What search engine does ChatGPT use, and why it is no longer Bing
When ChatGPT searches the web, it first queries an index that no public tool lets you consult: OpenAI's own. Bing served as a provider for a long time and is still named in the documentation for Enterprise and Edu accounts, but the measurements published in the summer of 2026 show it almost absent from the consumer retrieval path. The rest of the results come from Google results pages scraped by contractors. Optimizing for Bing alone therefore means working on the smallest of the three doors.
- OpenAI never wrote that it used Bing. Its official wording, published on October 31, 2024 and still online, speaks of “third-party search providers” (in the plural) and of content provided directly by its partners.
- The head of ChatGPT testified under oath in April 2025 that OpenAI was developing its own web indices, a search index and ranking signals, and that the quality of its providers other than Google was causing it serious problems.
- OpenAI runs its own crawler, OAI-SearchBot, whose only documented function is to surface websites in ChatGPT's search features.
- Until July 21, 2026, ChatGPT exposed in its streams the name of the pipeline behind each result. Three researchers documented it: the in-house index dominated, scraped Google came next, and Bing practically never appeared.
- Bing has changed trades. Microsoft shut down its consumer search APIs on August 11, 2025 and has written since that its index serves to ground AI answers rather than to rank pages for human readers.
- No tool tells you whether your page is in OpenAI's index. That is the underlying constraint, and it means measuring by citation.
On this page
Third-party search provider
A company that sells another company access to a web index through a programming interface rather than through a results page. The customer receives addresses, titles and snippets that it reuses in its own way. It receives neither the ranking, nor the layout, nor the features of the original results page. This is the distinction most articles on the subject flatten: going through a search provider has never meant displaying its results page.
What OpenAI says officially
The reference sentence dates from the launch of ChatGPT search, on October 31, 2024, and it has not moved since: “ChatGPT search leverages third-party search providers, as well as content provided directly by our partners, to provide the information users are looking for.” Third-party search providers, in the plural, plus content from publishing partners. Bing is not named there, and it does not appear in the consumer help page either, which says the same thing more briefly: “ChatGPT search sometimes partners with other search providers.”
Bing is named elsewhere, in the page meant for Enterprise and Edu accounts: “ChatGPT searches based on prompts and may share disassociated search queries with the Bing search engine to return web results.” The same page adds, however, two sentences later, that structured data “may be shared with Bing or third-party data providers”. In other words, even in the document most explicit about Bing, OpenAI keeps the option of going elsewhere.
The practical consequence can be checked every day: a site ranked first on Bing may never appear in ChatGPT, and a site absent from Bing's first pages may be cited there.
The crawler that gives the intent away
OpenAI documents four crawlers. The one that matters for visibility is called OAI-SearchBot, and its official description fits on one line: it is “used to surface websites in search results in ChatGPT's search features”.
| Crawler | Documented function | What blocking it leads to |
|---|---|---|
OAI-SearchBot | Surfacing websites in ChatGPT's search features | You drop out of ChatGPT search. It is the only block that takes visibility away from you |
GPTBot | Making foundation models more useful and safer, which means training | Your future content leaves the training sets. No effect on search |
ChatGPT-User | Certain actions triggered by a person in ChatGPT and custom GPTs | OpenAI specifies that robots.txt rules “may not apply”, since the action comes from a person |
OAI-AdsBot | Validating the safety of pages submitted as ads in ChatGPT | No effect on organic visibility. The crawler validates ad pages, it does not condition any citation |
One detail of this table is worth pausing on. A company that simply resold a provider's results would have no need for a crawler of its own. The crawler exists because it feeds something, and that something does not belong to the provider.
What OpenAI said under oath
The strongest piece of evidence in this file does not come from a technical blog. It comes from a courtroom.
In April 2025, at the remedies trial in the United States antitrust case against Google, Nick Turley, head of ChatGPT, testified under oath. The proposed findings of fact filed on May 29, 2025 by the United States and the plaintiff states summarize his testimony and point to the pages of the transcript. Four passages matter.
First passage: OpenAI began grounding ChatGPT on third-party search results in 2023 (Rem. Tr. 502:10-19). Second passage: it is developing a search capability “involving web indices, a search index, and ranking signals” (Rem. Tr. 460:6-461:1). The plural of “web indices” and the singular of “a search index” are in the text of the document.
Third passage, the most telling: “Today, OpenAI continues to experience ‘significant quality issues’ with non-Google search API partners” (Rem. Tr. 392:11-23). The inner quotation marks are the document's own, which repeats the witness's words there. Fourth passage: OpenAI explains that in 2024 it sought access to Google's search interface for its grounding needs and was turned down (Rem. Tr. 413:8-23, 414:4-416:3).
