Measurement

Which websites do AI engines actually cite? We counted 882 of them.

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

Someone asked in the AEO subreddit this week how often AI engines cite anything at all. It is a better question than it looks, because the entire category answers it with rankings nobody can reproduce. Reddit is the most-cited source. YouTube is second. LinkedIn is third. The same league table, restated for two years, sourced to studies that report a percentage and never a denominator.

We have a scan database full of stored AI answers, so I stopped arguing about it and counted. Every source link in every answer we have kept, across 55 completed scans run between 24 June and 24 August 2026, covering 3,031 answer cells. That is 882 distinct sources pointing at 432 distinct domains.

The league table does not survive it. Not because the ordering is slightly off, but because two of our six engines answered the same questions in the same scans and cited almost entirely different websites. On 12 of the 34 questions where both produced sources, they shared not one domain between them.

What we counted, and what we cannot count

The method is deliberately dull. Every completed scan stores the full verbatim answer each engine returned. Google AI Mode and Bing Copilot both hand back a numbered reference list attached to the answer, and ChatGPT returns inline links once it searches. We pulled every URL out of that stored text, normalised the host, and counted. No sampling, no weighting, no model in the middle deciding what counts as a citation.

One decision in that method changes every number, so it goes first. Both engines cite a reference inline in the prose and then repeat it in the list at the bottom. Counting raw links gives 1,389. Counting distinct sources per answer gives 882. The second is the honest unit, because a reader asking how many sources an answer used does not mean how many times a footnote marker appeared. The first draft of this piece used the inflated figure and a review caught it, which is worth saying because the correction made the headline gap between the two engines wider rather than narrower.

Here is the part most published citation studies leave out, and it is the part that decides whether you should believe any of this. Three of our six engines are effectively invisible in this count, and that is our instrument, not their behaviour. Claude returns its citations as structured data alongside the answer rather than inside it. Perplexity returns a separate source list that our stored answer text does not carry. Gemini is queried without a search tool at all. Publishing their totals here as citation rates would be a lie with a decimal point on it, so we do not.

So this is a study of two engines, Google AI Mode and Bing Copilot, plus a before-and-after on ChatGPT. Those two are the surfaces that put their working on the page, which is exactly why they are worth reading closely.

And the sample is small and lumpy, which you should know before the numbers rather than after. Eleven of the 55 scans produced a citation, across seven brands, but four scans supply 95.8 percent of the sources and one supplies 54.2 percent. Three brands carry the engine comparison. One of them is a consumer paints company, well outside the B2B firms we usually write about, and it is in the corpus because somebody scanned it. All of it is selected on the problem, because a company runs a visibility scan when it already suspects it is missing from AI answers. A study that does not tell you which engines it could not see, and which handful of tests carried the result, is not being modest. It is hiding its denominator.

Same question, two engines, no shared sources

Start with volume, because it reframes everything after it. Across our stored answers Google AI Mode cited 20 distinct sources per answer. Bing Copilot cited 2.5, with a median of 3 among the answers that cited anything. Same scans, same buying questions, eight times the sources on one surface.

That alone should kill the idea of a single citation share number. Being one of twenty sources is not the same event as being one of three, and averaging the two produces a figure that describes neither.

Then the overlap. We took every question where both engines answered and at least one cited something: 34 of them, drawn from four scans across three brands. AI Mode produced 683 sources on those questions, Copilot 120. The mean overlap between the two sets of domains, measured as the share of the combined pool they had in common, was 0.065. On 12 of the 34 questions the overlap was exactly zero. Two engines, one question, asked in the same minute, and not one website in common.

The 34 shared questions come from four of our own scans across three brands, run between 24 June and 24 August 2026, on a sample selected on the problem. The whole-corpus shares beside each engine are drawn from the wider count of 55 scans.

The two headline domains make the point sharpest. YouTube was AI Mode's most-cited domain, at 52 of its 699 distinct sources across the whole count, 7.4 percent. Copilot cited YouTube zero times in 120. Reddit ran 2.1 percent on AI Mode and zero on Copilot.

Given how lumpy the sample is, the number worth trusting here is not the average but the spread. YouTube's share of AI Mode's sources ran between 6.1 and 10.4 percent in every one of the four scans that contributed, and Reddit's between 1.5 and 2.4 percent. Copilot cited neither in any of them, including the consumer brand. Four tests is not many, but a result that holds separately in all four is not one scan's accident.

Google owns YouTube and indexes its transcripts, so a Google surface leaning on it is not a mystery. The finding is not that YouTube is powerful in the abstract. It is that YouTube is a Google play, and a video strategy sold to you as a general AI visibility tactic is being sold on evidence from one vendor's surface.

