ChatGPT Search

By Bill Hartzer · Updated 19 August 2026

What ChatGPT Search is, and why it is not one thing

ChatGPT Search is web retrieval inside ChatGPT. You type a question into the ordinary chat box; when the model judges that current web information is needed — or you ask for it — ChatGPT retrieves pages, reads some of them, and writes an answer with inline citations. There is no ranked list of ten blue links and no separate site to visit. The chat window is the surface.

For a very large number of people this is now the primary way they look things up, which is why the question of where its results come from matters far more than its modest standing as a named "search engine" would suggest.

That answer is unusually complicated, and it is why this page carries the verdict hybrid index. Most hybrid engines blend one crawler with one partner feed. ChatGPT runs three distinct sourcing paths — its own index, purchased web data scraped from Google, and a contractual relationship with Bing — and which one answers you depends on what you pay.

That deserves to be stated flatly, because almost nobody knows it: the same question asked by a free user and a paying user is answered from different indexes. This is not a quality tier. It is a different view of the web, and the product does not disclose it.

The crawlers, and the one publishers get wrong

OpenAI documents four automated agents and publishes IP ranges for all of them. The distinction between the first two is the most practically useful piece of information on this page.

OAI-SearchBot is the search crawler. OpenAI states plainly that sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers. This is the crawler that governs visibility.
GPTBot is the training-data crawler. Blocking it signals that content should not be used to train foundation models. It has nothing to do with whether you appear in answers.
ChatGPT-User fetches a specific page when a user or an automated action asks ChatGPT to visit it. OpenAI's documentation says robots.txt rules may not apply to it, on the same user-initiated reasoning Perplexity uses for its equivalent agent.
OAI-AdsBot validates the safety of pages submitted as advertisements. Its presence in the documentation is itself a dated fact about the business model.

Publishers reverse the first two constantly: blocking GPTBot does not remove a site from ChatGPT's answers, and blocking OAI-SearchBot does. To OpenAI's credit, the separation is deliberate and better designed than most competitors' — it lets a publisher say "index me, do not train on me," which is a distinction the rest of the field mostly refuses to offer.

Where the results actually come from

Three accounts of ChatGPT's sourcing coexist and they do not agree. Rather than averaging them, here is each with its source.

OpenAI's own account. The launch announcement of 31 October 2024 said ChatGPT search "leverages third-party search providers, as well as content provided directly by our partners." The help documentation, still live on 19 August 2026, says ChatGPT search "sometimes partners with other search providers," naming Bing and Shopify, and states that Bing is the sole third-party provider for Enterprise and Edu workspaces. It also says queries are rewritten before being sent to providers rather than forwarded raw.

Semrush, 6 August 2025. Experiments by Abhishek Iyer and Leigh McKenzie found a page reachable only via Google being cited by ChatGPT, with McKenzie finding that the paid tier used Google while the free tier did not — contradicting the then-universal assumption of Bing-only sourcing.

Search Engine Land, 17 August 2026. The most detailed public measurement to date, by Olivier de Segonzac, identifies an internal OpenAI retrieval hub it calls labrador, serving web, news, academic, Reddit and YouTube results, built from OAI-SearchBot's crawling. Its findings: only 1.5% of labrador URLs appeared in Bing's top 20 for the same query, with no snippet matches, so this is not a Bing passthrough. On free-tier instant answers, labrador supplied 74.7% of results and Google-fed sources only 3.1%. On paid-tier thinking mode, 75.3% of results came from scraped Google results purchased from third-party data providers, the study naming Bright Data and Oxylabs.

The reconciliation that fits all three: Bing is a contractual partner whose role has shrunk to specific surfaces, OpenAI's own index carries the cheap default path, and purchased scraped-Google data carries the expensive one. The blunt implication is an economic separation — a cheap internal index for free users, expensive fresh Google-derived data for paying ones.

Two caveats belong with this. The Bright Data and Oxylabs relationships are inferred from measurement, not confirmed by any party, and OpenAI publishes no current itemised supplier list. And the measurement is a single study of a demonstrably volatile stack. It is the best available evidence; it is not a permanent fact.

Why Bing's role shrank

Two events in 2025 pushed OpenAI away from the arrangement everyone assumed was permanent.

On 11 August 2025 Microsoft retired the Bing Search APIs entirely, decommissioning existing instances and pushing remaining consumers toward the considerably more expensive "Grounding with Bing Search" product inside Azure AI Agents. Any product buying Bing results the cheap way had to move.

