Guides · Updated October 31, 2026

How ChatGPT picks and cites its sources (and how to track yours)

The short answer: when ChatGPT searches, it rewrites the question into its own searches, reads a few results for each, and cites the pages whose title says plainly what it was looking for and that state the fact it needs. For software questions that's mostly “best X” lists and vendors' own product, pricing and docs pages. OpenAI doesn't publish how it chooses, and the pages change from one answer to the next, so the only way to know where you stand is to ask your buyers' questions many times and count.

Does ChatGPT always search?

No. It decides per question; across all ChatGPT use, about a third of questions searched in early 2026 (Semrush, via Search Engine Land). Buying questions search far more often: 83 of 100 did in one test (Grow & Convert). An answer without a search has no citations at all: it comes from what the model learned in training, where a small or new product barely exists.

How it picks the pages it cites

OpenAI documents its crawler (OAI-SearchBot) and that citations appear in the answer or its Sources panel, but not how pages are chosen. What people have measured:

• It cites what matches its own searches, not your question. One question becomes several searches, with years, “best X for Y”, product names and searches on vendors' sites. Cited pages' titles match those searches more closely than the question itself, and pages with plain-words URLs (/best-podcast-database) were cited more often: 90% against 81% (Ahrefs, 1.4M prompts).
• Ranking helps, but isn't the whole story. Re-running ChatGPT's visible searches, only about 40% of the pages it cited were anywhere in Google's or Bing's first 10 pages (Grow & Convert, 100 buying prompts).
• Most pages it reads aren't credited. Studies disagree on how many (from half to a few percent, depending on what counts as read), but they agree on the direction. Reddit is the clearest case: 68% of the pages ChatGPT retrieved and didn't cite were Reddit threads (Ahrefs). A thread can shape an answer without ever showing as a source.
• The shortlist forms early. In one capture of ChatGPT's own searches, products it wrote into a search reached the answer 69% of the time; products that only turned up in results, 2% (Mohanadasan, 60 conversations). Being on the pages that build the shortlist matters more than polishing your own.

Named, cited and linked are different

An answer can do three things for your product, and each is measured separately:

They come apart often. Across AI engines, most appearances of a domain were as a source with the brand never named; for ChatGPT, a domain was cited in 87% of its appearances and named in 21% (Semrush and Growth Memo, 3,981 appearances). The other way round, in 40 software categories, 92% of the tools ChatGPT named had a citation, but 87% of those citations were someone else's page; the vendor's own site, 13% (DerivateX, 233 recommendations, a small study). So counting your own domain's citations misses the lists that get you named, and credits pages used for a fact while someone else is recommended.

Which pages get cited for software

• “Best X” lists and roundups, the most-cited kind for software questions: 44% of cited pages in one study (Ahrefs, 26,283 pages). Cited lists share a shape: all were lists, 78% had the year in the title, 68% a comparison table (DerivateX, 143 pages). That's what cited pages look like, not proof that a table gets you cited.
• Vendors' own product, pricing, home and docs pages, checked once a product is on the shortlist. With lists, 78% of cited sources in buying answers (Grow & Convert).
• Review sites and Reddit, small as visible sources in 2026's software studies: a few percent each. Reddit is read far more than it's credited.
• App-store listings for apps: nobody has measured it, but SocialPhotos' Shopify App Store listing was its most-cited page.

Freshness matters less than the advice says: the median cited page was about 500 days old, and showing a date didn't make a page likelier to be cited (Ahrefs). Lists with the year in the title are the exception worth keeping current.

What our own answers show

PodSeeker (a podcast database for PR teams, ours) in the two weeks to October 30, 2026: 25 buying questions, 102 answers from the ChatGPT app, logged out, in the US.

• Its site was cited in 90 answers; it was named in 80. On no question was it named more often than its site was cited. On 5 questions, mostly how-tos, ChatGPT used its pages and named it less often or not at all: on “Should our in-house PR team use podcast booking software or hire a booking agency?” its pages were cited in 3 of 3 answers and it was named in none.
• 44 of its pages were cited. The most: the homepage (37 answers), pricing (31), two “best X” guides on its own blog, “Best podcast booking tools for PR teams” (29) and “Best podcast database for PR” (19), a docs page on how it works (17), the FAQ (14) and the API docs (11).
• 87 other sites were cited. The most: an independent blog's list of podcast tools (34 answers), two competitors' sites (32 each), G2 (10).

