What sources does ChatGPT cite for product recommendations?
For buying questions, ChatGPT cites a short list of third-party pages: review directories, vendor comparison pages, Reddit threads, "best X" lists and a few publishers. The brand's own site is a small share. In our ChatGPT audit of one B2B service category, 200 citations split into review sites 70, vendor pages 48, Reddit 42, news 12, listicles 10 and owned pages 9; one directory, clutch.co, took 58 of the 200 on its own[1]. Across engines the mix shifts: Reddit was pulled into 43 percent of Google AI Overview answers in our tracking but 19 percent of ChatGPT's[1].
How ChatGPT chooses what to cite
When ChatGPT answers a buying question with search on, it rewrites the question into a few searches, retrieves a handful of pages, and writes the answer from those pages with inline citations. It does not consult a ranking of brands; it reads what the retrieved pages say and names whoever they name. That is why the same brand can be recommended in one run and missing in the next: a different page was retrieved[2].
Two consequences follow. The pages it retrieves are mostly not the brand's own, and they are mostly not the pages ranking first on Google. Only about 12 percent of the URLs AI engines cite rank in Google's top 10 for the original prompt[3]. Google's AI Overviews are the exception, pulling about 38 percent of their citations from the top 10[4].
The source types, with counts
Our own audit ran 25 buying questions about Reddit marketing and GEO agencies through ChatGPT and classified every cited URL. Of 200 citations[1]:
| Page type | Citations of 200 | What it looks like |
|---|---|---|
| Review directories | 70 | Clutch, G2, DesignRush and similar profile pages; clutch.co alone took 58 |
| Vendor pages | 48 | Competitors' own service and comparison pages |
| Reddit threads | 42 | "Anyone used an agency for X?" threads, often one to two years old |
| News and trade press | 12 | Coverage of the category, rarely of a single vendor |
| Best-of lists | 10 | "Top 10 X agencies" articles on marketing blogs |
| The brand's own pages | 9 | Home and service pages, when they state plainly what the company does |
The split is for one category and one engine, so treat it as a shape, not a law. In consumer categories YouTube reviews and retailer pages take a bigger share; in software, G2 and Capterra comparisons do. The pattern that holds everywhere: a few third-party pages carry most of the weight, and the brand's own site is a minority of citations.
Why one list page matters more than a hundred posts
Citations concentrate. Across 28 buyer questions in our category tracked daily for 30 days, one "best GEO agencies" article was cited 562 times and another 628 times[1]. Those two pages decided which agency names appeared in thousands of answers. A brand that is on them is recommended; a brand that is not is invisible, however good its own site. The same concentration shows up in every category we have audited: the top five cited URLs usually account for most of the recommendations.
The mix differs by engine
Engines do not read the same internet. In one 30-day run of 31 prompts on three engines, the share of each engine's answers that pulled in the same three sources looked like this[1]:
| Source | Google AI Overviews | ChatGPT | Gemini |
|---|---|---|---|
| 43.4% | 19.1% | 24.3% | |
| YouTube | 38.9% | 2.0% | 1.7% |
| 36.3% | 5.1% | 0% |
Reddit is the most cited domain in AI answers overall[5], and in our category it was retrieved in 31 percent of all answers[1]. But a YouTube review that moves Google's AI Overview does almost nothing for ChatGPT. If your buyers use one engine more than the others, the source mix for that engine is the one to work.
What ChatGPT does not cite
- Ads. Paid placements are labelled and the engines do not retrieve them as sources.
- Pages it cannot read. Sites that block OpenAI's search crawler in robots.txt, or that render their text only after JavaScript runs, do not get cited[6].
- Removed content. A seeded Reddit comment that moderation or Reddit's LLM-based detection removes stops being cited the next time the engine re-reads the thread.
- Pages with nothing quotable. A home page that opens with a slogan gives the engine nothing to lift. ChatGPT takes 44 percent of its citations from the first third of a page[7].
How to find the sources for your own category
- Write 10 to 15 questions your buyers ask before choosing. Use their words, not your positioning.
- Run each on ChatGPT, Gemini and Google AI Mode several times and save every cited URL.
- Group the URLs by page type and count them. The top five pages are where your category's answers are written.
- Mark which of those pages name you. The ones that do not are your target list.
Our free audit does this on five engines with repeated runs and sends back the mention map and the source trace in 48 hours. What to do with the result.
Frequently asked
Does ChatGPT cite the brand's own website?
Sometimes, and it is a small share. In our audit, owned pages were 9 of 200 citations against 70 for review directories. Owned pages get cited when they answer the question plainly in the first paragraph, which is why rebuilding them is part of the work, but most of the answer is written from third-party pages.
Is Reddit really the biggest source for ChatGPT?
Reddit is the most cited domain in AI answers overall, and in our category it was retrieved in about 31 percent of answers across engines. For ChatGPT alone the share was lower (19 percent of answers in one 30-day run) than for Google AI Overviews (43 percent), so the right answer depends on the engine your buyers use.
Do Google rankings decide what ChatGPT cites?
Mostly not. Only about 12 percent of the URLs AI engines cite rank in Google's top 10 for the same prompt. Google's own AI Overviews are the exception, with about 38 percent of their citations from the top 10. ChatGPT runs its own searches and retrieves a different set of pages.
How do I see which sources ChatGPT cites for my category?
Run your buyers' questions with search on, open every citation, and tally the URLs by page type. Do it more than once per question because retrieval varies. Or request our free audit, which runs the same method on five engines and sends back the source list.
Sources
- Redaccel tracking in Peec AI, 28 non-branded questions about GEO and Reddit agencies, ChatGPT, Gemini and Google AI Overviews, 25 August to 24 September 2026, about 2,600 answers; the engine split from a 30-day run of 31 prompts on three engines; and Redaccel's own ChatGPT audit of redaccel.com, 25 prompts, 200 citations. Method on GEO agencies compared and the audit checklist.
- Jiang, Zhang, Xing, Chang and Xiao, "Understanding ChatGPT Citations: Evidence from Production Web Search Across Model Generations", arXiv, October 2026.
- Ahrefs, "Only 12% of AI Cited URLs Rank in Google's Top 10 for the Original Prompt", 2025.
- Ahrefs (Louise Linehan), "Update: 38% of AI Overview Citations Pull From The Top 10", March 2026.
- Semrush, "The Most-Cited Domains in AI: A 3-Month Study"; corroborated by Statista, 2026.
- OpenAI, crawler documentation (OAI-SearchBot, GPTBot, ChatGPT-User), checked October 2026.
- Search Engine Land, "44% of ChatGPT citations come from the first third of content".