Read the full sequence. A company has depended on search providers since 2023, judges their quality seriously deficient, asks Google for access to its index, is told no, and builds its own. This is not a blogger's hypothesis: it is sworn testimony, cited in a court document filed by the plaintiffs, with the transcript references.
The two months when ChatGPT showed its plumbing
What remains is to find out what the mix gives in practice. A window opened, then closed.
Until July 21, 2026, the ChatGPT interface let through in its streams an undocumented field, result_source, which named the retrieval pipeline of each result. Four values were observed: labrador, which corresponds to OpenAI's index, bright and oxylabs, two contractors that scrape Google results pages, and serp, residual. On July 21, OpenAI removed the field.
Three pieces of work made use of the window, and you need to know which one carries what weight. Suganthan Mohanadasan documented the field first, on a single account and thirty conversations, describing his own counts as indicative. Nicolas Sitter worked on a very large volume in a single vertical, hotels, with 50,899 captures between December 2025 and June 2026.
The third is by far the most solid, and it is the one we rely on. The French agency Resoneo analyzed 1,249 answers in July and August 2026, with an extension that captures the raw streams, replay through the programming interface, several countries and several account tiers, then published its method and its data. A short version appeared in Search Engine Land on August 17, 2026 under the byline of Olivier de Segonzac, co-founder of Resoneo. So this is not confirmation by a third party, it is the same study in two formats, and we cite it as such.
Three results come out of it. The split first: 74.7% of results came from the in-house index in thinking mode on a free account, while scraped Google rose to 75.3% in thinking mode on a paid account. The match with Bing next, or rather its absence: no snippet served by the in-house index matched a Bing snippet, which Search Engine Land sums up bluntly as “No labrador snippet matches a Bing snippet.” The method last: the team published canary pages on its own domain, with the full server logs, to observe what OpenAI's crawlers went to fetch instead of inferring it.
One observation from the same study has immediate practical value. In instant mode, the model receives only three elements per result: the title, the address and a snippet. When that snippet comes from the in-house index, it is cut a little beyond 200 characters and taken from the start of the page's rendered body, and the meta description tag is ignored. When it comes from scraped Google, that tag is picked up one time in three. The study also notes that one page in seven has no H1, which leaves the template to choose the snippet's starting point at random. And 24% of the titles served by the in-house index ran past 75 characters, the limit beyond which Bing truncates its own.
The contradiction to know about. Resoneo writes that it never observed a bing value: “We have never observed it, on any account, in any regime, on any date.” Neither did Nicolas Sitter. Suganthan Mohanadasan, for his part, saw it. Resoneo puts forward a hypothesis to explain the gap, making clear that it is only that: the value would be exposed only on Teams accounts, because of the relationship with Microsoft. Nobody can settle it anymore, since the field has disappeared. So treat Bing's share as small and not nil, rather than as equal to zero.
What Bing has become
Even when you go through Bing, you no longer go through the Bing of before, and Microsoft says so in black and white.
Act one, on August 11, 2025: the consumer Bing search interfaces are shut down, replaced by “Grounding with Bing Search” in Azure, which delivers grounded answers to agents rather than raw content to developers.
Act two, on May 6, 2026, in a Bing blog post signed by three Microsoft AI leaders. The central sentence is worth quoting word for word: “Search indexing was built to help humans decide what to read. Grounding is being built to help AI systems decide what to say.” The same text spells out the shift in the unit of value, “The unit of value shifts from documents to groundable information”, meaning discrete, supportable facts with clear provenance. And it sets the new measurement requirement: “Search optimizes for likelihood of relevance. Grounding must measure strength of evidence.”
Act three, in June 2026: Microsoft announces Web IQ, presented by Knut Risvik, Distinguished Engineer for Search and AI, as “a search engine for AI systems”, which “does not just return documents; it returns passages and structured evidence objects”. The post announces neither pricing nor a general availability date.
The tension is worth noting. In February 2026, Jordi Ribas, Corporate Vice President for Search and AI at Microsoft, wrote that “Microsoft grounding powers nearly every major AI assistant in the market”. The measurements from the summer of 2026 invite you to read that sentence as a commercial claim rather than as a description of ChatGPT's pipeline.
What it changes in the work
Write for the start of the page body, not only for the description
When the snippet comes from OpenAI's in-house index, it is taken from the start of the rendered body and the meta description tag is ignored. When it comes from scraped Google, the description is picked up one time in three. So both need care: put the answer in the first two sentences under an H1 that names the subject, and check that every page does have an H1.
Let OAI-SearchBot through, and separate that decision from the one on GPTBot
These are two distinct choices. Blocking training does not remove you from search, and blocking search does not protect you from training. Check the rule on all your templates, not only at the root. The full reasoning is in Business owners: should you let AI crawlers access your content? and the same trap exists at Google, as shown in Blocking Google-Extended does not remove you from Google Search.
