How ChatGPT search chooses which sites to cite

By Roshan, founder of SearchMaxx - Sep 30, 2026 - updated Oct 18, 2026 - 7 min read

ChatGPT can only cite pages its search crawler is allowed to reach. When it searches, it rewrites the question into shorter queries, sends them to partner search providers, and ranks what comes back "using multiple factors". OpenAI has not published those factors. Studies show it favors pages whose titles match the question and, for factual questions, the official source.

Most advice on this topic is guesswork dressed up as a formula. Our position is narrower: do the few things OpenAI has documented, then use the independent data to choose what to write. Where we infer something below, we say so.

What OpenAI has documented

OpenAI's public statements fit on one page. Here they are, with sources.

We read both help pages in October 2026. OpenAI updates them often, so check the live versions.

How a search happens

OpenAI's help page describes the steps:

1. ChatGPT decides whether to search. It "may search the web automatically when your question would benefit from current information." Users can also turn search on by hand.
2. It rewrites the question. "ChatGPT search sometimes partners with other search providers. When it does, ChatGPT search typically rewrites your query into one or more targeted queries that it sends those providers."
3. It can search again. "After reviewing the initial results, ChatGPT search may send additional, more specific queries to other search providers."
4. It ranks and cites. Responses "may include citations", and a Sources panel lists cited pages and other relevant links.

The same page links to the privacy policies of Microsoft and Shopify as search providers whose policies apply. (OpenAI Help Center)

Inference: Microsoft runs Bing. So pages that Bing has indexed are likely candidates when ChatGPT queries that provider. OpenAI has not said what share of results comes from any one provider. It also notes it may learn a URL "from a third-party search provider or by crawling other pages", which suggests it mixes provider results with its own crawling.

That is the whole public record. OpenAI does not say how it weighs authority, freshness, links or structured data. Anyone who tells you the exact formula is guessing.

What independent studies have measured

Ahrefs, which sells SEO and AI tracking software, has published the largest studies we know of on ChatGPT citations. Its results are vendor research, not peer-reviewed, and each covers one model at one point in time. Still, the patterns are consistent enough to act on.

Most answers do not search at all

In 66.6 million free, logged-out ChatGPT responses from August and September 2026, only 14.4% triggered a web search. When it did search, it ran about 1.6 queries per response. (Ahrefs, October 12, 2026)

Ahrefs notes that paid and reasoning models search more, and that "best X" style prompts trigger more lookups. Buyer questions about software tend to be that kind. But the takeaway holds: much of what ChatGPT says about your category comes from the model itself, not from a live search.

Google rankings do not predict ChatGPT citations

Ahrefs reran 118,931 of ChatGPT's search queries on Google. Only 6.82% of the URLs ChatGPT returned were in Google's top 10 for the same query. 83.39% did not appear in Google's results at all. (Ahrefs, August 20, 2025)

So ranking first on Google does not mean ChatGPT will cite you, and the reverse is also true.

Titles that match the question win

Ahrefs studied 1.4 million prompts to compare pages ChatGPT retrieved and cited against pages it retrieved and skipped. (Ahrefs, April 15, 2026)

• About half of retrieved URLs were cited.
• Cited pages had titles closer in meaning to the prompt. The similarity score was 0.602 for cited titles against 0.484 for skipped ones.
• Search results with natural-language URL slugs were cited 89.78% of the time, against 81.11% without.
• Reddit made up 67.8% of the URLs ChatGPT retrieved but did not cite. Ahrefs reads this as ChatGPT using Reddit for context while rarely citing it.
• Cited search pages had a median age of about 500 days. Very new pages were mostly in the skipped group.

These are associations, not proof of cause. Ahrefs measured relevance with open-source embedding models, because ChatGPT's own scoring is private.

For facts, the official source wins

The October 2026 Ahrefs study also looked at which sources ChatGPT cited between March and September 2026 in the US. (Ahrefs)

• 84.4% of answers to fact questions cited at least one official source. For comparison questions it was 68.3%.
• 6.8% of ChatGPT's queries used the site: operator. Of those aimed at the most-searched domains, 75.9% targeted brand sites.
• The words "official" and "price" became more common in its queries over the period.
• Reddit citations fell 88%, from about 1 in 5 answers in March to about 1 in 43 by late September. Wikipedia fell 80%.
• OpenAI publishing partners' share of citations among the top 200 domains rose from 7.5% in March to 15.8% in July, then settled at 12.5%.

Ahrefs reads this as ChatGPT moving toward original sources and away from sites that summarize them. That is its interpretation of two months of one model. It fits OpenAI's stated aim of "relevant, reliable information", but OpenAI has not confirmed it.

What this means for a software company

Put the documented facts and the measured patterns together and three points follow. The second and third are our inferences.

1. Access is the only hard rule. If OAI-SearchBot is blocked, by robots.txt or by a firewall, you will not be cited. This is documented.
2. Your own site should be the best source for facts about you. When a buyer asks what your product costs or whether it has SSO, ChatGPT increasingly looks for the official page. If that page is vague, missing, or hidden behind JavaScript, it will cite someone else's description of you.
3. Comparison questions are where third parties matter. For "best X for Y" questions, about a third of answers cited no official source at all in Ahrefs' data. Being named on the review sites and comparison articles that rank for those questions is how you show up there.

A checklist you can run today

The robots.txt step is covered in depth in which AI crawlers to allow. One file you can skip worrying about is llms.txt. OpenAI has not said its crawler reads it, as we cover in llms.txt in 2026.

Remember that a citation is not the only win. Many ChatGPT answers name products without any search or link. For how that turns into visits, and how to measure it, see zero-click search in 2026.

The short version

OpenAI documents access, not ranking. Let OAI-SearchBot in, make sure your pages send real text, and put the facts about your product on your own site in pages titled the way buyers ask. Then earn mentions on the comparison pages ChatGPT reads for "best X" questions. Everything beyond that is a pattern in someone's data, worth acting on, but not a rule.

Our free check tests the first two in about 10 seconds. It tells you whether OAI-SearchBot and the other AI crawlers can reach your site, and how much of your page they can read.

Want this done for your site?

Every day we make one change to your website that helps AI engines name you, and a person checks each one. Email me your website and the question your buyers ask AI.

Email Roshan to get started

Or check your site free: