What we measured, stated before the findings
Forty buyer questions, frozen word for word — sixteen "best X" prompts, eight "X alternatives", eight head-to-head comparisons and eight "how do I choose" questions. Seven engines: ChatGPT, Perplexity, Gemini, Claude, Google AI Overviews, Google AI Mode, and Bing as a proxy for Copilot. Two runs, 3 and 7 September 2026. Coverage was 40 of 40 questions on every engine in both weeks with zero failed checks AS PUBLISHED — but on 2026-09-21 all 80 Bing rows (40 in each week) were retroactively marked failed, because that column had been searching a truncated question. Coverage today is 40 of 40 on the six answer engines and 0 of 40 on Bing. That matters exactly as much as the original sentence claimed it did: a rate computed over answers that never arrived is a rate with the wrong denominator, and for four weeks Bing was such a denominator.
Every number below is reproducible: the panel publishes its results as JSON, including the prompt count, the per-engine coverage and the instrument fingerprint that says whether two weeks are comparable at all. If a prompt or an engine changes, the comparison is marked incomparable rather than being quietly presented as a trend. You can read the week yourself on the public index.
And the limits, up front. Two weeks is two weeks: where we say a finding held, we mean these two runs agreed, not that it is stable. One question set, one market, one moment. The shapes below are worth your attention; the decimals are not load-bearing.
WITHDRAWN 22 SEPTEMBER 2026 — THE BING COLUMN, AND EVERYTHING WE CONCLUDED FROM IT. Our Bing/Copilot integration sent the provider a TRUNCATED question — in effect the first word — for its entire life, so every Bing answer in this panel answered a different question from the one we printed beside it. "Alternatives to Peloton for home cycling" was searched as "Alternatives", and came back with a directory. We found it on 21 September, marked all 80 rows failed, and the public index now shows Bing as a gap rather than a number. Every Bing figure in this article is struck below. The answer-engine findings do not depend on that column and stand; where a combined figure included it, the figure is recomputed here over the six answer engines and labelled.
One further caveat on the recomputed numbers, because it is the honest reason they do not all match the originals digit for digit: the measurement instrument is frozen per week, but our CLASSIFIER is applied at read time, and it has changed since 11 September. It gained an "authority" category for government and academic domains, and app stores moved into "directory". So the category counts below are today's analysis of the same frozen answers — not a re-measurement, and not a quiet edit of the old ones. Where a number moved for that reason rather than the Bing withdrawal, it says so.
There is no shared canon — each engine eats a different diet
The single most useful result is also the most inconvenient one for anybody selling "AI optimisation" as one job. Ask the same question of six answer engines and you get six different source lists, concentrated differently.
- Gemini cited reddit.com for 22 of 40 questions and youtube.com for 20 — more than half its answers lean on community and video.
- ChatGPT's heaviest sources were youtube.com (10 questions), techradar.com (9) and forbes.com (6): review publishers and video, almost no forums.
- Perplexity spread thinner than any other engine — its top domain, forbes.com, appeared in 4 of 40, with reddit.com level at 4.
- Claude cited review publications and almost nothing else; it reached community content least of the six.
- Google AI Overviews and AI Mode were the two most video- and forum-heavy surfaces, and the two least reliant on the long tail.
The practical consequence: "we got cited by AI" is not a fact about your brand, it is a fact about one engine. A tactic that earns you Gemini citations through community presence can be worth nothing on Claude, which in this panel cited community content the least of the six. If you are going to work on this, measure per engine or you will be optimising for an average nobody experiences.
Four out of five cited domains fit no category at all
We classify every cited domain into media, community, social, directory or reference — and we keep an explicit "unknown" bucket rather than forcing a guess. In the published week, 1,414 engine-and-domain pairs produced 1,137 unknowns. That is 80.4%, against 79.9% the week before. (Those pair counts never included Bing — the panel has always kept the proxy column out of its combined totals — so the withdrawal does not move them. The small drift from the 1,142 and 80.8% first published is the classifier change described above, not a re-measurement.)
Read that carefully, because it is easy to over-claim. "Unknown" is our classifier's blind spot, not a property of the web. What is in there is not mysterious: vendor pages, product documentation, niche blogs, retailer listings, the specific page that happened to answer the specific question. The finding is not "AI cites unknown things", it is that the named categories everybody argues about — Reddit, media, directories — describe a minority of what actually gets cited, and the majority is ordinary pages doing an ordinary job.
One more caveat we owe you: the category is assigned per DOMAIN, not per page. A directory listing and an editorial article on the same domain are one row in our data. That is a real limitation and it is why we do not report a media-versus-marketing split at all.
Review directories are well under one percent — and we got the reason wrong twice
"Get listed on G2 and Capterra" is standard advice, and on this panel the directories are a rounding error: seven of 1,414 engine-and-domain pairs in the published week (0.5%) and nine of 1,416 the week before (0.6%) — six distinct directory domains out of 851, and nine out of 858. Two of those are app stores, which today's classifier counts as directories and the original did not.