The Reddit number the category repeats is not real

The claim in circulation is that Reddit is the single most-cited source across AI engines, at roughly 40 percent of citations. We carried a version of it on this site until this week, sourced to a trade summary of a synthesis with no method published anywhere we could find it, and we removed it this week. Our own data could not get anywhere near it: 15 of AI Mode's 699 sources, and none at all on Copilot.

The one study we can find that publishes both its method and its raw counts agrees. Dan Petrovic measured OpenAI's grounding metadata, which exposes the candidate sources an engine was offered as well as the ones it chose. Reddit was supplied as a candidate 491,024 times and cited 3,012 times, a selection rate of 0.61 percent. Same study, per engine: Reddit is about 2 percent of Google's cited sources, second behind YouTube, and appears zero times across 139,601 Anthropic grounding sources.

Two datasets of very different sizes, different questions, different brands, same shape: Reddit is a minor source on Google's surfaces and essentially absent elsewhere, and YouTube outranks it where both appear. If you have been buying a Reddit programme on the strength of that ranking, you have been buying a strategy for one engine at the price of a strategy for six.

You can see which domains the engines cite on your own buying questions instead of reading somebody's league table. The scan is free and takes about a minute.

Run a free AI visibility scan

432 domains, and two thirds of them cited once

The most useful number in the whole exercise is the least quotable. Those 882 sources point at 432 distinct domains. 292 of them, 68 percent, were cited exactly once. The ten biggest domains together account for 20.7 percent, which means roughly four fifths of the citation surface is tail.

I am not going to publish the top ten as a ranking, because with four scans carrying 95.8 percent of the sources that list describes those four markets and not the world. What is worth reporting is the SHAPE of the tail, which is the same in all of them. Under the recognisable names sit hundreds of mid-tier industry blogs, vendor comparison round-ups, niche trade publications and consultancy explainers. Not one of them is a platform you can join. They are pages somebody wrote that happened to answer the question better than anything else the engine had.

This is the argument against the entire directory-submission genre of AEO advice. There is no list of twenty places to be listed that covers the citation surface, because the citation surface was 432 domains wide in a handful of scans and it grows with every question you ask. What you can do is be the kind of source a long tail is made of: a specific page, answering a specific question, that an engine can lift a sentence from.

1.6 percent of the sources were the company's own website

For every scan we know which company was being measured, so we can ask how often the engine, answering a buying question in that company's own category, cited that company's own domain. The answer across 882 distinct sources is 14 times. One point six percent. That is on a sample selected on the problem, so treat it as the floor for a company that already suspects it is invisible rather than as an industry average.

More than 98 percent of the evidence the engine assembled about a market came from somewhere other than the companies competing in it. This is the mechanism behind every piece of advice we give that sounds like a platitude until you see the count. Your website is where you make the claim. The engine builds its answer out of the places where somebody else confirms it.

There is a second, quieter finding in the same data. Of the 882 sources, only 48 pointed at a homepage, 5.4 percent, and the most common shape by a distance was a URL two levels deep. Engines do not cite companies. They cite pages that answer one question. A homepage answers none.

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

The last column of the count is the one that should change how you buy this software. Until 18 August 2026 our ChatGPT probe answered from training weights with no web search. Across 998 stored answers it produced not one source link. Then we forced retrieval through the Responses API, and every one of the next five answers carried links, 26 distinct sources in total.

Five answers is a direction and not a level, and I will not dress it up as more. But zero out of 998 is not ambiguous, and the switch that changed it was ours, not OpenAI's. The same engine, the same questions, a different call, and a completely different answer about whether ChatGPT cites anybody.

Every AI visibility tool on the market is making that choice for you, silently, and most of them will not tell you which way. It is the single most consequential setting in this category and it appears on nobody's pricing page. Ask your vendor whether their ChatGPT probe searches the web. If the answer is no, or if they do not know, their ChatGPT number is a measurement of a language model's memory, not of the product your buyers actually use.

What to do with this

Stop buying strategies from blended rankings. A single table of the most-cited sources in AI search is averaging engines that shared no source at all on a third of the questions we could compare. Ask which engines, measured how, over which questions, and counted in what unit. Four questions, and most published rankings fail the first.

Pick your channel per engine, not per category. YouTube earns citations on Google's surfaces, in every scan we have that produced any. Reddit is a minor Google play and close to nothing elsewhere. If your buyers are on Copilot, a video budget is not the lever, and you would only know that by measuring your own questions.