Do not ease off on Google
A large share of what ChatGPT retrieves is Google repackaged by contractors, and the proportion rises on paid accounts in thinking mode. A site damaged on the Google side pays for it twice.
Open Bing Webmaster Tools all the same
Bing's share of retrieval looks small, but the tool has become the only source of real data on AI grounding. Since February 2026, its AI performance report exposes the grounding queries that were used to retrieve your content, and since June 2026 your site's citation share on a given query. Nobody else provides that.
Make your facts groundable
Microsoft's vocabulary can be put to work directly: discrete, supportable facts with clear provenance. In concrete terms, figures accompanied by their source and their date, self-contained definitions, comparison tables that read well out of context, and markup that does not depend on JavaScript.
Measure by citation, never by position
There is no rank to track in ChatGPT. Build a fixed sample of questions representative of your customers, ask them at regular intervals, and note what is cited. The utm_source=chatgpt.com parameter that ChatGPT adds to its outbound links gives you the traffic, not the visibility. What each tool sees, and what none of them sees, is detailed in Measuring your AI visibility: what each tool actually sees.
Why the comparison with Google does not hold
At Google, the chain is single and instrumented: one index, one console, reports, documentation that says what counts and, since a worldwide rollout on August 31, 2026, a performance report that delivers impressions in the AI surfaces of Search, plus a setting to opt out of them that Google says it does not use as a ranking signal elsewhere.
At OpenAI, there is none of that. No console, no coverage report, no submission tool, no way of knowing whether a page is indexed. The FAQ meant for publishers points to the standard tools, the robots.txt file and the noindex tag, and to the tracking parameter in outbound links. It even admits, about the Atlas browser, that “if we obtain the URL of a disallowed page from a third-party search provider or by crawling other pages and have signals the page is relevant to a user's query, we may surface just the link and page title”. The wording describes the hybrid architecture without naming anyone.
So three transparency regimes coexist within the same industry: Microsoft publishes citation share, Google publishes impressions, OpenAI publishes nothing. That is no reason to ignore ChatGPT, it is a reason to promise no rank and to check by hand. The boundary between the two trades is covered in SEO and GEO: what really differs between the two.
The sentence to remember if you keep only one
What ChatGPT queries is composite: an OpenAI index on the front line, Google scraped by contractors as a top-up, a Bing share that is small and cannot be measured, and a mix that changes without notice.
Saying “ChatGPT is Bing” is not only inaccurate. It produces an incomplete task list, where you optimize for the smallest of the three channels and drop the rest.
If you want to know where you stand before committing to anything, the GEO audit starts precisely with the access files and a citation survey.
What we do not know
An honest article on this subject has to say where public knowledge stops.
- The current mix. The proportions measured hold for the summer of 2026, on the samples described. The most solid measurement comes from a single team. Since the field was removed, nobody can verify the split today.
- What feeds the in-house index. Search Engine Land acknowledges it: the canary pages saw OAI-SearchBot almost silent while the index kept serving results. The crawler that fills the index has not been identified.
- Bing's exact share. One researcher observed the value, two others never did, and the observation window is closed.
- Confirmation by OpenAI. No publication from the company claims the existence of an index of its own. Everything rests on the court testimony, the documentation of its crawler and the observation of its streams.
What stays in-house: open the robots.txt file and check line by line what is granted to OAI-SearchBot, on every template and every subdomain; check that every page carries an H1; and reread the first two hundred characters of the body of your most commercial pages, since that is what the model receives. These three checks take half a day, and everything else depends on them.
Getting cited by generative engines and using it to sell is at the heart of the Strengthen your visibility in AI answers goal.
The other half of this file is here: Claude searches with Brave: what it changes for getting cited. And if you would rather start with a status check, see our approach to AI visibility.
Frequently asked questions about ChatGPT's search engine
Does ChatGPT use Google?
Partly, but not directly. The measurements published in the summer of 2026 show that a large portion of ChatGPT's results comes from Google results pages scraped by specialized contractors, and that this portion rises sharply in thinking mode on paid accounts. OpenAI also testified in court that it asked Google for official access to its search interface in 2024 and was turned down. So you cannot neglect Google, even if your objective is citation in ChatGPT.
Should you open a Bing Webmaster Tools account?
Yes, but more for measurement than for indexing. Bing's share of what ChatGPT retrieves looks small. On the other hand, Bing Webmaster Tools is currently the only tool that shows real data on the grounding of your pages in AI answers: the grounding queries used to retrieve your content since February 2026, and your citation share on those queries since June 2026. No other player publishes that.
Does blocking GPTBot remove me from ChatGPT?
No. GPTBot is used for training the models. The crawler that conditions your presence in ChatGPT search is called OAI-SearchBot. These are two separate decisions, and many sites confuse them. You can refuse training by blocking GPTBot while staying visible, as long as you let OAI-SearchBot through.
