Across 4 companies in 4 different sectors, placements in the first five products an AI assistant names went from 19 of 570 buyer questions to 460 of 570. Mentions of any kind went from 40 of 570 to 509 of 570. The four had nothing in common except the method: a legal technology company, a direct to consumer retailer, a technology services firm and a digital marketing agency.
The four runs
- Claw Legaltech, legal technology, india. First five: 12 of 200 to 171 of 200. 377 pages published, first citation 14 days after publishing.
- An Indian direct to consumer retailer, ecommerce. First five: 4 of 150 to 118 of 150. 240 pages, first citation in 11 days.
- A technology services firm, it and technology services. First five: 2 of 120 to 92 of 120. 190 pages, first citation in 16 days.
- A digital marketing agency, marketing services. First five: 1 of 100 to 79 of 100. 160 pages, first citation in 9 days.
The three things that were the same every time
They all started lower than they believed. Every one of these businesses ranked well on Google for its own category. All four were close to absent when an assistant was asked the same thing in plain language. Ranking and being named are different jobs, and a strong search position tells you almost nothing about the second one.
This was the single most common surprise. A marketing team that reports green on every search dashboard can be invisible in the place buyers are now asking, and nothing in their existing reporting would show it.
The first movement came from pages that answered one question each. Not service pages and not brochures. One page per buyer question, with the answer in the first paragraph and the specifics underneath. The first citation arrived between 9 and 16 days after publishing in every case, which is far faster than the same pages took to rank in ordinary search.
Specific beat polished. The pages that got quoted carried a real count, a named list or a dated price. Pages written to sound impressive were read and passed over, because a vague page gives an assistant nothing it can repeat.
The six kinds of page that earned citations
Across all four runs, the pages that got cited fell into six shapes. This list is the most directly useful thing in this study, because it is what we would write first for a new client.
- A price page with real numbers and a date. Even a range, even with conditions. In three of the four runs a pricing page was among the first pages cited.
- A named coverage or scope list. Which courts, which regions, which integrations, which certifications. Named, counted and dated, never "comprehensive".
- A "how to choose" page that does not conclude with the client. Written to help somebody decide, including when the answer is a competitor or no purchase at all.
- A limits page. What the product or service does not do, and who should buy something else. Quoted more often than we expected, because a cautious recommendation wants it.
- A documented method or process page. Step by step, with what goes wrong. Assistants prefer a source that shows its working.
- A page answering the awkward commercial question the sector normally avoids in public: total cost, what the guarantee excludes, why a quote varies.
What was different between them
The technology services firm was slowest and finished lowest, at 92 of 120. Its category is crowded with analyst pages and large consultancies, and assistants reach for those first. The marketing agency moved fastest, from 1 of 100, which is roughly where most service businesses sit without knowing it.
So the honest read is that the method worked in every sector we tried it in, and the ceiling depends on who is already trusted in that category. A business competing with government portals or large analyst firms should expect to do well on the buying questions and to keep losing the purely informational ones to sources that deserve them.
A worked example of the difference specificity makes
In the legal technology run, two pages were published in the same week on closely related subjects. One described the product's coverage in general terms. The other named the courts and tribunals covered, stated the count, and dated it. The second was cited. The first was not.
Nothing else about them differed materially: same domain, same week, same subject area, similar length. The difference was that one of them gave an assistant a sentence it could repeat and stand behind, and the other gave it an adjective.
What we are not claiming
This is not a controlled experiment. Four companies is four companies, and nobody ran a matched group that published nothing. The assistants themselves changed during every one of these periods, and this data cannot separate our work from their changes.
We are also the vendor reporting on its own clients, which you should weigh accordingly. And three of the four are described by sector rather than named, because we do not have permission to name them, which is a real limitation on how much you can check.
Rather than argue about it: the scan on this site is free, it runs the same method, and it will tell you your own starting number in a few minutes. That number is the only one about your business that matters.
Common questions
Why are three of the four companies not named?
We do not have permission to name them. Claw Legaltech is named because it is our own sister company. We would rather publish an unnamed case with real figures than name nobody and publish nothing, but you should discount an unnamed case accordingly.
Can I see the prompt sets?
Not yet. They were built and run by hand before this product existed and are being prepared for publication. Until then these are a vendor's own figures. The free scan gives you your own, which you do not have to trust anybody for.
Is 460 of 570 a realistic expectation for my business?
No, and you should be suspicious of anyone who says it is. That figure is four companies that did the work across sectors with different ceilings, and the weakest of the four finished at 92 of 120. Your ceiling depends on who already owns your category.
How many pages does this take?
These four published between 160 and 377 pages. That is more than most businesses expect, and the count matters less than the shape: one page per real buyer question, with something specific in it. Fifty pages that each answer a real question beat five hundred that do not.
What happens if we stop?
We do not have data on that, because none of these four stopped. Our honest expectation is that existing pages keep working while they stay accurate, and that a page with a stale price or an out of date claim eventually becomes a liability rather than an asset, because it can be checked.
Does it matter that you are the vendor?
Yes. Every assistant we tested discounts vendor material, and several said so explicitly. That is exactly why the method leans on checkable facts rather than persuasion, and why the most valuable page in any of these runs was usually the one that helped a buyer decide without pushing them towards the client.
What to do first
Run the free scan and get your own baseline, because every number above is about somebody else. Then take your ten most valuable buyer questions, ask them of two assistants without naming your company, and write down which sources get cited. That list of sources is your real competitive set, and it will tell you whether your next page belongs on your own site or on somebody else's.
How all four were counted
The tracked set is 200 questions, and that is the denominator under every number on this page. The questions are the ones a buyer in this sector actually types, and not one of them names the company. That rule is the whole measurement. A question that names the business gets a polite answer about the business, so it tells you about your prompt rather than about your position.
Each question was asked on 5 assistants, because they answer the same question differently on the same day from different sources. "Named" means the company appears anywhere in the answer. "In the first five" means it appears among the first five products the answer lists, which is the number that matters, because an assistant that names three or four options has no sixth place to give you.
Baseline reading: 27 July 2026. Current reading: 26 September 2026. Pages published in between: 377.
The four things we recorded for every question
- Whether the company was named at all, anywhere in the answer, including in a list rather than as a recommendation.
- Whether it was in the first five products named, and in what position.
- Every source the answer cited, by domain, tagged as official, the company's own site, a competitor's own site, a directory or listing, news, or a community thread.
- Whether the assistant actually ran a live search, asked directly and recorded as a count, because an answer given from memory and an answer given from a live search are two different measurements and averaging them produces a number that means nothing.
What this is, and what it is not
These runs were operated by hand, before this product existed, using the same method the product now runs on a schedule. They are our own results on our own clients, so read them as a vendor reporting its own work. The prompt sets are being prepared for publication so that anyone can repeat the run and check us.
They are also not a controlled experiment. Nobody ran a matched group that published nothing, so we cannot separate our work from everything else that happened in those months, including the assistants changing what they do. What the numbers show is that companies which did this work moved, not that nothing else could have moved them.