Measurement

Two AI engines, the same questions, and almost no shared sources

By Arnav Mukherjee, founder of TofuBofu · August 27, 2026

TLDR

  • There's no list of sites to get on. Two engines shared no source on 12 of 34 identical questions.
  • Pick your engines, then measure those. A ranking with no engine attached is describing one engine and hiding it.
  • Reddit isn't preferred, it's abundant. OpenAI cited it 3,012 times out of 491,024 chances.
  • Other people's pages do the talking about you. Your own site was 1.6% of the sources.
  • Treat all of it as a count, not a trend. Our own measurement changed 4 times inside this window.

Someone asked in the AEO subreddit how often AI engines cite anything at all. The category answers that with a league table you've seen a hundred times. Reddit first, YouTube second, LinkedIn third, restated for two years, sourced to studies that never publish a denominator.

We've got a database of stored AI answers, so I counted instead of arguing. The league table didn't survive, and it didn't fail on ordering. It failed because two of our six engines answered identical questions in the same scans and cited almost entirely different websites.

Same question, same minute, two different webs

Google AI Mode and Bing Copilot both hand back a numbered reference list, so both can be counted the same way. Across four scans covering three brands, there were 34 question-cells where both engines answered and both cited something. Same questions, same scans, same minute.

On 12 of those 34, the two engines didn't share a single domain. Mean overlap across all 34 was 0.065. AI Mode contributed 683 sources to that comparison and Copilot 120.

If you take one thing from this: there's no such thing as "the sources AI cites." There are sources this engine cited, for this question, in this minute. Every piece of advice built on a single pooled ranking inherits that error, and most of this category's advice is built on one.

One question, two evidence bases One buying question Google AI Mode 683 sources, 20 per answer Bing Copilot 120 sources, 2.5 per answer Shared: none, on 12 of 34 Four scans, three brands, 34 question-cells where both engines answered and cited.

432 domains, and two thirds cited exactly once

Across 55 scans and 3,031 answer cells we pulled 882 distinct sources pointing at 432 domains. The ten most-cited domains between them accounted for 20.7% of all sources. 292 of the 432 were cited exactly once.

That's a long tail, not a cartel. It's also the reason we publish no ranked list of most-cited domains anywhere on this site, and won't. With a sample like ours a league table would describe four scans while appearing to describe the web, which is precisely what we're objecting to in everybody else's.

Does the Reddit number the category repeats check out?

In our data Reddit was 15 of 699 AI Mode sources, 2.1%, and 0 of 120 on Copilot. YouTube was 52 of 699, 7.4%, and again zero on Copilot. Both held their shape across all four contributing scans, so neither is one scan's artefact: YouTube ran 6.1% to 10.4% and Reddit 1.5% to 2.4%.

We're not saying those are the true global figures. Our sample is small and it's concentrated. We're saying something narrower and harder to dodge: one of the two engines we can measure cited both platforms zero times, so any single Reddit percentage quoted without an engine attached has already thrown away the variable that decides it.

Somebody with a far bigger sample got there first, and by a better route. Dan Petrovic at DEJAN worked from OpenAI's grounding metadata, which uniquely exposes the candidates a model was offered as well as the ones it used. Across 27,351 probes, Reddit was supplied as a candidate 491,024 times and cited 3,012 times.

His words: "OpenAI's models produced only around 3,000 Reddit citations out of the nearly half a million times the domain was supplied as a candidate grounding source." A 0.61% selection rate. Reddit isn't preferred. It's abundant, and abundance is what the league tables have been measuring.

Your own website was 1.6% of the sources

When an engine answered a buying question in a company's own category, it cited that company's own domain 14 times out of 882. And only 48 of the 882 sources, 5.4%, were a bare homepage with no path at all.

Both numbers say the same thing from different angles. Engines cite pages that answer a question, not companies that exist. When one describes you to a stranger, it's reading somebody else's page about you roughly 98 times in 100.

ChatGPT cited nothing in 998 answers, then cited in all five

Before ChatGPT's retrieval change we'd stored 998 answers carrying zero citations. After it, five answers, all five with citations. Five is a direction and not a level, and we won't quote it as a rate. The zero is the strong half, and the mechanism is what matters: an engine answering from memory has no sources to hand you, so there's nothing to optimise for until it decides to look something up.