Write the page, not the profile. Two thirds of the cited domains appeared once. The citation surface is a tail of specific pages answering specific questions, and being on more directories does not reach it.

Get somebody else to say it. More than 98 percent of the sources in our count were not the scanned company's own domain. Your own site is the claim. Everything the engine quotes back is the corroboration, and if the corroboration does not exist the engine reaches for a competitor who has it.

Frequently asked questions

Which websites do AI engines cite most?

There is no single answer, because the source pool is engine-specific. In our own stored answers, youtube.com was Google AI Mode's most-cited domain at 52 of its 699 distinct sources, or 7.4 percent, while Bing Copilot cited YouTube zero times across 120 sources on the same questions. Below the top few domains the list collapses into a long tail: 432 distinct domains, of which 292 were cited exactly once, and the ten biggest together accounted for 20.7 percent. Read the domain names with care rather than as a ranking: four scans supply 95.8 percent of these sources and one supplies 54.2 percent, and the sample is selected on the problem, because a company runs a scan when it already suspects it is missing from AI answers.

Is Reddit the most-cited source in AI search?

No, and the figure the category repeats is not supportable. In our own stored answers Reddit was 15 of Google AI Mode's 699 distinct sources, 2.1 percent, and none of Bing Copilot's 120, and its share ran between 1.5 and 2.4 percent in every single scan that contributed. That matches the only study we can find that publishes its method and its raw counts: Dan Petrovic measured Reddit supplied as a candidate grounding source to OpenAI 491,024 times and cited 3,012 times, a selection rate of 0.61 percent, and found Reddit at roughly 2 percent of Google's cited sources and zero times across 139,601 Anthropic grounding sources. Reddit matters on Google's surfaces and barely registers elsewhere.

How many sources does an AI answer cite?

It depends on the engine far more than on the question. Across our stored answers Google AI Mode cited 20 distinct sources per answer and Bing Copilot 2.5, with a median of 3 among the Copilot answers that cited anything at all, on the same buying questions asked in the same scans. An eight-fold difference means a brand can be one of twenty sources on one surface and one of three on another, so being cited by the generous engine is a far weaker signal than being cited by the strict one. Count distinct sources, not links: both engines repeat a reference inline and again in a trailing list, and counting the repeats inflates the total by more than half.

Does ChatGPT cite its sources?

Only when it searches, and whether it searches is a property of how it is asked. Until 18 August 2026 our ChatGPT probe answered from training weights with no web search, and across 998 stored answers it produced not one source link. We then forced retrieval through the Responses API, and every one of the next five answers carried links, 26 distinct sources in total. Five answers is a direction rather than a level, but the zero across 998 is not ambiguous. If an AI visibility tool tells you ChatGPT never cites anybody, ask whether it is measuring the engine or its own API call.

Do AI engines cite my homepage or my inner pages?

Inner pages, overwhelmingly. Of the 882 distinct sources we counted, only 48 pointed at a homepage, which is 5.4 percent, and the most common shape by far was a URL two directory levels deep. Engines are not citing companies, they are citing specific answers, so the asset that earns a citation is a page that resolves one question rather than a page that describes your firm.

How often do AI engines cite the brand's own website?

Almost never. Across our 882 distinct sources, 14 pointed at the domain of the company being scanned, which is 1.6 percent, and this is a sample selected on the problem because these companies scanned themselves already suspecting they were missing. The engine is building its answer about you out of other people's pages. That is the whole argument for third-party corroboration, and it is why a content programme that only ever publishes on your own domain moves the number so slowly.

Can I trust a published ranking of AI citation sources?

Only if it tells you which engines it covers, how it collected the citations, and what the denominator is. Our own count shows why: two engines answered the same 34 questions and shared no cited domain at all on 12 of them, with a mean overlap of 0.065. Any ranking that blends engines is averaging pools that barely intersect. Ask four questions of any such list. Which engines, measured how, over which questions, and counted in what unit? A ranking that cannot answer all four is a marketing asset, not a measurement.

Sources and further reading

TofuBofu scan database, 55 completed scans between 24 June and 24 August 2026, 3,031 answer cells, 882 distinct sources recovered from stored answer text. Recounted 27 August 2026. Eleven scans across seven brands produced citations and four of them supply 95.8 percent, a sample selected on the problem.
Dan Petrovic, DEJAN: No, AI doesn't prefer Reddit. Search does.: the selection rate measurement, and the per-engine spread across OpenAI, Google and Anthropic grounding sources.
G2, The Answer Economy: 2026 AI Search Insight Report: 51% of B2B software buyers now begin research with an AI chatbot more often than with Google, from 1,076 respondents surveyed in March 2026.