We measured 2,592 AI answers: which engines cite anyone at all

Between 13 July and 25 August 2026 we ran 186 distinct buyer-shaped prompts against four AI engines on behalf of 19 sites, and stored every answer verbatim along with the sources each engine returned. That produced 2,592 answers that came back successfully, 191 where the engine was unavailable, and 6 outright errors.

The unavailable runs matter and they are excluded from everything below. An engine that did not answer is not a measurement of absence, and counting it as a zero is the single most common way a visibility number gets quietly inflated in the wrong direction.

For each successful answer we asked one narrow question: did the engine cite any source at all, anyone, not necessarily the brand we were measuring. The spread was far wider than we expected.

Finding it by hand was luck. The fix was not, and it is worth describing because it is the part most easily skipped.

A day where an engine cites nobody far more often than it normally does tells you nothing about whether a particular brand lost citations, so it must not enter a baseline. Our breakout detector compares a prompt against its own median citation rate over prior runs; if instrument zeros accumulate in that median, the median sags, and once they are half the history the median hits zero. At that point the detector reports that a prompt has never cited you before, about a prompt that cited you in every real run. It is not an inflated number, it is a false sentence.

So a run is now marked suspect and dropped from both the totals and the baselines when at least half of that engine's answers cited nobody, and that share is at least three times the same engine's own median across at least three other runs, and the run's 95% lower bound clears both. Across 19 sites that rule fires on exactly one incident, the 26 July ChatGPT runs, and on nothing else.

We do not claim to know whether it was an extraction fault on our side or an engine that stopped citing that day. We cannot tell those apart, so we exclude the run and say why rather than pick the flattering explanation.

Every tool in this category, ours included until this month, reports one number: the share of answers that cite you. That number collapses two situations that call for opposite work.

An answer can cite nobody, which means there is no incumbent and the job is to become the source the engine reaches for. Or it can cite other people and not you, which means there is an incumbent, by name, and the job is to displace a specific page. Told only that you appear in 12% of answers, you cannot tell which of those you are facing.

Splitting them on one site in our corpus, across 517 measured answers: 31 cited the site, 400 cited somebody else, and 86 cited nobody. Per engine, displacement ran between 74% and 86%. Almost every buyer question that site cares about already has an incumbent, which is a completely different brief from the one the headline percentage implies.

Nineteen sites is not a market. They are the properties of one operator, weighted toward content publishing, and they are not a random sample of anything. A different vertical would very likely produce different displacement rates, and we would not be surprised to be wrong about the magnitude.

What the sample can support is narrower and, we think, more useful: the per-engine citation behaviour is a fact about the engines rather than about our sites, it is measured over hundreds of answers per engine across many days, and the direction of the differences is far too large to be sampling noise. Perplexity citing on 751 of 751 answers is not a close call.

Every share we publish inside the product carries a 95% Wilson interval for exactly this reason, and a change has to clear a significance test before we call it a change. The figures above are counts over a stated window, not projections, and we have not dressed them up as more than that.

The queries are below. If you run an AI-visibility tool and your own corpus disagrees with ours, we would rather know: the useful version of this field is one where these numbers get argued about with data, not one where every vendor publishes a proprietary score and nobody can check anything.

You can also run the same instrument against your own domain for free at seoaio.ai/audit. It shows the whole result without an account, including the parts where it has nothing to report.