And askadia.ai, launched October 26: not cited in any of 13 answers to its own questions, while Peec AI's site was cited in 5. A new site starts at zero; we'll report here when that changes.

How often citations change

Far more often than recommendations. Week to week, 74% of the sites ChatGPT cited for a question were different, 85% of the pages; the typical question had no site cited in all 17 weeks (SISTRIX, 82,619 prompts). Day to day, only 1.1% of sources were cited all 7 days of a week (GetMentions, 67,144 answers). So “my page is cited” is a rate: cited in k of n answers, per question, over a week or more. One check proves nothing.

How to get your pages cited

1. Let it in. robots.txt allows OAI-SearchBot, and nothing important sits behind a login, a noindex or a script that never renders (the checklist).
2. Title and URL in your buyer's words. “Best podcast database for PR”, not “Discover our platform”. ChatGPT cites pages that plainly match its own searches.
3. The answer near the top. 44% of citations come from the first 30% of a page (Growth Memo, via Search Engine Land).
4. The pages it checks on a shortlist: pricing in plain numbers, what it does and for whom, integrations, an honest comparison or alternatives page.
5. Be on the lists it already cites. See which third-party pages appear in answers that name your competitors, and earn a place on them: an honest pitch, a free account to review.
6. Don't change dates for show. Update a list when it changes; a new date alone did nothing measurable.

Schema markup and llms.txt showed no measured effect on citations; prompts hidden in pages for AI are manipulation and can get a site excluded (what doesn't work).

How to track your citations

There's no citation report from OpenAI. Three views, each partial:

• Analytics (clicks). Visits from ChatGPT in GA4 (how to set it up). Since May 2026, a visit to your homepage is most likely a click on your name; a visit to a deeper page, a click on a citation.
• Server logs (reads). Requests from ChatGPT-User are ChatGPT fetching a page while answering someone; OAI-SearchBot is it indexing. A read isn't a citation, but your pricing page being fetched means you're being checked.
• Asking, repeatedly (citations). The only view of which pages are cited and whether you're named: your buyers' questions in the ChatGPT app, logged out, several times a week. For each answer, note whether you were named and every page cited: yours, your pages elsewhere (an app-store listing), lists that name you, lists that don't. Count per question, “named 3 of 5, cited 4 of 5”.

Pitfalls: some answers don't search and cite nothing; answers through OpenAI's API cite different pages from the app, with little overlap; about 10% of cited links led to error pages in one test; and citations churn, so judge over weeks.

Google's AI Overviews and AI Mode pick their sources from Google's own index, so the same question can cite different pages there (ChatGPT vs Google AI Mode). Ask them the same way, logged out, and count each assistant on its own.

How ChatGPT shows sources: 2026's changes

• January 30: some product and topic names became clickable, opening a panel of facts and links.
• May 7: brand names in answers began linking to the brand's site, mostly the homepage, tagged utm_source=chatgpt.com; software referrals from ChatGPT jumped (Qwairy, 140,000 answers; Profound).
• August 13: in some sessions the Sources button moved into the “More actions” menu.
• October 7: GPT-6 and answers with charts and generated tools; OpenAI hasn't said how sources appear in them.
• October 8–16: many ChatGPT links lost their utm_source tag, so analytics undercounted it for a week.

Questions

Does ChatGPT search use Bing?

Partly. OpenAI says ChatGPT sends its rewritten searches to third-party search providers, naming Bing as one, and that a page must be open to its crawler, OAI-SearchBot, to be shown as a source. Its cited pages overlap only partly with Bing's or Google's results for the same searches.

Can I see which sources ChatGPT cites for my product?

Only by asking: open the Sources panel (or “More actions”, then View sources) on each answer. Because sources change from answer to answer, ask each question several times, logged out, and count.

Why is a competitor's page cited when mine is better?

Usually because its title matches ChatGPT's own search more plainly, it ranks for that search, or it's on the lists ChatGPT reads. Better, to ChatGPT, means easier to match and quote.

Does being cited mean ChatGPT recommends me?

No. A cited page may only supply a fact while the answer recommends someone else. Track whether you're named, and in what position, separately from whether you're cited.

Related: Why ChatGPT doesn't recommend your product · How to rank in ChatGPT · What is AI visibility, and how to measure it