Ask a vendor for its second run, and for how the question bank differed

The useful question to put to anybody showing you an AI visibility figure is not where the number came from. It is whether there is a second one. We published 85.6% as the vendor and agency own site share of 10,652 citations, from 597 usable readings on 15 October 2026, and then 82.0% of 20,143 citations, from 1,260 readings on 16 October 2026, with the second question bank written to go into eight directions the first never went. A reader who trusts neither of us can still read those two rows against each other, and that is the whole point of printing both.

One figure asks for belief. Two figures from banks designed differently ask for arithmetic, and arithmetic is a thing a buyer can do alone.

What the second bank changed on purpose

A second run that repeats the first one tests the engines. A second run that changes the instrument tests the finding. Ours changed the instrument in eight directions at once: vertical depth, questions naming a competitor, role and seniority, the problem stage before a buyer knows the category exists, price and procurement depth, comparison with the adjacent search suites, surface depth, and time.

Field	First run	Second run
Date	15 October 2026	16 October 2026
Questions per market	102	210, in 21 batches
Readings usable of attempted	597 of 612	1,260 of 1,260
Citations recorded	10,652 across 2,708 domains	20,143 across 4,947 domains
Vendor and agency own site share	85.6%	82.0%
Engines	ChatGPT and Gemini	ChatGPT and Gemini

Several of the eight new directions were chosen because they should have produced a different answer. Procurement questions and competitor named questions are the shapes most likely to pull in directories, comparison sites and third party reviews. The own site row moved by under four points.

The five things to ask for, in order

These work on any supplier's figure, including ours, and four of the five are answerable in a sentence.

• The counting rule. What one citation was counted as. Ours is one domain in one reading, deduped, and the identical second run counted by raw address occurrence gives 36,576 rather than 20,143. Without this line, two figures cannot be compared at all.
• Attempted and usable. Both numbers. 597 of 612 is a different claim from a run that quietly dropped some attempts and reported only the survivors.
• The engines by name, with a denominator each. Not a category word. Two named apps with 306 and 291 readings behind them is a reading; one blended score across them is not.
• The second run, and how its bank differed. If the answer is that the bank was the same, the second run tested the engines rather than the finding.
• The figure that went against them. A supplier whose published run contains no uncomfortable row has either been lucky or has edited. Ours is a citation count of 3 of 597 and then 6 of 1,260 for our own domain.

That list doubles as a procurement question set, and it belongs in the same part of a buying process as the terms that decide what a software agreement actually covers, which is worked through in the six layers, the buying order and the terms that decide.

What two agreeing figures do and do not establish

They establish that the finding did not belong to the wording of the first bank. That is a real and limited thing.

• They do not make a share a constant. Both readings are dated, and a share belongs to its test rather than to the channel forever.
• They do not extend to a new category. Both banks ask about one market, ours. A different field can sort differently and the only way to know is to count it.
• They do not cover more than two apps. ChatGPT and Gemini in both runs. Our product reads five surfaces, and the three we did not read contributed nothing to either figure.
• They do not turn a source count into a recommendation rate. A domain in a source list and a business named in the answer text are different events, measured separately, and a page that adds them has produced a number describing neither.

Why we publish both runs rather than the better one

Because a single figure from an interested party is worth about as much as the party's reputation, and we would rather be checkable than trusted. Of the 102 build spec conversations we ran on 15 October 2026, 99 asked for a disclosed method and 78 asked for the question set itself to be handed over so a stranger could run it again. Those two requests are satisfied by publishing two runs with their designs, and by nothing else.

We sell this measurement and the publishing work that follows it. The Google Search Console side of the same habit, two domains with their windows and the single export date printed on the page, is our proof page.

Questions about replication in AI visibility research

What were the two own site shares?

85.6% of 10,652 citations across 597 usable readings on 15 October 2026, and 82.0% of 20,143 citations across 1,260 readings on 16 October 2026.

Why not just publish the newer figure?

Because the older one is what makes the newer one checkable. A page that replaces its own numbers has thrown away the only evidence it had that the finding was stable.

Is a four point difference a disagreement?

It is the distance between two banks of questions designed eleven days apart with one twice the size of the other. Both land in the same part of the table and point at the same decision, which is what a replication looks like.

What if a supplier has only ever run one study?

Then ask for the counting rule, the attempted and usable counts and the engine denominators, and treat the headline as one reading on one date until a second instrument has been near it.

Which engines were read in both runs?

ChatGPT and Gemini, on both dates, with the market written into the question text rather than set as a parameter. Our earlier reading is written up at 597 readings across ChatGPT and Gemini in three markets, bank included.

The question that separates a measurement from a brochure

Ask for the second run. The answer tells you whether you are being shown a finding or a result that happened once, and it costs nothing to ask. If the second run exists and its bank was built to break the first one, you can do the comparison yourself and stop relying on whoever handed you the page.