Where AI citations actually come from: 1,793 traced

We track which AI engines mention which brands, which means we keep the sources behind every answer. This is what 1,793 of them look like.

Between 5 and 15 August 2026 we ran the same five buying-intent prompts against ChatGPT, Perplexity, Google Gemini and Google AI Overviews — 42 answers per engine, 168 in total — and recorded every URL each answer cited.

How many sources each engine cites

The engines are not comparable surfaces. One of them does not cite at all.

Perplexity

22.6

Google AIO

12.7

ChatGPT

7.4

Gemini

0 — cites nothing

Gemini's zero is the finding most people get wrong. It is not a sampling artefact or a scraping failure — Gemini's consumer answers do not attribute the pages behind them at all. Every one of its 42 answers named brands and cited nothing. There is no citation to win on that surface, so any report showing you a Gemini "citation share" is measuring something it cannot see.

Perplexity is the opposite case. At 22.6 sources per answer it hands out three times more citations than ChatGPT. If your goal is to be cited rather than merely mentioned, Perplexity is where the odds are best, by a wide margin.

What kind of page gets cited

Every citation was bucketed by URL shape: editorial paths (/blog/, /guides/, /resources/), documentation subdomains (docs., help., support.), community and review sites, and everything else.

More than half of every source these engines reached for was somebody's article. Not a product page, not a docs site, not a review platform — an article, usually a roundup comparing several products at once.

The documentation number is worth sitting with, because "publish great docs" is standard AI-visibility advice. In a commercial, comparison-shaped query, docs accounted for barely one citation in eighty.

Which domains the engines actually read

Two things worth pulling out.

Concentration beats volume. Zapier earned 41 citations from two URLs. Reddit earned 37 from twenty-two. Semrush needed 32 URLs to reach 123. One genuinely authoritative page outperforms a large library of thin ones, and it keeps outperforming it across every engine at once.

Engines have their own reading habits. TechRadar appeared 33 times and only ever on ChatGPT. SE Ranking's Visible subdomain appeared 30 times and only ever on Perplexity. A source that dominates one engine can be invisible on another, which is why single-engine measurement produces confident, wrong conclusions.

The individual pages doing the work

Nine of the top ten are roundups. Not one is a homepage.

How to get cited by AI

The practical consequence of all this is unintuitive, and it is the opposite of what most GEO advice recommends.

Getting into somebody else's roundup beats publishing your own. The pages above are the ones the engines read. A brand named in the Zapier post inherits 39 citations' worth of exposure without ranking anything. Building a page that displaces it means out-ranking Zapier — which, for most companies, is not the cheaper path.

Pick the engine before the tactic. Perplexity gives out three times the citations ChatGPT does. Gemini gives none. "Optimise for AI search" is not one job, and a tactic that works on one surface may have no surface to work on.

Depth beats breadth. The concentration numbers say one page cited forty times is worth more than twenty pages cited twice, and it is much less work to maintain.

Measure before you change anything. Run your own buyers' questions on each engine, many times, and record who is named and what is cited. That is an AI visibility audit, and it tells you which of the pages above matter in your market.

How we measured this

What we did. Five buying-intent prompts about AI visibility tools, run against ChatGPT, Perplexity, Gemini and Google AI Overviews, 42 answers per engine, 168 answers total, 5–15 August 2026. Every cited URL was recorded with its position, then bucketed by URL shape.

What it does not cover. One category — AI visibility tools — and one language, English. Commercial comparison queries are the shape most likely to pull roundups, so the 53.5% editorial share would very likely be lower for informational or navigational questions. Bucketing by URL path is a heuristic: a /blog/ URL is usually editorial, but not always. Engines change their retrieval behaviour without notice, so treat every figure here as a reading from August 2026 rather than a constant. Vidima built this test and sells a tool in the category it measures.

Common questions

Does Google Gemini show its sources?

Not in its consumer answers. Across 42 Gemini answers we recorded zero citations — not a low number, zero. Gemini produces the answer without attributing the pages behind it, so there is no citation to earn on that surface, and any tool reporting Gemini citation share is measuring something else.

Which AI engine cites the most sources?

Perplexity, by a wide margin. It averaged 22.6 sources per answer against Google AI Overviews' 12.7 and ChatGPT's 7.4. Perplexity is where a citation is most winnable, simply because it hands out the most of them.

What kind of pages do AI engines cite?

Editorial content. Blog posts, guides and roundups accounted for 53.5% of all 1,793 citations. Vendor product pages, news and everything else made up 41.6%, community and review sites 3.6%, and documentation just 1.2%.

How do you get cited by AI?

Be named on the pages that already get cited. In our sample a single Zapier post was cited 39 times, and Zapier earned 41 citations from just two URLs. Earning a mention in an existing roundup is a far shorter path than trying to rank a new page of your own.