Why does my site never appear in ChatGPT's answers?
Three causes come up most often. First, an unintended block on OAI-SearchBot, sometimes on a single template. Next, a page whose first lines of body text contain no usable answer, since the snippet served by OpenAI's index is taken there, over a little more than 200 characters, and the meta description tag is ignored in that case. Finally, the absence of a main heading, which makes the snippet's starting point unpredictable: the measurements from the summer of 2026 note that one page in seven has no H1. You also have to accept that no tool will confirm indexing, unlike at Google.
Is the llms.txt file useful for this?
No. None of the crawlers documented by OpenAI reads that file, and no publication from the company refers to it. What counts is your robots.txt file, the actual presence of your content in the indexes concerned and the structure of your pages.
Can we know today where ChatGPT's results come from?
Not directly anymore. Until July 21, 2026, ChatGPT exposed in its streams an undocumented field that named the pipeline of each result, which allowed three researchers to document it. OpenAI removed the field that day. The known proportions therefore hold for that window, and the most solid measurement comes from a single team, the agency Resoneo. The architecture described remains the best public information available, without being verifiable in real time.
Is this worth the effort for a Quebec SMB?
The actions that count here cost almost nothing and also serve your classic SEO: check that search crawlers are not blocked, put the answer in the first two sentences under a clear main heading, and document your figures with their source. What costs a lot is paying for a promise of rank in ChatGPT, since no rank exists and nobody can measure it.
- OpenAI, “Introducing ChatGPT search”, October 31, 2024, updated December 16, 2024 and February 5, 2025: openai.com
- OpenAI, crawler documentation, accessed September 6, 2026: developers.openai.com
- OpenAI, help centre, “Searching the web with ChatGPT”, accessed September 6, 2026: help.openai.com
- OpenAI, help centre, “ChatGPT search for Enterprise and Edu”, accessed September 6, 2026: help.openai.com
- OpenAI, help centre, “Publishers and developers FAQ”, accessed September 6, 2026: help.openai.com
- “Plaintiffs' Remedies Proposed Findings of Fact”, United States v. Google, remedies phase, filed May 29, 2025, testimony of Nick Turley from April 2025, references Rem. Tr. 392:11-23, 413:8-23, 460:6-461:1 and 502:10-19: justice.gov
- Olivier de Segonzac, “Inside ChatGPT's retrieval stack: The index, cache, and pages it actually reads”, Search Engine Land, August 17, 2026: searchengineland.com
- Resoneo, “What ChatGPT pulls, what it shows, what it cites”, July 2026, updated August 2026, the full study of which the article above is the short version: think.resoneo.com
- Suganthan Mohanadasan, “ChatGPT Changed How It Picks Sources While You Were Reading My Last Post”, July 14, 2026, first public record of the field and the only observation of a bing value: suganthan.com
- Nicolas Sitter, “ChatGPT's Hidden result_source (2026): How It Sources Hotel Answers”, 50,899 captures in the hotel vertical: nicolassitter.com
- Microsoft, end-of-life announcement for the Bing search interfaces, retired August 11, 2025: learn.microsoft.com
- Krishna Madhavan, Knut Risvik and Meenaz Merchant, “Evolving role of the index: From ranking pages to supporting answers”, Bing blog, May 6, 2026: blogs.bing.com
- Knut Risvik, “Announcing Microsoft Web IQ”, Bing blog, June 2026: blogs.bing.com
- Jordi Ribas, “Elevating the Role of Grounding on the AI Web”, Bing blog, February 12, 2026: blogs.bing.com
- Microsoft, “Introducing AI Performance in Bing Webmaster Tools”, February 10, 2026: blogs.bing.com
- Microsoft, “New AI Visibility Insights in Bing Webmaster Tools”, June 16, 2026: blogs.bing.com
- Google, “Introducing Search Generative AI performance reports in Search Console”, June 2026, worldwide rollout on August 31, 2026: developers.google.com
- Rachel Handley, “ChatGPT Definitely Uses Google to Search the Web”, Semrush, August 6, 2025, cited as a historical milestone: semrush.com
Gabriel almost always takes your first call and carries out your audit. He builds the strategy starting from your growth goal: where to put your budget, which market to test and how to connect each lead to a real sale in your CRM. He mainly leads engagements for three goals: Optimize the profitability of your digital campaigns, Develop a new market, and Generate demand and growth. With Geneviève, he also works on organic search (SEO), AI visibility (GEO) and conversion rate optimization (CRO). The sales a Google Ads or Meta Ads campaign brings in depend on the page that receives the click. He writes mainly about marketing strategy, paid advertising and measurement.
About Falia →
SEO
Paid advertising
Conversion
AI visibility
Web design