Your customers are already asking assistants the questions that decide who gets the work. The measurement only tells you something if you ask the same questions they do — and almost nobody does.

This is the practical companion to why AI answers differ every time. That article covers how to measure honestly: fixed questions, a schedule, a rate instead of a screenshot. This one covers the part that comes first — which questions to fix. It is where most measurements quietly go wrong, and the failure is entirely understandable: the person choosing the questions knows the business from the inside, so they ask from the inside. They type the company name, or the product name, or the industry term for what the company does. The assistant answers, the company appears, and the measurement reports a visibility that no customer will ever experience, because no customer asks that way.

Which questions should I ask AI assistants about my business?

Five to ten questions your customers would actually type into an assistant, written in their words rather than yours. That means a service with a place attached ("emergency plumber in Leeds"), a problem described before it has a name ("water heater making noise"), a comparison ("best accounting software for a small shop") and phrasings that carry buying intent. Your product names and slogans are the words customers use last, so they belong at the bottom of the list, if they appear at all.

Each of those shapes catches a different moment in the same journey. The service-plus-place question is someone who knows what they need and wants a name. The problem-first question is earlier: the customer does not yet know what the fix is called, only what is wrong, and whoever the assistant names at that moment gets to define the problem for them. The comparison question is someone weighing options, and the buying-intent phrasing — "cost of", "near me", "which should I get" — is someone close to a decision. A question set that only covers one of these moments only measures one of them.

The reason product names go last is not modesty. A customer who types your product name into an assistant has already found you; the answer to that question was never in doubt and never in play. The questions worth measuring are the ones asked by people who do not know you exist yet — which is exactly why they are the hardest questions for an insider to think of.

Where do I find the words customers actually use?

In records of what customers already say, not in a brainstorm about what you would like to be asked. Three sources cover most businesses: the queries in your Search Console report, which are the questions your site is already being shown for; the way customers describe their problem on the phone, before anyone has corrected their vocabulary; and the phrasing in your reviews, where customers explain in their own words what they came for.

Search Console is the strongest of the three because it is measured rather than remembered. Every row in the queries report is a real search by a real person that surfaced your site, spelled the way they spelled it. The rows that get impressions but few clicks are especially useful: those are questions the market is asking in volume, in words that are demonstrably not quite yours yet.

The phone and the reviews fill in what Search Console cannot: the phrasing of people who have not yet reached a search box. Whoever answers your phone hears the raw version of the problem several times a day — "it's making a noise", "the email says something about a certificate" — and that raw version is much closer to what gets typed into an assistant than anything a marketing meeting will produce. If a phrase makes you wince because it is technically wrong, that is usually a sign it belongs on the list. The measurement is about the question as asked, not the question as it ought to be asked.

Should I ask Google's AI features and chat assistants the same questions?

Yes — and record the results separately, because Google's AI features and the chat assistants are different channels with different retrieval and different answers. Being named in one says nothing about the other, in either direction. A question is not covered once it has been asked somewhere; it has a result per surface, and the surfaces can disagree for months at a time.

The disagreement is structural, not noise. Google's AI features draw on Google's index and Google's crawling; a chat assistant draws on its own retrieval, fetched by its own crawlers, which your site may treat completely differently — sometimes without anyone having decided to. So the useful record is a small grid: each question down the side, each surface along the top, a result in every cell. The grid is what tells you whether an absence is about the question (absent everywhere) or about the channel (absent in one place, named in another) — and those two findings call for entirely different fixes.

What should I record when an AI assistant answers?

Three outcomes per question: named, absent, or unknown — the runs where the answer could not be read at all. And whenever the answer is absent, write down who was recommended instead. An absence with a name attached is the single most useful thing this measurement produces, because it tells you which business is actually winning the question in the AI channel.

The unknown category earns its place quickly. AI features do not render for every request, and answers sometimes arrive in a form that cannot be read reliably. Neither of those is an absence — it is a reading the instrument failed to take — and a record that quietly files unknown under "not mentioned" will report you as less visible than you are. The honest habit is boring and worth keeping: if you could not read the answer, the outcome is unknown, and unknown licenses nothing except asking again next time.

The competitor names are the part almost everyone throws away, and they should not. In one measurement we ran, an assistant answering a question about monitoring recommended a competitor by name — and that single answer identified the actual rival in the AI channel better than any ranking report had. The businesses contesting your keywords in classic search are ones you already watch; the name that comes back from a generated answer is frequently a different company entirely, because being citable is a different property from ranking. An answer to a commercial question is rarely empty. When it does not name you, it names someone, and that someone is a fact worth a column of its own.

How many questions should I track in AI answers?

Fewer than you think. Every question you track costs attention each time it runs — and once tooling asks on your behalf, it costs money too — so five well-chosen questions asked on a fixed schedule beat fifty questions asked once. The value is in the repetition: re-ask the same questions, unchanged, and read the rate over time rather than any single answer.

The temptation to be thorough is the trap here. A list of fifty questions feels rigorous on the day it is written, and then the second run never happens, because reading fifty answers across several surfaces is an afternoon. What survives contact with a normal working week is a short list — short enough that the results actually get read, stable enough that the trend line means something. A single answer to any one question is one instance of a system that varies by design; "named in six of the last ten runs" is a measurement. If a question stops mattering to the business, retire it openly and note the date, rather than rewording it — a reworded question is a new question wearing the old one's history.

What to do with what you find

The findings only pay for themselves if each outcome has a next step, and the steps are mercifully short.

  • Absent on a question: first check that the pages which should answer it actually do — the customer's question as a heading, a plain answer in the first paragraph, in the customer's words. The longer treatment is in pages that answer real questions. Then check that the AI crawlers can reach those pages at all, because the best answer page on the web does nothing for a crawler that is turned away at the door. What makes a page the kind of source generated answers cite is covered in how to get cited by AI search.
  • Named: find the page that earned the citation and keep it healthy — reachable, fast, current. A page that wins a question in the AI channel is an asset with a measurable job.
  • Unknown: nothing, yet. Wait for more runs before concluding anything.

What to do this week

Open your Search Console queries report and pull out the questions your site is already shown for. Add the two or three phrasings you hear most often on the phone, in the caller's words. Cut the list to five to ten, spread across the four shapes — service and place, problem-first, comparison, buying intent — and ask them once on each surface you care about, recording named, absent or unknown per cell, with the recommended names written down wherever you were absent. Then put the same questions on a schedule and leave the wording alone, because everything useful lives in the rate.

TrustCtrl runs this as part of its site checks: your chosen questions asked on a schedule across Google's AI features and the chat assistants, each run recorded honestly as named, absent or unknown, with the sites that were cited or recommended listed alongside. It sits next to the checks that decide whether you can appear at all — whether the AI crawlers can reach you, whether your pages answer plainly — so the report pairs where you stand with what is standing in the way.