In October 2025 OpenAI restructured into OpenAI Group PBC under the control of the non-profit OpenAI Foundation, and Microsoft's exclusivity over OpenAI — as cloud provider and as search partner — was loosened in the same reorganisation. Microsoft remains a large minority shareholder.

Both changes point the same direction: OpenAI had strong reasons to own its retrieval and buy the rest wherever it was cheapest. A product widely described as "Bing with a chat interface" now has roughly 1.5% URL overlap with Bing's top results.

The index itself also works differently from a conventional engine in one measurable way: the snippets ChatGPT retrieves are query-independent and cut at index time, roughly 200 characters following the page's H1. A classic search snippet is assembled per query to match what you asked. That is a meaningfully different retrieval object, and it shapes what the model has to reason over.

How it makes money, and the advertising turn

ChatGPT Search is not monetised as search. It is a retention and capability feature of ChatGPT, which is monetised four ways: consumer subscriptions with a free tier and paid tiers offering higher limits, better models and — materially — the more expensive retrieval path; seat-priced business subscriptions for Team, Enterprise and Edu, which carry the Bing-only path and stronger data commitments; metered API billing for developers; and, since 2026, advertising.

The advertising is the consequential change. Per its Wikipedia record, advertising in ChatGPT was announced on 17 January 2026, targeted at free-tier adult users in the United States, and ads were visible in production by March 2026. The existence of OAI-AdsBot in OpenAI's crawler documentation corroborates this independently.

This matters more than a line item. The largest AI answer surface on the internet now has an advertiser on the other side of it — which was precisely the structural critique of Google that AI search was marketed as escaping. And it is harder to audit here than there: a sponsored blue link is visibly labelled and skippable, while the influence of a commercial relationship on a paragraph of generated prose is not inspectable at all.

Money flows outward too. OpenAI has signed content-licensing deals with major publishers including the Associated Press, Axel Springer, Condé Nast, the Financial Times, Le Monde, News Corp, Reuters and The Atlantic. It is the only company in this category that has broadly licensed content rather than being sued over it, and against Perplexity's litigation docket that is a materially more defensible position.

Prices are not stated here. Tiers have changed repeatedly and any figure would age badly. The durable structural point is that what you pay changes which index answers you.

Privacy, regulation, and the missing off switch

OpenAI says search queries are rewritten before being sent to third-party providers rather than forwarded raw. That is a genuine mitigation, with the corollary that OpenAI decides what the provider sees. Location sharing is user-controllable in settings. Business tiers carry contractual commitments that inputs are not used for training by default; consumer tiers have historically defaulted the other way with an opt-out available.

The strongest external fact in the record is regulatory. Italy's data protection authority, the Garante, ordered ChatGPT offline in Italy in March 2023 on GDPR grounds, restoring it in April 2023 after OpenAI made changes. No other product in this category has been taken offline in a country by a regulator.

The advertising change is a privacy fact as well as a commercial one. Ad-supported delivery to free-tier users implies targeting signal derived from conversation, and OpenAI's precise targeting inputs and retention for advertising are not publicly documented. That is a thing to watch rather than a settled criticism. The sensitivity ceiling has risen too: ChatGPT Health launched in January 2026, so health questions go into the same box that performs web search and, on the free tier, now carries ads.

Finally, there is no clean way to turn retrieval off. OpenAI's documentation describes no toggle to disable web search; it fires when the model decides it is needed. You can steer it message by message, but you cannot opt into a deterministic, inspectable list of results. That is the exact inverse of an opt-in AI posture, and worth knowing before choosing this as a default research tool.

Who it suits, and who it does not

ChatGPT is better than any ranked list at answering messy, multi-part, context-dependent questions, and the conversational follow-up loop is a real advance over reformulating a query five times. At roughly 900 million weekly active users as of February 2026 — a ChatGPT figure, not a search figure — it is in practice one of the most consequential information-retrieval surfaces ever built. Its agentic and long-form research features, including Operator from January 2025 and Deep Research, extend that reach further.

It is a poor choice for several specific people:

Anyone who needs to audit where an answer came from. The gap between OpenAI's disclosure and independent measurement is large, and the sourcing changes with your subscription.
Anyone doing work where a fluent wrong answer is expensive and who will not click the citations. The Search Engine Land study found pages ChatGPT actually opened were cited 74% of the time, versus 7% for pages merely pulled into the candidate set — reassuring in one direction, and a reminder in the other that a citation is a claim about provenance, not a guarantee of fidelity.
Anyone who wants to browse and compare sources themselves. There is no results list to scan.
Anyone who left conventional search to escape advertising, at least on the free tier in the US since March 2026.
Anyone who wants a choice of model. Unlike Perplexity, Kagi and the former You.com, retrieval and generation are both OpenAI's. You get OpenAI's models or nothing.

