Which businesses benefit most from GEO? A 100 question, five category test

The honest answer is a count, not a list of industries: the businesses that benefit most from GEO are the ones whose buyers ask questions where an engine both recommends a supplier and cites a page to support it. So the figure this page is built around is the number of tested responses in which a business was both recommended and supported by a citation, stated as x of n. We have not run the 100 question, five category test as of 29 September 2026, so we publish no such count over 100 questions, and we will not put a placeholder where a measurement should be. The nearest real counts we hold are these. In AI visibility tooling, on our own domain, September 2026, Perplexity: 0 of 6 responses recommended or cited aiknowsus.com, so the recommended and cited count in that sample was 0 of 6. In legal technology, clawlaw.in, 18 August 2026, ChatGPT: the company was the top source on 1 of 18 blind commercial questions, and in that one answer its own comparison page was named in the answer as still being the vendor's own editorial page.

Those two fractions are small and they are real, and a small real fraction is more useful to you than a confident industry list. If somebody tells you that legal services, healthcare and home improvement benefit most from GEO, ask them for the numerator, the denominator, the engines and the dates. Without those, they are describing a hunch.

How it was measured

The test has five categories of twenty questions, run on four engines, three times each. That is 100 questions times 4 engines times 3 runs, which is 1,200 responses. The arithmetic is printed here because a rate over 1,200 responses and a rate over 100 questions differ by a factor of twelve, and a report that does not say which one it used cannot be compared with anything.

The five categories, and what each one is for. A business does not benefit from GEO in general. It benefits in the categories where its buyers actually ask, so the split is the finding.

• Category questions. Best or top supplier in a category, in a country or city. This is where a recommendation is most likely and where being absent costs the most.
• Problem questions. The symptom in the buyer's own words, before they know your category exists. Engines often answer these with advice and no supplier at all, which is itself a finding about your sector.
• Comparison questions. Two or three named suppliers set against each other, with your own brand never named in the prompt.
• How to questions. How to do the job without buying anything. Expect official sources to win here, and expect that to be correct. On 17 September 2026 Claude said plainly that no commercial product should rank above the official portal for a question about using that portal, and in that run official court portals took every position above any commercial product on the how to questions.
• Commercial questions. Price, contract length, total cost, what is excluded. These are the questions most sectors avoid answering in public, which is why a page that answers them tends to be the only thing available to quote.

The scoring rule that produces the headline count. Recommended and cited together is one state and it must be scored as one. A response counts only when both of these are true in the same response: the business is put forward as an answer to the question, and a page on that business's own domain appears as a source supporting it. Score the four neighbouring states separately, because they lead to different work.

• Recommended and cited. Both. This is the headline numerator.
• Recommended, not cited. Named as an answer with no page of yours behind it. Real and common. On 17 September 2026 Claude confirmed that clawlaw.in pages had ranked in its raw results and shaped what it wrote, while the site landed as a name inside a list rather than as a linked recommendation, because its own comparison pages read as vendor advocacy.
• Cited, not recommended. Your page is used as a source for an answer that recommends somebody else. This is the state that tells you your content is good and your commercial positioning is missing.
• Named without either. A mention in passing.
• Used without attribution. The argument is yours, the credit is nobody's. On 17 September 2026 Claude admitted it had used two specific arguments drawn from clawlaw.in pages and dropped the attribution in both places, calling it a citation lapse rather than a ranking judgement.

The blind rule, which decides whether the whole test is worth anything. The business being measured is named in none of the 100 questions, and in no wrapper or system instruction. We have the cost of breaking this rule measured. On 27 July 2026, two runs of the same 78 questions happened on the same day for clawlaw.in. The run whose wrapper named the brand came back ranking it first on almost every question. The blind run put the company second by breadth and absent altogether from the litigation due diligence questions it most wanted to win. The first run was discarded.

The fields to log, one row per response. Question identifier, question text, category, engine, run number, date and time with time zone, country and city, language, signed in state, every business named in order, every URL offered, the five states above, whether the engine claimed to have searched, the labeller's name, and a link to the saved capture. Twelve hundred rows. The sheet is the test.

Search claims are recorded as claims, never as facts. Two dated reasons. On 27 July 2026 ChatGPT, asked 78 blind questions, then confirmed it had run no live web search for any of them. And in three separate batches of the September 2026 aiknowsus.com audit, Perplexity withdrew its own earlier statement, saying it could not honestly substantiate the claim that it had run a live search for each question, and in another batch that its claim to have searched all five was not adequately supported.

What the numbers were

Every count we hold, with its sector attached. Read the sector. A legal technology fraction is not evidence about a clinic, a manufacturer or a marketing agency, and moving it across sectors is the exact error that makes published GEO figures useless.

