Four of our own readings, three engine sets, and why none of the four numbers may be added to another

We hold four separate AI readings taken in October 2026 and they cannot be summed, averaged or ranked against each other, because three of them were run on ChatGPT and Gemini and the fourth was run on ChatGPT and Perplexity, and because two of them count a different event from the other two. Adding any pair of them produces a figure that describes no run we actually took. This page is the register: what each reading asked, on which apps, on which date, with which denominator, and the one sentence each one is entitled to. A reader checking a claim of ours should be able to find out in thirty seconds which of the four it came from.

The reason to publish the register rather than just follow it is that merging is the easiest thing in this field to do by accident. Four readings in one month, three of them sharing an engine pair, two of them sharing a unit, and every one of them producing a small whole number. The numbers look alike. They are not alike.

The register, as one table

Read the two right hand columns together and the structure of the problem is visible. The engine pair differs on the fourth row. The unit differs on the fourth row. And the scale ranges from 10 answers to 1,260 readings, so even within the three rows that do share both an engine pair and a unit, the denominators are three different sizes.

The two audits share an engine pair and still cannot be added

Audit one recorded our own domain in the source list of 3 readings of 597. Audit two recorded it in 6 of 1,260. Same two apps, one day apart, and the honest way to put those together is side by side rather than in a sum.

The question sets were designed separately and deliberately differently. Audit one asked 102 questions per market. Audit two asked 210 per market in eight directions audit one never went: vertical depth, competitor named, role and seniority, the problem stage before a category is known, price and procurement depth, comparison with adjacent suites, surface depth and temporal. So 3 of 597 and 6 of 1,260 are two independent measurements of the same thing, and what they are good for is exactly that. Audit one put vendor and agency own sites at 85.6 per cent of its citations and review directories at 0.8 per cent. Audit two put the same two classes at 82.0 per cent and 0.7 per cent. Two separately written question sets, eleven days apart in design, agreeing. That replication is the finding. A total of 9 out of 1,857 would have thrown it away and replaced it with a number nobody measured.

The fourth reading is on different apps and counts a different event

The blind naming scan of 16 October 2026 was run on ChatGPT and Perplexity against clawlaw.in, a legal search company built by our founder, and 6 of the 10 answers named it. The questions never contain the company name. That reading sits on our proof page with its pair of worked examples, one question won and one question lost on both apps on the same day.

Two things separate it from the three above it, and either one alone would be enough. The apps are not the same pair, so Perplexity appears in this reading and nowhere else in our data, while Gemini appears in the other three and not in this one. And the event is not the same event: being named in an answer's text is a different outcome from a domain appearing in a source list, the two have different denominators, and they fail for different reasons. The rule is set out at length in being read and being chosen are two measurements, and the practical effect is that no sentence of ours may slide from one to the other.

What a merged number would actually be claiming

• A combined engine coverage we never had. Three apps appear across the four readings and no single reading covered three. Our product watches five surfaces: ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews. A reading of two is a reading of two, and Claude and Google AI Overviews were asked nothing in any of the four.
• A denominator belonging to nobody. Summing readings taken on different question banks produces a total whose questions overlap in ways the total cannot express.
• A unit that does not exist. Source list appearances and text namings added together make a quantity with no definition, and therefore no way to be checked or disputed.
• A date that is not a date. Each reading carries one, and a merged figure would carry a range during which the engines were changing anyway.

How we keep the four apart in practice

A rule nobody can break accidentally has to be mechanical rather than attentive. Ours has three parts and the third is the one that does the work.

• Name the dataset in the sentence. Not the phrase our research, but 1,260 readings on ChatGPT and Gemini on 16 October 2026. The engines enumerated rather than called AI, which 98 of 102 build spec conversations asked for by name.
• One reading per table and per paragraph. A table with a row from two readings in it has merged them visually whatever the footnote says.
• Write the refusal down where a writer will hit it. We already run a house rule stopping two of our own domains appearing in one sentence, published as the comparison we refuse to make on our own proof page. This is the same mechanism applied to readings instead of domains.

Keeping four readings apart: the questions that come up

Can results from two AI apps be combined into one score?

Not into a figure that means anything. Two apps retrieve differently, cite differently and change on their own schedules. A blended score hides which of the two moved, which is the only part you can act on. Report one column per app, each with its own denominator and date.

Why not add the two citation audits, since they used the same two apps?

Because their question banks were written separately and with different aims, so the combined total has a denominator that describes neither run. Kept apart they are a replication, which is worth more than any single larger number: 3 of 597 and 6 of 1,260, with the source class shares agreeing to within a few tenths of a point.

Is the blind scan comparable to the citation audits at all?

Only as a second kind of evidence about the same market, never as a number on the same axis. It used ChatGPT and Perplexity rather than ChatGPT and Gemini, and it counts a business being named in the answer text rather than a domain appearing in a source list.

How many engines do your published numbers cover?

Two per reading. ChatGPT and Gemini in the two audits and the city scan, ChatGPT and Perplexity in the blind scan. Our product reads five surfaces and that is a separate statement from what any one of these datasets measured, which is why the two never appear in the same clause.

Which reading should I quote if I only quote one?

Whichever one measured the thing you are claiming, named in the sentence. For how often a domain turns up as a source, audit two is the largest at 1,260 readings. For whether a business gets named, the blind scan is the only one of the four that measured it, and the full method behind the citation side is published as 597 readings across ChatGPT and Gemini in three markets.

A register is cheaper than a retraction

Four readings in one month is enough to confuse, and the only reliable protection is writing down which is which before anybody needs to quote one. Keeping them apart costs us the single large impressive figure that a merge would have produced, and it buys the thing that figure could never have: any one of our four numbers can be taken apart by a reader without finding a join. We are paid to run these readings for other companies and then to change what the answers say, so the register is also a specimen of the work. If a supplier's headline number cannot be traced back to one named app pair, one date and one denominator, ask which of their readings it came from.