Twice the questions found 4,947 domains instead of 2,708

We ran the same kind of measurement twice, a day apart, with deliberately different question banks. The first ran on 15 October 2026 at 102 questions a market, kept 597 usable readings, and the citation log came out at 10,652 across 2,708 distinct domains. The second ran on 16 October 2026 at 210 a market, kept all 1,260 readings, and its log came out at 20,143 across 4,947. The source pool did not grow between Wednesday and Thursday. The question bank did.

So when somebody shows you a figure for how many sites AI engines cite, the first thing to ask is how many questions were asked. The domain count is downstream of that, and a study reporting one without the other has handed you half a measurement.

The two runs, in the same table

Field	15 October 2026	16 October 2026
Questions per market	102	210, in 21 batches
Questions in total	306	630
Readings attempted	612	1,260
Readings usable	597	1,260
Citations recorded	10,652	20,143
Distinct domains	2,708	4,947
Engines	ChatGPT and Gemini	ChatGPT and Gemini

The second bank was written to go where the first had not: 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 questions about time. A bank that reaches into eight new directions finds domains the first bank had no reason to reach.

Which figures moved and which held

This is the part that makes the comparison worth publishing. Doubling the bank moved the counts and left the proportions almost where they were.

• Counts roughly doubled. Citations went from 10,652 to 20,143 and distinct domains from 2,708 to 4,947.
• The own site share barely moved. Vendor and agency own sites took 85.6% of citations in the first run and 82.0% in the second.
• The directory share barely moved. Review directories took 0.8% and then 0.7%.
• The usable rate improved. 597 of 612 attempted, then 1,260 of 1,260.

Counts are properties of the sample. Proportions survived a bank built to break them, and that is why the proportions are the figures we put in front of a reader and the counts are the figures we date and caveat.

What this means for reading anybody's study, including ours

• Ask for the question count before the domain count. Without it the second number cannot be compared to anything.
• Ask whether the questions are published. Of the 102 build spec conversations we ran on 15 October 2026, 78 asked for the prompt set to be handed over so a stranger could re-run it, and 80 asked for the sample size.
• Ask for attempted as well as usable. 1,260 of 1,260 usable is a different claim from a study that quietly dropped some of its attempts.
• Ask what one citation was counted as. Ours is one domain in one reading, deduped. The same second run counted by raw address occurrence gives 36,576 rather than 20,143.
• Treat two studies with different banks as two studies. They can agree on proportions and still have incomparable counts, which is exactly what ours do.

Sizing your own bank

The practical version is not a formula. List the parts of your business where you could plausibly be strong or absent independently of the others, put several questions in each, and keep adding until a new question stops surfacing a domain you had not already seen. That saturation point is your bank size, and it is specific to your category. Deciding whether to run that yourself or buy it is a separate decision with its own traps, set out in the five silent wrong answers an in house build gives.

Both runs read two apps, ChatGPT and Gemini. Three further surfaces are part of what our product covers and part of what neither run measured. Neither run holds a per industry rate, because the vertical questions named no industry. Anything at city level was measured by a different scan of ten named places and is absent from both.

We are paid for this measurement and for the writing that follows it, so both runs are published including the figures that do not flatter us: 3 readings of 597, then 6 of 1,260, had a page of ours in the source list. The Google Search Console side of the same disclosure is our proof page.

Questions about sample size and source diversity

How many distinct domains did each run find?

2,708 on 15 October 2026 from 102 questions per market, and 4,947 on 16 October 2026 from 210 per market.

Is a wider question bank worth the extra running cost?

It makes the counts larger and it makes a proportion steadier. What it cannot fix is a bank that only asks one kind of question, which is why the second run was written in eight directions the first had missed rather than simply doubled.

How many questions should I ask about my own category?

Enough that a new question stops finding you a domain you had not seen. Our two banks were 102 and 210 per market and the second was still finding new ones, so the saturation point for this category sits above both.

Can I compare my domain count to yours?

Only if the question counts and the counting rule match, and they almost certainly will not. Compare your second reading to your own first reading instead, which is a comparison you control both ends of.

Which engines were in both runs?

ChatGPT and Gemini in both, two apps on both dates. The first run's method, question design and failure modes are published as 597 readings across ChatGPT and Gemini in three markets.

The four directions worth adding to a bank you already have

If your own bank was written in one sitting, it almost certainly leans one way. Four additions did most of the work in our second run, and each one is cheap to write.

• Questions that name a competitor. Not your own name, ever, but a rival's. These surface the comparison pages your category is actually decided on.
• Questions from before the category is known. A buyer describing the problem rather than naming the product. Different vocabulary, different sources, and usually the earliest point you can be present.
• Procurement and price depth. What it costs, what the contract covers, what happens at renewal. These are the questions no operator documents and the ones a buyer has to resolve.
• Time. What changed this year, what is current, what has been replaced. A dated page is the only thing that can answer them.

Adding those four to an existing bank is a morning's work and it reliably finds domains the original set had no route to. Our second run also added role and seniority, vertical depth, comparison with adjacent search suites and surface depth, which is how 102 per market became 210.

Report the bank, then the finding

A domain count with no question count behind it is a number waiting to be misquoted, and the gap between 2,708 and 4,947 on two consecutive days is the size of the misquote available. Put the bank size, the attempted and usable reading counts and the counting rule at the top of your own report, and the finding underneath them. Anybody can then tell what your number is a fact about.