Written out in full, because a measurement product that will not show its method is asking to be taken on faith.
Each measurement uses a corpus of buying questions for your vertical — the phrasing customers actually use when asking an assistant for a recommendation, not keyword strings. For every non-English market the questions are authored natively by a speaker of that language rather than translated from the English set.
This matters more than it sounds. German furniture buyers search Nachbau and Dutch buyers say namaak; neither is what a dictionary returns for "replica". A translated corpus measures the translation, and quietly reports a market you did not ask about.
A citation rate is never a property of a brand alone. It is a property of a brand and a set of questions, and where those questions come from changes the number — on one brand we measured, by 24 points in one market and by 2 in another, with nothing else varied. Our third paper names this parameter prompt provenance and proposes that every visibility figure travel with four facts. We publish ours here, on every report, and we think every product in this category should.
| Disclosure | AnswerRange |
|---|---|
| Provenance class of the prompt set | Machine-generated from category search demand by default (P2). Observed search queries from your own Search Console when connected (P3). Questions you write yourself before the run (P1), labelled as yours. Every report states which it used, and how many of each. |
| Prompt count, and whether it changed since last period | 12 per market per run, fixed between runs so periods are comparable. A change of corpus is recorded on the report and the comparison is reset. |
| Replicate count per prompt per period | 2 per platform; 1 on ChatGPT. |
| Whether branded queries are included | No. Queries containing your brand, including misspellings within one edit, are removed before a Search Console corpus is built. An assistant naming you in reply to your own name is not visibility. |
Observed-query corpora have a floor: a site with too little search demand cannot support one, and is measured on the category corpus instead — and told so. A report never silently substitutes one class for the other.
Five are measured: ChatGPT, Google AI Overviews, Google AI Mode, Gemini and Perplexity. All are reached through provider APIs and Google's AI Overview results. These are not identical to the consumer apps — personalisation, session history and app-only retrieval layers differ. We state that rather than bury it, because it bounds what the number means.
Every question is asked twice per platform, once on ChatGPT. This is not redundancy: these systems do not answer identically twice. Our audit found one platform contradicting itself on roughly 40% of identical repeated questions. A single ask would report that coin-flip as fact — and a product that repeats nothing cannot tell you how much of its own month-over-month movement is the platform changing its mind.
An answer counts as citing you when your registrable domain appears among the sources the platform
attributes. Domains are normalised through a public-suffix rule, so shop.example.co.uk
resolves to example.co.uk and not to the bare suffix. Brand mentions without a link are
not counted — they are real, but they are a different measurement, and conflating the two inflates
the figure.
Citation rates are reported with a 95% confidence interval computed by bootstrap resampling over questions rather than over individual answers. Resampling answers would treat three repeats of one question as three independent facts and report an interval far narrower than the measurement deserves. Clustering on the question is the conservative choice and the correct one. When every answer is a miss — or every one a hit — the bootstrap has nothing to resample and would report a zero-width interval, which is a claim of certainty a zero cannot make; at that boundary the exact binomial (Clopper–Pearson) interval on the raw counts is reported instead, so 0 of 108 reads as 0–3.4%, never as 0–0%.
A movement between two measurements is tested with a two-sided Fisher exact test on the counts, not by eyeballing whether two bars overlap. Overlapping intervals routinely hide a real difference, because the comparison has its own, narrower sampling distribution; the test decides, and the report says when the bars disagree with it and why that is not a contradiction.
A report compares five platforms at once, and five tests at the usual threshold give a 23% chance that at least one comes back "changed" when nothing did. So every comparison in a report is corrected together, Holm–Bonferroni, holding the chance of any false alarm across the whole report at 5%. Everything that does not clear that is labelled within noise. Telling a merchant they dropped when they did not is the expensive mistake, because they will spend money reacting to it.
If the corpus changed between two runs — you connected Search Console, say — the runs are not compared at all. The later one becomes a new baseline and the report says so. Changing where the questions come from moves the number by itself, and we will not report that as a change in your visibility.
Three papers, open access with their datasets — 5,382 observations and the analysis code, under CC BY 4.0. The method above is the method in them.