The four readings this site holds, and what each one is allowed to answer

This site publishes four separate readings, and they are not four versions of one number. Audit one is 597 usable readings of ChatGPT and Gemini, taken on 15 October 2026. Audit two is 1,260 readings of the same two engines, taken on 16 October 2026. A city scan is 120 readings across ten named cities, taken on 15 and 16 October 2026. A blind scan of ten buyer questions was run on 16 October 2026 on ChatGPT and Perplexity, which is a different pair of apps from the audits. Four readings, four dates, two engine pairings. Add any two of them and the result measures nothing, so what follows is the register: what each reading can settle, and what it cannot.

The register, in one table

Every figure below is the figure as measured. None of them has been rounded, combined or carried over from one row to the next.

Reading	Date	Apps	Size	The question it settles
Audit one	15 October 2026	ChatGPT and Gemini	597 usable readings of 612 attempted, 102 questions per market	What kind of source gets cited when a buyer asks a category question
Audit two	16 October 2026	ChatGPT and Gemini	1,260 readings, all usable, 630 questions	Whether audit one holds on a question set twice the size, built in eight directions audit one never went
City scan	15 and 16 October 2026	ChatGPT and Gemini	120 readings, ten cities, six buyer questions each	Who wins a named city, city by city
Blind scan	16 October 2026	ChatGPT and Perplexity	10 questions, 6 of them named the company	What a buyer actually hears when the brand name is left out of the question

Audit one and audit two are the same shape and not the same study

Audit one asked 102 questions in each of the United States, the United Kingdom and India and produced 10,652 citations across 2,708 distinct domains. Vendor and agency own sites took 85.6% of those citations. Review directories took 0.8%.

Audit two asked 210 questions per market, 630 in total, and produced 20,143 citations across 4,947 distinct domains. Vendor and agency own sites took 82.0%. Review directories took 0.7%.

Those two sets of figures sit beside each other and they do not merge. The question sets are different by design: audit two went deliberately into vertical depth, competitor named questions, role and seniority, the problem stage before a buyer knows the category exists, price and procurement, comparison with the adjacent search suites, surface depth and time. A 210 question set and a 102 question set are two populations, and a percentage computed across both would describe a population nobody sampled.

What replication is, and why it is not addition

Here is the one statement that legitimately spans two of the four readings, and the form it has to take. Audit one found our own domain cited in 3 of 597 readings. Audit two found it in 6 of 1,260, which is 0.5%. Two independent question sets, designed eleven days apart, returned the same answer about the same site.

That is a replication, and it is reported as two counts with two denominators sitting next to each other. It is not 9 of 1,857. The moment you add them you have invented a study that nobody ran, with a question set that does not exist, and you have thrown away the only interesting thing about the pair, which is that they were built separately and agreed anyway.

The same discipline applies inside a single reading. Gemini cited us 0 times in 630 readings in audit two and 0 times in 291 readings in audit one, so the honest sentence is that it is 0 in 921 across both, with both denominators present. A single engine column is never averaged into the other one, and the one column per engine framework sets out why a blended score hides the result that matters.

The city scan answers a question the market audits cannot

Audit one measured three markets. A market is not a city, so for months a page about one city could carry only a market figure with a label on it. The city scan fixed that directly: 10 cities, 6 buyer questions each, both engines, 120 readings, all of them usable.

It found that every city has its own winner and the winners are small, city specific domains. In New York the local sources that came back were localseo.nyc at 3 readings, thenycseo.com at 3 and goodfirms.co at 3. In Manchester it was mainchester.co.uk at 3, popula.co.uk at 3 and rank4ai.co.uk at 3. In Mumbai, searchmetric.in at 4. In Bangalore, onecity.co.in at 4 and aiseo.solutions at 4.

And the limit on that reading is as useful as the finding. There are ten cities in it. There is no figure for an eleventh. A page about a city we did not measure carries the market figure and says that is what it is.

The blind scan is a different pair of apps, so it is a different sentence

The ten question blind scan published on our proof page ran on ChatGPT and Perplexity on 16 October 2026. The audits ran on ChatGPT and Gemini. One app is common to both and one is not, which means no sentence may describe the two together.

It also means the scan answers something the audits do not. The audits count what gets cited. The blind scan records whether a company is named at all when the question never contains its name. On clawlaw.in, 6 of the 10 answers named it. On one question both apps named it first. On another, neither app named it, and one of them named two other platforms instead. That won and lost pair, on one domain on one day, is the part a reader believes, which is why it is published together and never apart.

Keep the register before you need it

If you are running your own readings, the register is the artefact to build first, because it is almost impossible to reconstruct later. One row per reading: the date, the apps by name, the question count, the reading count, and one line on what the reading is licensed to say. Without it, two readings taken a fortnight apart start getting quoted as one trend by somebody who was not in the room.

The same applies to the answers themselves rather than the totals. When a shortlist comes back from an assistant, the useful thing to keep is the log of what was named and what was cited, not the score, and keeping the log rather than the number is the practice that makes a second reading comparable to the first.

We sell this measurement and the work that moves it, and the register is published rather than summarised for that exact reason: a company with a commercial interest in a number should be the one making it hardest to misread.

Questions about the four readings behind this site

Why not publish one combined visibility figure?

Because there is no question a combined figure answers. The four readings used two different app pairings, four dates and four question sets of different sizes. A single number built from them would have no denominator anybody could state, and 80 of 102 build spec conversations on 15 October 2026 asked for the sample size N specifically.

How many engines do these four readings cover?

Three apps across four readings, and never more than two in any one of them. The audits and the city scan are ChatGPT and Gemini. The blind scan is ChatGPT and Perplexity. Our product reads five surfaces, and that is a separate fact about the product rather than a description of any dataset here.

What counts as one citation in these figures?

One domain appearing in one reading, deduped. Counting raw url occurrences across the same audit two data gives 36,576 rather than 20,143. Both are arithmetic and only one of them is the published measure, so a citation count quoted from this site is always the deduped one.

Why are 15 of audit one's attempted readings missing?

612 readings were attempted and 597 were usable, so the denominator printed is 597. Audit two attempted 1,260 and all of them were usable. Printing the attempted count beside the usable count is what lets a reader see that the gap was 15 and not 150.

How often does a register like this need rewriting?

Every time a reading is added, and the date of the newest row goes at the top of any page that quotes it. All 102 build spec conversations asked for a date, a last verified line and a refresh frequency, which makes it the single most requested thing in the whole specification and the cheapest to supply.