What sources AI engines cite: 1,019 answers for 18 websites

By Abiot Y. Derbie - Founder, RankEcho - Updated 2026-11-02

Key findings

The figures describe this sample: the questions each audit asked, the engines' APIs, and the dates. They show what the engines did, not why.

• Being named is most of the battle. When the question named the site, 64% of answers cited it. When it asked for the best tool, alternatives to a rival, or help with a problem, 15% did.
• Engines rarely use the same sources. For the same question, two engines' cited domains overlapped by 8% to 15% on average.
• Community sites are a Perplexity and Gemini habit. YouTube and Reddit were 7.1% of Perplexity's citations and 3.5% of Gemini's, but 1 of 1,581 for Claude and 1 of 546 for ChatGPT.
• The long tail is long. 2,136 of 3,227 cited domains (66%) were cited only once.
• Gemini hides many of its sources. 42% of Gemini's citations were Google redirect links that do not show the site behind them.

The sample

The study counts every finished audit RankEcho ran between 2026-06-06 and 2026-11-01 for 18 websites, 36 audits in all, excluding RankEcho's own site. Each audit asked a set of buyer questions generated for that site, through each engine's API. That gave 1,019 answers to 437 questions and 9,504 cited links.

The sites are mostly small businesses and software companies, several of them outside the US. Most were audited once. Not every engine was asked for every site: ChatGPT answered for 3 sites and Claude for 7, so their figures rest on fewer sites and should be read with more caution. Google AI Overviews were left out because their stored results are not reliable enough to count, and so were answers where the engine's web retrieval was skipped.

Only totals across all sites are published. No site, question or answer from the sample is named, and the audited sites' own domains are left out of the domain table below.

Finding 1: questions that name you get you cited

Each audit mixes questions that name the site with questions a stranger would ask. The gap between them was the largest effect in the data.

The gap holds on every engine

On every engine, a question that named the site was several times more likely to be answered with a link to it. ChatGPT showed the widest gap: 76% on named questions and 7% on the rest, from only 3 sites.

The questions that do not name you are the ones new customers ask. A site's citation rate on its own name says little about whether engines recommend it to strangers, so report the two separately.

Finding 2: each engine reaches for different sources

Share of each engine's resolved citations by kind of site. Google redirect links are left out of Gemini's base.

What the source mix suggests

Perplexity and Gemini search the open web broadly and pick up video and forum threads. Claude leaned towards research papers and almost never cited YouTube or Reddit. ChatGPT cited the fewest sources per answer.

So work that helps on one engine may do nothing on another. A useful YouTube video or an honest Reddit answer can reach Perplexity and Gemini answers; a well-referenced page with clear facts is a better bet for Claude. Treat these as hypotheses to test on your own questions, not rules.

Finding 3: the third-party sites that came up most

These domains were cited for the most sites in the sample, out of 18. They are the closest thing to sources that matter across topics; nearly everything else was specific to one site's subject.

Finding 4: engines rarely agree

When two engines answered the same question, the sets of domains they cited overlapped little. The overlap below is the Jaccard index: domains both cited, divided by domains either cited. Most pairs shared at least one domain, usually the site itself or a large platform, but the rest of their sources differed.

What this means for your site

• Measure questions that name you and questions that do not separately. The second group is where new visitors come from, and it is much harder.
• Report each engine on its own. With overlap this low, a blended score hides where you are found and where you are not.
• Look at the sources cited instead of you on the questions you care about. In this sample most were specific to the topic, so a generic list of sites to be on is a weak substitute for your own data.
• Be careful with Gemini source lists: many links hide the site behind a redirect.

Limits

This is an observational sample of RankEcho audits, not a random sample of websites or questions. Each site's questions were written for that site, so question mixes differ between sites. Engines change their models and search behaviour often, and the same question can be answered differently on the next run. The figures do not show why an engine chose a source, and they should not be read as a prediction for another site, engine or date.

Source kinds come from RankEcho's domain classifier and a short list of known domains, so some sites may be counted in a broader group. The aggregate table can be downloaded below; the row-level answers are not published because they belong to the audited sites.

Frequently asked questions