• Recommended and cited: 0 of 6 responses. AI visibility tooling, aiknowsus.com, September 2026, Perplexity. Six blind questions in our own category. The engine audited its own answers and reported it had cited or recommended us in none of the six.
• Top source: 1 of 18 responses. Legal technology, clawlaw.in, 18 August 2026, ChatGPT, eighteen blind commercial questions.
• Named: 21 of 78 responses. Legal technology, clawlaw.in, 27 July 2026, ChatGPT, 78 blind questions, one run each. In the same reading ProVakil was on 28 of 78 and Legistify on 18 of 78, and the engine confirmed it had searched for 0 of the 78, so that ordering describes what the model had absorbed.
• Not cited: 161 recorded statements, across 24 batches and 72 conversations, September 2026, Perplexity, on our own domain, in the engines' own self audits of their own answers.
• Third party review sources reaching an answer: 0 of 6 questions. Legal technology, clawlaw.in, 17 September 2026, Claude. One well known review site appeared in the raw results and was discarded, because the list it offered was of American products and so was not an answer to an India question.

What those numbers actually suggest about who benefits. Not which industries, but which conditions. Three of them show up directly in the observations above rather than in anybody's theory.

• Buyers in your sector ask questions that have a supplier as the answer. Where the honest answer is an official portal or a free method, a commercial page should not and often will not win. That was Claude's own position on 17 September 2026.
• The evidence a buyer needs is publishable. On 17 September 2026 two enterprise vendors were ranked above clawlaw.in specifically because they publish explicit court and tribunal coverage lists, and the assistant noted that its ordering reflected price transparency and source authority rather than product quality. A business whose scope and price can be stated in public has an advantage over one whose cannot, regardless of which is better.
• Independent third party coverage in your sector is thin or absent. Where review sites and directories do not reach the answer, your own pages and official sources are what is left, which is an opportunity. In that same run, 0 of 6 questions had any third party review or directory source reach an answer.

What has not been run, as of 29 September 2026. We have not executed the 100 question, five category test on any domain or in any sector. So this page publishes no recommended and cited count over 1,200 responses, no rate by category, no rate by business size, and no comparison between sectors. Those cells are empty on purpose, and they stay empty until the run exists.

What this cannot tell you

• It cannot rank industries. Two counts from two sectors do not support a league table of sectors, and we are not going to build one out of them.
• A count from our domain is about our domain. 0 of 6 in AI visibility tooling says nothing about your category.
• Recommended and cited is a strict state and deliberately so. A business can be helped by being named without a link. The strict count is the one that is comparable, and the looser ones are reported beside it rather than instead of it.
• The used without attribution state is only partly visible. Both cases we hold were admitted by the engine when it was asked, on 17 September 2026, and would otherwise have gone unrecorded.
• Answers are unstable, which is why runs are the denominator and why three runs is the minimum.
• Nothing here promises a position in an AI answer. A test tells you which questions you are absent from. It does not buy you a place in the answer, and neither do we.

Sources and change log

Figures come from the clawlaw.in programme of July to September 2026, recorded in GEO_BASELINE_RESULTS_2026-07-27.md, GEO_GAP_ANALYSIS_2026-08-18.md and the assistant audit files of 17 September 2026, and from the aiknowsus.com audit of September 2026 across 24 batches and 72 conversations. Counts were produced by a script over those files. The capture files are not yet public, so the counts over them cannot be checked from outside today, and we would rather say that than imply they can.

Disclosure. We sell an AI visibility product, AI Knows Us, and every number above about our own domain is a number about our own product's market. We have published our 0 of 6 because a vendor who will not publish its own zero has no standing to ask you to publish yours.

Version. 29 September 2026, first publication. Updated when the 1,200 response run completes and whenever a figure is corrected.

Common questions

Is GEO worth it for a small business?

Run twenty questions before you decide, and look at two things in the answers. Whether the engine names any supplier at all, and whether the suppliers it names publish scope and price in public. If it names nobody, your sector is not ready and your money is better spent elsewhere this quarter. If it names suppliers who publish their scope, you are looking at a gap you can close with writing.

Which sector benefits most?

We do not know and we hold no measurement that would tell us. What we hold is two fractions in two sectors, 0 of 6 in AI visibility tooling and 1 of 18 in legal technology, and both are small. Anybody publishing a ranked list of sectors without denominators is guessing.

Our engine answers never name any company. What does that mean?

It usually means the question has a method as its answer rather than a supplier. That is a real finding and it should change your plan: write the method page rather than the supplier page, and accept that on those questions the official source outranking you is correct. Claude said exactly that about court portals on 17 September 2026.

Can we benefit if we are cited but never recommended?

Sometimes, and you should measure it separately so you can see it. Cited without being recommended means your content is being trusted while your commercial case is not being made. That is a positioning problem on your own pages, not a content problem, and it is invisible if you track one blended visibility score.

How many questions is enough for a first read?

Twenty, in the words your buyers use, with your brand in none of them, run three times each on two engines. That is 120 responses, which is one focused afternoon, and it is enough to tell you whether the 1,200 response version is worth planning.

Why publish your own zero?

Because the alternative is a page that asks you to trust numbers we will not apply to ourselves. Our count on our own category questions was 0 of 6 in September 2026 on Perplexity, and across 24 batches and 72 conversations the engines recorded 161 times that they had not cited us. Zero is where this work starts for almost everybody.

What to do first

Write twenty questions in the five categories, four per category, with your brand in none of them. Run them on two engines three times each and score only one thing to begin with: in how many responses was any supplier both recommended and supported by a link. That single fraction tells you whether GEO is a live opportunity in your sector before you spend anything on it, and it is a number you produced yourself.