Common misconceptions

"ChatGPT Search is Bing with a chat interface." Closer to true in 2024, badly wrong in 2026. Independent measurement found only 1.5% URL overlap with Bing's top 20. OpenAI runs its own crawler and index and buys scraped Google data for its expensive path.

"ChatGPT does not have its own index." It does. The retrieval hub identified as labrador, fed by OAI-SearchBot, supplied roughly three-quarters of free-tier results as measured in August 2026.

"ChatGPT is fully independent of Google and Bing." Also wrong. Paid-tier thinking mode ran about 75% on purchased scraped-Google results, and Bing remains the named provider for Enterprise and Edu workspaces.

"Blocking GPTBot keeps you out of ChatGPT's answers." No. GPTBot governs training; OAI-SearchBot governs search visibility. OpenAI states the distinction explicitly and publishers still get it backwards.

"ChatGPT Search has no ads." Advertising was announced on 17 January 2026 and was in production by March 2026 for free-tier adult users in the United States.

"Everyone gets the same answers." Free and paid users are served from materially different retrieval backends.

"You can compare ChatGPT's market share with Google's in StatCounter." You cannot. StatCounter measures browser-referred search traffic and does not break ChatGPT out; circulating figures come from panels and surveys with incompatible methodology. Putting them in the same table is the most common quantitative error in AI-search commentary. There is also no standalone Wikipedia article for ChatGPT Search — the main ChatGPT article is where it is covered.

Frequently asked questions

Does ChatGPT Search use Bing?

Partly, and much less than most people assume. OpenAI's help documentation names Bing as a partner and says it is the sole third-party provider for Enterprise and Edu workspaces. But independent measurement published on 17 August 2026 found only 1.5% overlap between the URLs OpenAI's own index returned and Bing's top 20 for the same queries, with no snippet matches. Bing's role has shrunk to specific surfaces rather than being the engine behind ChatGPT.

Does ChatGPT have its own search index?

Yes. OpenAI runs a crawler called OAI-SearchBot which feeds an internal retrieval hub identified in independent research as labrador, serving web, news, academic, Reddit and YouTube results. That index supplied around 74.7% of free-tier instant answers in an August 2026 measurement. OpenAI does not describe the index publicly in that level of detail; its own documentation refers more vaguely to third-party search providers and partner content.

Do free and paid ChatGPT users get different results?

Yes, and this is the least-known fact about the product. Measurement published in August 2026 found free-tier instant answers drew about 74.7% of results from OpenAI's own index, while paid-tier thinking mode drew about 75.3% from scraped Google results purchased through third-party data providers. That is not a quality tier but a different view of the web, and nothing in the product tells the user which one they are getting.

How do I stop ChatGPT from showing my site?

Block OAI-SearchBot, not GPTBot. OpenAI states that sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers. GPTBot is a separate crawler governing whether content is used to train foundation models, and blocking it has no effect on search visibility. A third agent, ChatGPT-User, fetches pages when a user asks ChatGPT to visit them, and OpenAI's documentation notes robots.txt rules may not apply to it.

Does ChatGPT have ads?

Yes, since 2026. Advertising in ChatGPT was announced on 17 January 2026, targeted at free-tier adult users in the United States, and ads were visible in production by March 2026. OpenAI's crawler documentation independently corroborates this with a dedicated agent, OAI-AdsBot, that validates pages submitted as ads. The structural consequence is that the largest AI answer surface online now has an advertiser on the other side of it.

Can I turn off web search in ChatGPT?

Not cleanly. OpenAI's help documentation describes no toggle that disables web retrieval outright; it fires when the model judges current information is needed. You can steer it on a per-message basis, but you cannot switch the capability off and get a purely offline answer, and you cannot ask for a deterministic list of results instead of prose. This is the opposite of engines that treat AI answers as something you invoke.

Is ChatGPT a search engine?

It performs search but it is not structured like one. There is no ranked results page, no way to browse and compare sources yourself, and no separate search product to visit. It is more accurate to call it an AI product with retrieval added than a search engine with AI added. That framing also explains its shape: retrieval, ranking and generation are all decided for you, and the output is prose with citations rather than a list.

How many people use ChatGPT Search?

Nobody knows, including from public sources. ChatGPT reported roughly 900 million weekly active users as of February 2026, but that is a whole-product figure and an unknown fraction of those sessions trigger web retrieval at all. OpenAI has published no figure for search queries specifically. StatCounter does not measure ChatGPT in its search-engine series, so any percentage comparing it to Google comes from a different methodology entirely.