Every number we publish, and how to check it

If we ask you to trust a number, you should be able to take it apart. So this page has where our own AI visibility stood at our last reading, what happened on seven sites we do not own, and the things our numbers cannot tell you. The results that went against us are here at the same size as the ones that did not.

01 — Our own visibility, measured

3 of 4 engines identify and cite us by name

The branded question — “What is CitePack?” — asked of each engine three times, first-party rather than through a middleman. Every individual answer stored.

Perplexity

CITED · 3/3

Identifies us correctly and cites citepack.com in all three samples.

Google (Gemini)

CITED · 3/3

The first engine to resolve us, and steady since.

ChatGPT

CITED · 3/3

On 18 July it returned a different entity with our name. On 28 July it named us and cited citepack.com in all three samples. Those are two dated readings, ten days apart. We shipped things in between and ChatGPT re-indexed on its own schedule, and nobody can separate those afterwards, so we are not going to tell you what caused it.

Claude

COLLIDING · 0/3

Still reaches a 2007 academic paper of the same name. An identity problem, not a content one.

And the number that matters more. Asked for the best tool in our category, without our name in the question, no engine named us in any of the 12 probes. Recognized, not recommended. That is the harder of the two. Our own reading is that it is won mostly on pages we do not own, and we are not going to claim it before we have it.

Measured 2026-07-28 · 36 probes · m=3 per query family. Small samples, reported as counts and never as rates or trends. An engine we could not probe is recorded as not measured, never as a zero.

02 — The instrument on sites we do not control

Seven sites, measured twice, two weeks apart

A measurement you can only ever pass is not a measurement. So before we asked anyone to trust ours, we ran it across seven live sites nobody here owns, twice, between 6 and 17 July 2026, and published whatever came back. That is 28 engine-level pairs, a before and an after for each.

Site & engineBeforeAfter
stripe.comPerplexitynot citedcited
trello.comGoogle (Gemini)citednot cited
trello.comPerplexitycitednot cited
notion.so, linear.app, and four moreall four enginesno change

One reading improved. Two got worse. The rest held.

What this shows, precisely. We did not modify any of these sites — we measured them. So this is evidence that the instrument is sensitive to real change in the world and reports it in whichever direction it goes. It is not evidence that CitePack causes citations, and we keep that line sharp on purpose. A vendor who blurs it is selling you a result they cannot produce.
03 — Why a single check proves nothing

The same question, two minutes apart

Answer engines are non-deterministic. Asking an identical question twice on the same day, up to 60% of the cited sources come back different. This is the single most important fact about measuring AI visibility, and it is why we sample rather than check.

A tool that shows you a confident score from one question is not wrong occasionally. It is wrong in a way it cannot detect. The full method

What our numbers do not show

  • We cannot promise an engine will cite you.Nobody controls what a model retrieves. We measure the current state, ship the fixes that make you resolvable and quotable, and re-measure afterwards — reporting the result either way, including when nothing moved.
  • We cannot tell you our pages cause citations.Our own branded reading changed between 18 and 28 July, and in those ten days we shipped several different things while the engines re-indexed on their own schedule. Nobody can separate those out afterwards, so we do not try. You get the dates and the readings, and you draw your own line.
  • Small samples are counts, never trends.Three samples per engine per question tell you the state today. They do not support a percentage, a trajectory, or a forecast, and we do not render one.
  • Unmeasured is reported as unmeasured.Never quietly counted as a no. On a report you paid for, a fabricated negative is worse than no answer.
  • Everything here is stored and recomputable.Each probe — engine, question, sample, answer — is written to an append-only log, because an engine’s index cannot be rewound. Nothing on this page is an estimate, a projection, or a borrowed vendor statistic.

Now run it on your own site

The free scan reads your page, scores the seven on-page signals, and asks Perplexity one question about your category, then tells you whether your domain was among the sources it cited. The other three engines, and the three question families on this page, come with the $49 report. Free, no signup.

See if AI cites your site.

Paste a URL. We ask a real engine what your customers ask, show you the answer it gave and who it named, and — if something is stopping it reading you — the line in your own file. Nothing is held back.