What AI assistants tell buyers about your industry, and how to measure the category instead of your brand

Almost every tool in this field is sold as brand monitoring. You type in your company, it reports how often your name comes up. That measurement is easy to produce and it answers the wrong question, because a buyer asking an assistant for help does not type your name. They describe a problem, and whoever the answer names wins.

Measuring that properly is a category benchmark, not a brand check. The method is not difficult. Getting it unbiased is, and this page is mostly about the four ways a panel leaks.

Why industry measurement is a different measurement

A mention is not a recommendation

Three different events get reported as visibility. Your name appearing anywhere in the text. A page on your domain appearing among the cited sources. Your company being put forward as a choice. They have different denominators and they move independently. In our own last run, our domain appeared in 0 of 19,793 citation URLs and our name appeared in 0 of 635 readings, while rivals were named in 451 of those readings without us. Had any one of those three figures been reported alone, it would have told a different story.

A branded question produces the brand

This is the single most common error, and it is not subtle. Ask an assistant what it thinks of your company and it will discuss your company. Any visibility figure produced from a question containing your own name is measuring nothing except that you typed your name. Our own audit ran 18 batches of questions in which our company is never mentioned, and only the last batch names it, run last so it cannot contaminate the rest.

Competitor selection changes the answer

Pick the three rivals you worry about and the comparison questions you write will be about them. The panel has to be built from how buyers describe the problem, before anyone looks at who wins, or the result is a measurement of your own assumptions.

How to build a buyer-question panel that is not rigged

1. Write the questions before you look at any answer. Lock the list, date it, and keep it. A panel edited after the first results is no longer a baseline.
2. Cover three intents separately. Discovery, where the buyer is naming a problem. Comparison, which almost always produces names and often three. Purchase intent with a real constraint in it, a budget, a team size, a city, a compliance requirement.
3. Stratify. By industry, by location where location changes the answer, and by buyer situation. Twenty questions spread across three intents beats sixty variations of one.
4. Prevent leakage and duplication. No brand names anywhere in the unbranded buckets, including your rivals, and deduplicate near-identical phrasings before you count, because four phrasings of one question inflate every rate they touch.
5. Fix the market in the words. On the route we used, the country, language and location parameters were accepted by the interface and then ignored, which we measured on our own traffic. A US benchmark has to establish the market in the question and confirm it in the answer.

How to score the answers

Five metrics, each a fraction with a visible denominator.

• Mention rate. Readings in which a brand appears, over eligible readings. Eligible matters, because a question that names nobody cannot be a loss for anyone. In our run, 184 of 635 readings named no company at all, so scoring against 635 would have understated every brand in the category including the leaders.
• Recommendation rate. Readings in which the brand is actually put forward as a choice, which is a judgement call and needs a written rule applied by a human before counting.
• First-choice share. Only scored where the answer gives a meaningful order. Many answers do not, and inventing a position for them is the quickest way to a false benchmark.
• Source-domain diversity. How many distinct hosts the answers lean on. In our run, 19,793 citations resolved to 3,307 hosts, and half of all citations came from only 51 of them. A category whose answers rest on a few dozen sources is a category you can work on. One resting on thousands is not.
• Repeatability. The share of repeated runs in which the same brands appear. Without it, every other number above is a single draw from a distribution.

Count one brand once per reading, with repeats inside a single answer collapsed. The counting rule changes the totals more than the data does, and two benchmarks under different rules cannot be compared.

A worked example from our own category

We ran this on the category we sell into. On 22 October 2026, 635 readings across ChatGPT and Gemini, 324 US questions in 19 batches, 550 readings with live search. The result was 1,053 namings across a roster of 24 vendors, with the top of the table held by brands with very large published estates and by three specialists whose own domains also sit high in the citation count. The full roster, with readings per name, is in which brands AI assistants put forward when US buyers ask. Our own name appeared 0 times, which is the most useful line in it for judging whether to believe the rest.

Two things that run taught us about panel design. Ten repeated phrasings of one buying question produced one finding, not ten, so the deduplication rule was doing more work than the sample size. And the questions that produced no company at all clustered by intent rather than by industry, which means the fix for those is a different question, not a different page.

What a category benchmark cannot tell you

It cannot tell you why. A roster ranks outcomes, and the causes sit upstream in whether a question calls for a name at all, whether you can be retrieved, whether more than one independent source says the same thing about you, and whether your claims are specific enough to repeat. It also cannot establish that any intervention caused a change, because that needs a comparison group measured over the same period, with model version changes logged separately. We publish our own failure on exactly that point. Going from 134 to 641 recommendation-shaped pages between 20 and 22 October 2026 moved our naming count from 0 to 0.

Who should run one, and who should not bother

Run a category benchmark if your buyers ask open questions and your category has a settled vocabulary. Do not bother if the category questions in your field come back as guidance with no companies in them, because the measurement will be accurate and will not tell you anything you can act on. Start with a one-off hand-run panel of twenty questions before buying anything. If the answers name nobody, the problem is not your visibility.

Where we stand to gain, and where a rival is better

We sell this measurement, so treat the method here as the argument of an interested party and the numbers as checkable. We read five surfaces, ChatGPT, Gemini and Google AI Overviews, Copilot, Perplexity and Claude among them, which is five product families and not nine. Our pricing page, 119,203 bytes, renders no SSO, no SAML, no SOC 2, no ISO, no DPA and no white label, checked by word boundary over the whole file. Read on 22 October 2026, tryprofound.com/pricing publishes up to 9 answer engines on its Enterprise plan together with SSO, SAML and SOC 2 compliance, and publishes no numeric price for it. On breadth of engines and on security paperwork they are ahead of us.

Our own published US prices are SEO Start 149, Growth 299, Scale 599, SEO Scale 699, Inbound 1099, Outbound 1299, Agency 1299 and Everything 1799 dollars a month. SEO Start reads 0 questions, 0 credits and 0 engines, so nobody should read AI visibility as starting at 149. Growth at 299 is the cheapest plan that reads all five surfaces. Our pricing page also says the Agency plan covers 10 client brands while its own comparison table answers No to more than one business in all eight columns, and both readings are live right now.

Questions about industry-level AI measurement

How many questions does a category panel need

Twenty well-spread questions run three times each beats two hundred run once, because the repeat is what separates a change from ordinary variation.

Should I include my own brand name anywhere

In one clearly separated bucket, run last. Mixing branded and unbranded questions in one rate is the error that invalidates most vendor visibility figures.

Why do the same questions give different brands on different days

Generation is probabilistic, the decision to search is taken per question, and a searched answer reflects whatever was indexed that week. That is why repeatability is one of the five metrics rather than an excuse.

Is a category benchmark worth publishing publicly

It is the most citable thing a company in a measurable category can publish, provided the prompt set, the counting rule and the denominators go out with it.