We originally published a sharper claim than the data supported, and this is the first of two corrections to it. We said directory citations tracked query SHAPE on the answer engines — that "alternatives" prompts pulled them in while "best" prompts almost never did. They do not. On the six answer engines, "best" produced a directory in 6 of 96 prompt-engine cells on 3 September and 3 of 96 on 7 September; "alternatives" in 2 of 48 and 2 of 48. That is the same rate inside the noise, in both directions, in both weeks.
WITHDRAWN 22 SEPTEMBER 2026 — THE SECOND CORRECTION, AND THE MORE EMBARRASSING ONE. Having dropped the shape claim for answer engines, we relocated it: we published that "the shape effect is real, but it belongs to Bing", citing Bing returning a directory for 6 of 8 and 5 of 8 "alternatives" prompts against 2 of 16 and 5 of 16 "best" prompts. That is now withdrawn in full. Those Bing answers were not answers to those questions — the column was searching the first word of each one, so "alternatives" prompts were literally searches for the word "alternatives", and a directory is what that returns. We did not find a property of Bing. We measured our own bug and named it after Bing.
What survives is the level, not the mechanism: directories are a very small share of what answer engines cite, whatever shape the question takes, and we no longer claim any surface where question shape pulls them in. We have no measurement of Bing at all — the column is a gap in the record until it is rebuilt and proved against the question it was asked. The rule we took from this the first time was to re-query a number before repeating it in public. The rule we take from it now is narrower and more useful: check that the answer you are reading is an answer to the question you asked. The provider told us — every response echoed the keyword it had actually searched, and for four weeks nobody compared it to the one we sent.
Being cited and being named are two different things
There is a gap most dashboards collapse. An engine can cite your page as a source without ever saying your name in the answer — and a buyer reads the prose, not the footnotes.
On three of the six — Gemini, Perplexity and Claude — the source that gets cited is almost never the source that gets named: under 3% in every case. Those are close to disjoint outcomes, which has a direct consequence for how you read any AI-visibility report: a rising citation count and a falling mention count are not a contradiction, and a tool that reports one number for both is hiding the more useful of the two. Only the two Google surfaces name what they cite more often than one time in four.
CORRECTED 11 SEPTEMBER 2026, AMENDED 15 OCTOBER 2026. We first published this as 10.9% (365 of 3,347). That figure counted a name as “mentioned” even when the question itself contained it — ask “Booking.com vs Expedia” and an engine repeats both names whatever it cites. Dropping every sample where the cited domain’s name was already in the question gave 8.9%. The old number is still true as “the name is in the text”; it was never true as “the engine introduced the name”, and the second is what the figure is used for. Amended 15 October 2026: the 8.9% still counted the Bing column we withdrew on 22 September; over the six answer engines the figure is the 8.8% above. This paragraph also gave a before-and-after pair for “X vs Y” questions. It was counted with the Bing column too, we could not re-derive it without that column, and it has been removed.
What survives is still a FLOOR in the other direction. We score the visible answer excerpt, which is truncated, so the true rate is higher than 8.8% — by how much we cannot say without storing whole answers, which is a different instrument and would have no back-data. We would rather publish a floor and label it than publish an estimate and imply it.
What we would actually do with this
- Pick the engines your buyers use and measure those. An average across six engines is a number nobody experiences.
- Stop treating directory listings as an AI-visibility tactic. They may be worth it for other reasons; this is not one of them.
- Look at the pages in the long tail for your own category — the vendor pages and docs that got cited are the closest thing to a template you will find. Score one of your own against the signals that decide whether a page can be quoted.
- Track mention and citation as separate outcomes, and say which one moved.
- Freeze your question set before you measure anything, and re-freeze it deliberately. A changed prompt makes last week incomparable, and a comparison nobody flagged as broken is worse than no comparison.
We publish this panel weekly and the raw week is public. If it disagrees with what you measure, we would rather hear it than not: the correction above exists because somebody asked us how we had counted.
Frequently asked questions
Which AI engine cites the most sources?
This panel does not rank the engines by how many domains they cite. What it shows is how each one concentrates: Perplexity spread thinnest (its top domain appeared for 4 of 40 questions), Gemini leaned on reddit.com (22 of 40) and youtube.com (20), and Claude cited community content least of the six answer engines and had the highest share of long-tail domains outside every named category (92.1%). The mix varies enough between engines that a single "AI visibility" score averages away the difference between them.
Do AI answer engines cite G2, Capterra and other review directories?
Rarely. Across two weeks of 40 buyer questions on six answer engines, directories were seven of 1,414 and nine of 1,416 engine-and-domain pairs — well under one percent either way. We previously answered that Bing was the exception; that claim was withdrawn on 22 September 2026 when the Bing column turned out to have been searching a truncated question, so we now have no measurement of Bing to generalise from or to.
If an AI engine cites my page, will it say my brand name?
Usually not. In our measurement the cited source's name appeared in the visible answer text in 8.8% of cases once we stopped counting names the question had already supplied, and on Gemini, Perplexity and Claude it was under 3%. That figure is a floor, since we score a truncated excerpt. The practical point stands: citation and mention are separate outcomes and should be tracked separately.
Is two weeks of data enough to act on?
It is enough to see structure and not enough to see trend. The engine-by-engine concentration and the size of the long tail held across both runs; anything we describe as movement between two runs would be noise. We publish a new week every week and mark any pair as incomparable when the instrument changed.