What can't our sample tell you?

Only 11 of the 55 scans produced any citation at all, across seven brands. Four of those eleven supply 95.8% of the sources and one supplies 54.2%. Three brands carry the engine comparison, and one is a consumer paints brand that sits outside the B2B set the rest of our data describes. It's in there because it was scanned, not because it was chosen.

The count also covers two engines rather than six. AI Mode and Copilot attach a reference list, and ChatGPT returns inline links once it searches. Claude, Perplexity and Gemini store differently in our system, so their totals here would measure our storage rather than their behaviour. We don't publish them as citation rates and neither should anyone quoting us.

And the one almost nobody in this field will tell you: our own instrument changed four times inside this window.

So read every number here as a count, never as a trend. Scans either side of those dates were taken with different instruments, and any movement between them could be ours rather than the engines'. The product itself knows this: a customer whose scan history spans one of those dates gets told on their own trend chart that the step may be us. We'd rather say it here than let you infer a direction we can't support.

One counting decision moves every figure, so it's worth stating. Both engines cite a reference inline and repeat it in the list at the bottom. Raw links give 1,389. Distinct sources per answer give 882. We use 882, because a reader asking how many sources an answer used means distinct ones.

Stop reading other people's samples. See which sources the engines actually cite on your buying questions, in your category, this week.

Run a free AI visibility scan

What should you do with this?

Frequently asked questions

Which websites do AI engines cite most?

There isn't one list, and that's the finding. We counted 882 distinct sources across 432 domains in 55 stored scans. The ten most-cited domains together accounted for 20.7% of all sources, so roughly four fifths of the citations came from outside any top ten. 292 of the 432 domains were cited exactly once. Any article that hands you a fixed league table of the most-cited sites is describing one sample, usually one it won't show you.

Do different AI engines cite the same sources?

Barely. Google AI Mode and Bing Copilot answered the same 34 question-cells in the same four scans, at the same minute, for three brands. On 12 of those 34 they shared no cited domain at all, and mean overlap across all 34 was 0.065. That's the same question, asked at the same moment, producing two different evidence bases. Treating AI citations as one pool is the mistake underneath most advice in this category.

Is Reddit really the most-cited source in AI answers?

Not in our data. Reddit was 15 of 699 Google AI Mode sources, 2.1%, and 0 of 120 Bing Copilot sources. YouTube was 7.4% of AI Mode and again zero on Copilot. We're not claiming those are the real global numbers, because our sample is small and concentrated. We're claiming that a single figure quoted without an engine attached is meaningless, since one of the two engines we can measure cited both platforms zero times.

How often do AI engines cite a company's own website?

14 times out of 882 sources, which is 1.6%. When an engine answers a buying question in your category, it's overwhelmingly reading somebody else's page about you. Read that as a floor for firms that already suspected they were invisible rather than as an industry average, because that's who runs a first scan. A related number points the same way: only 48 of the 882 sources, 5.4%, were a bare homepage with no path. Engines cite pages that answer a question, not companies.

How many sources does an AI answer use?

It depends entirely on the engine. Google AI Mode averaged 20 distinct sources per answer. Bing Copilot averaged 2.5, with a median of 3 among answers that cited anything, so the mean isn't carried by a skew. That's an eightfold difference in how much evidence sits behind an answer, and it means a page has far more room to be one of AI Mode's sources than one of Copilot's.

Does ChatGPT cite sources?

Only when it searches, and for a long stretch it didn't. Across 998 stored ChatGPT answers before its retrieval change we captured zero citations. In the five answers after it, all five carried them. Five answers is a direction and not a level, and we won't publish it as a rate. The mechanism is the useful part: an engine answering from memory has no sources to give you, so there's nothing to optimise for until it decides to look.

Can you trust an AI citation study, including this one?

Ask what it counted and whether the denominator is published. Ours: 55 scans between 24 June and 24 August 2026, 3,031 answer cells, and only 11 of those scans produced any citation at all. Four scans supply 95.8% of the sources and one supplies 54.2%. Three brands carry the engine comparison, and one of them is a consumer paints brand that sits outside our usual B2B set. That concentration is why we publish no ranked list of most-cited domains: with a sample like ours, a league table would describe four scans and pretend to describe the web.

Sources and further reading