reported, not published: a 60-dollar CPM and a commitment around 200,000 dollars. Trade coverage attributed those figures to reporting rather than to OpenAI.

Within six weeks the pilot was generating 100 million dollars annualized, per an OpenAI spokesperson cited through Reuters.

The monetization turn arrived as a pair of posts. Fidji Simo, CEO of Applications, announced the test in "Our approach to advertising and expanding access to ChatGPT".

OpenAI would test ads in the US on Free and Go "in the coming weeks", with five published principles.

Alongside it, ChatGPT Go went worldwide at 8 dollars per month in the US, per the companion post. Go had launched in India in August 2025 and expanded through Asia that autumn.

It is the tier, with Free, that ads run on.

TechCrunch's framing captured the scale: ads were opening onto an audience OpenAI counted in the hundreds of millions. PPC Land's marketer briefing is the practical summary of what the announcement did and did not promise.

Ads went live on 9th Feb.