ChatGPT SEO: how ChatGPT picks its sources and how to become one

ChatGPT SEO means making sure ChatGPT finds your pages and cites them when it searches the web, and that it knows your brand when it answers without searching. Those are two routes, and a study from October 2026 sums them up in one line: pages are cited when searched; brands are also named when remembered.

Plenty of claims about ChatGPT go around without evidence. This article sticks to what OpenAI documents itself and what studies with a published method have measured, each with a link. I checked every source on 11 November 2026.

When ChatGPT searches the web

Search is available on every plan, including the free one, according to ChatGPT’s pricing page. Whether ChatGPT actually searches for a given question is usually up to the model. For the web search tool in its API, OpenAI says so explicitly: the model can choose to search the web or not based on the input.

When ChatGPT searches, it doesn’t just pass the user’s question to a search engine. It writes its own queries. A study by Yunxuan Jiang and colleagues (October 2026) observed this across four ChatGPT versions: the fast Instant models mostly search in a single round, while the Thinking model often runs several. Pages retrieved more often across repeated runs are also more likely to be cited.

Without search, ChatGPT answers from its training data. There are no links then, but brands still get named. More on that below.

Where ChatGPT’s sources come from

This is where the most common assumption has no evidence behind it. You often read that ChatGPT uses Bing’s index. OpenAI, on the other hand, documents its own search crawler, OAI-SearchBot, and its API documentation refers to “cached/indexed results”, meaning its own cached and indexed web content. How much Bing contributes in the background is not stated in any documentation I could access publicly.

What can be measured is the outcome. The largest study so far is a paper by Mahe Chen and Nick Koudas (October 2026): 158,099 answers from generative search systems such as ChatGPT and Gemini to commercial questions, with 468,037 citations, collected in two markets over 90 days each. The authors also replayed the questions through the vendors’ APIs to see which pages ChatGPT’s search retrieved in the first place. Three findings matter in practice:

• Only what its own search retrieved gets cited. Every page ChatGPT cites comes from its own search results.
• Those results are different pages from Google’s and Bing’s. 97% of the pages ChatGPT retrieves are not in Google’s top 50 for any of its queries, and 99% are not in Bing’s top 30.
• But often from the same websites. For 51% of the retrieved pages, the website does appear in Google’s top 50, just with a different page.

The study by Jiang and colleagues reaches the same picture: even Google and the OpenAI API combined recover only about a quarter of the URLs ChatGPT cites. The match is much better at the level of websites than of individual pages.

What this means for you: ranking well on Google doesn’t guarantee ChatGPT will cite that exact page. A website that is strong on Google for a topic has a good chance of being in there with some page, though.

OpenAI’s crawlers: what you need to allow

OpenAI describes its bots on a dedicated page. Three of them matter for ChatGPT:

Bot	Purpose	What blocking it does
OAI-SearchBot	Surfaces websites in ChatGPT’s search	Blocked sites don’t appear in ChatGPT search answers, at most as navigational links
ChatGPT-User	Fetches pages when a user asks ChatGPT to	According to OpenAI, robots.txt rules may not apply because the fetch is user-initiated. It plays no part in search
GPTBot	Collects content for training the models	Signals that your content should not be used for training. It has no effect on search

There is also OAI-AdsBot, which only checks the landing pages of ads in ChatGPT.

OpenAI says these settings are independent of each other. You can allow OAI-SearchBot to appear in search and block GPTBot if you don’t want your content in training. After a robots.txt change, OpenAI says it takes about 24 hours for search to adjust. OpenAI also recommends allowing OAI-SearchBot’s published IP ranges. That matters if a firewall or CDN blocks bots: then even a perfect robots.txt doesn’t help.

The free AI crawler check shows which bots your robots.txt currently lets in. How to control individual bots is covered in AI crawler controls in 2026.

What makes a page citable

Chen and Koudas compared pages of the same website with each other. That matters, because otherwise strong websites mask everything else. The result is sobering for many GEO checklists:

• What works: pages whose titles and paragraphs match the questions people actually ask get cited more. Content works mainly by matching, in meaning, the queries ChatGPT writes itself.
• What only appears to work: FAQ blocks, structured data, summary boxes and page speed go with more citations only across different websites, not between pages of the same site. Stronger sites probably just have them more often.
• What matters more than any page property: whether the page shows up in ChatGPT’s own search at all. In the study, that explains far more of the difference between pages (32%) than all page properties combined (2.8%).
• What the model already knows: ChatGPT writes brands it knows into its queries, prefers websites it knows and names brands even without a page to back them up. In controlled experiments, the brand name alone reordered products with identical facts.

That leads to five things you can do:

1. Let OAI-SearchBot in, in robots.txt and in your firewall or CDN.
2. Serve content in the HTML. OpenAI doesn’t document whether OAI-SearchBot runs JavaScript. What is in the delivered HTML it can certainly read. More under technical SEO.
3. Phrase titles and subheadings as real questions and put the answer right below. One page per question beats one page that brushes past ten.
4. Get known where the model learns. Mentions in trade media, comparisons and directories make your brand one ChatGPT knows. That’s classic PR and link building, with a new reason.
5. Provide verifiable facts. Figures with a source and a date did best in the original GEO study too.

ChatGPT shopping: where it stands for European shops

For online shops, check OpenAI’s merchant page. As of November 2026 it says shopping is live for ChatGPT users in the US, with more regions to follow. OpenAI is moving away from buying directly inside ChatGPT (Instant Checkout); purchases happen on the merchant’s own site. If you sell through Shopify or Etsy, OpenAI says you’re already integrated. For shops in Germany, Austria and Switzerland there is nothing to set up yet. Keep an eye on the page.

How to measure whether ChatGPT mentions you

First-party data covers only part of it:

• Visits from ChatGPT show up in Google Analytics 4 since May 2026 in its own “AI Assistant” channel, as long as the browser sends a referrer.
• Fetches by ChatGPT are in your server logs: OAI-SearchBot for search, ChatGPT-User for pages users have ChatGPT open.
• Mentions without a click aren’t measured by anyone with first-party data. OpenAI offers no report for site owners. The AI Performance report in Bing Webmaster Tools counts citations in Copilot, in Bing and with unnamed partners. Microsoft doesn’t mention ChatGPT there.

That leaves sampling: asking the same questions regularly and recording whether you appear. Keep in mind what a study from April 2026 stresses: answers vary across runs, prompts and time. A single query tells you little; only repeated measurements show a picture. How to bring all the sources together is covered in measuring AI search visibility.

What you can skip

• llms.txt for ChatGPT. OpenAI has nowhere confirmed that ChatGPT reads the file on other people’s websites. What the file can do is in llms.txt: when it’s worth it.
• Blocking GPTBot out of fear for your visibility, or allowing it in the hope of more. Search depends on OAI-SearchBot, not GPTBot.
• Schema and FAQ blocks as a ChatGPT lever. They have their purpose for Google. The study found no sign that they lead to more ChatGPT citations within a website.
• “Bing SEO for ChatGPT”. That ChatGPT simply takes Bing’s results is not established, and the measurements point the other way.

When an audit is worth it

If ChatGPT names competitors in your market and not you, it’s rarely down to a single file. Usually it’s a mix of blocked bots, pages that don’t answer a specific question and a brand that barely appears outside its own website. The AI search visibility audit for €990 checks exactly that: 50 to 100 buyer prompts in ChatGPT, Perplexity, Gemini, Copilot and Google’s AI Overviews, your share of mentions against three competitors, crawler access and rendering. You get a 90-day roadmap your team can work through.