Your buyers ask an AI first. Does it say your name?

What AI visibility measurement is

AI visibility measurement is counting how often AI assistants name your company when buyers ask them for a recommendation. You ask a fixed set of buyer questions, repeatedly, from clean logged-out sessions, and count the answers that name you. The result is a rate with a sample size, not a rank — and it moves when the pages assistants cite change.

It is sometimes called answer engine optimisation, generative engine optimisation, AEO, GEO or LLM visibility. The names differ; the measurement problem is the same one. What separates a measurement from a guess is whether the question set was frozen, whether each question was asked more than once, and whether the counting rule was written down before anyone looked at the answers.

When you check on yourself, you will probably get a better number than we gave you

This is the most important thing on the page, so it goes before the price. In September 2026 we measured one Singapore point-of-sale vendor several ways. The five questions never changed a character. The answer to “how visible are we” changed a great deal.

Adding six words to the question moved the result more than anything else we tried — same assistant, same five questions, each in its own fresh session. The only difference was an instruction wrapped around the question, the kind anyone types when they want a straight answer.

One variable. Nothing about the assistant, the session, or the wording of the question itself changed — only the instruction around it. So a founder who checks on themselves and gets a flattering answer is not being lied to by the assistant, and is not catching us out either. They asked a different question than we did.

This is why the rules below exist, and why they are published before the number rather than after it.

One more thing from the same pilot, with its sample size attached because a number without one is decoration: across 1,053 citations in 80 recorded answers, 48% went to just ten domains, and eight of those ten belong to companies selling in that category. The assistants were not reading neutral reviews. They were reading whoever had published.

What this cannot tell you

Every number has an edge. You should know where ours is before you pay for it.

• We measure a stranger, not your existing customer. We ask logged out, with no history — which is exactly the person who has not bought from you yet. What an assistant says to someone with months of chat history is different, and nobody can measure that from outside.
• We count naming, not revenue. Being named more often is not proof of more sales. Naming is the part that can actually be measured, and it is the part that happens before anyone reaches your site; the sale is still yours to close.
• Gemini runs without web search. Its terms don’t permit the automated grounded route, so that row measures what the model already knows. Every report says so on the same line as the number.
• Assistants change underneath us. When a provider swaps a model mid-month, a jump may be theirs rather than yours. We flag it instead of taking credit for it.

One measurement, repeated until it means something

A single answer is not evidence. Ask the same question twice and you often get two different lists. So nothing counts here until it has been asked three separate times, on three separate days.

40 questions × 3 repeats × 4 counted assistants = 480 answers in your number. Gemini adds 120 more, reported beside it rather than inside it, for the reason in the Limits above. Every answer is kept and readable.

Published, in order, before anything is counted

These run in this sequence because each step is only trustworthy if the one before it was fixed first. Changing step 1 after step 4 would quietly invalidate every month that came before it.

1. Freeze the question set.Forty questions, written before the first run and never edited mid-measurement. A question added later cannot be compared against last month.
2. Freeze the configuration.Model, search depth and country are settled by test runs, then locked. We record the model behind every answer, so a silent provider upgrade shows up as a flag rather than as movement.
3. Ask cold, three times.Fresh session, no history, no account, no follow-up question. Three separate days, so you see variation instead of one lucky answer.
4. Count by rule, not by eye.What counts as a mention is decided in code, written down in advance, and applied identically to you and to every competitor.
5. Report counts, not quotes.We tell you how many answers named you. We do not republish what an assistant said about a named competitor — those statements are not ours to repeat.

The full method is published here — sample size, decision rules, exclusions, and enough detail to take the measurement yourself without us. As far as we can find, no other tool in this category publishes theirs, which is why we also compare them openly.

It also names our dependencies. One of the four counted rows reaches its assistant through a third-party search-data provider, and that provider is a single point of failure we do not control. If that route closes, the row pauses, the report says so on the page where the number would have been, and the other three continue. We are not going to tell you a spare is ready when it is not.

What you are actually paying a person to do

The measurement is the cheap half and it runs itself. The month is the other half, and it is why this is priced like a person rather than like software.

That last step is the part most of this industry avoids. An agency that does the work usually also decides how the work is judged. Here the work is graded by a published method that was fixed before the work started, and the report prints the months where nothing moved.

What we do not promise: that you will be named more often. Nobody controls what an assistant says, and anyone who tells you otherwise is selling something. What is promised is the work, and a number honest enough to tell you whether it worked.

What lands in your inbox

Seven sections, the same seven every month, so this month can be laid beside last month without anyone re-explaining the format.

1. Summary.Named in X of 480 answers, the leader’s figure beside it, and the winnability split below.
2. By assistant.One grid per assistant, one square per question, shaded by how many of the three runs named you. Plus the blunt line: the questions where no assistant named you at all.
3. By kind of question.Category, use case, comparison, alternatives, problem — eight questions each, and the five rates that add back to the headline.
4. Pages the assistants cited most.Which pages decide your category, how often each was cited, and whether it mentions you at all.
5. A 90-day fix list, in priority order.Every action carries the number from sections 1–4 that justifies it. No action appears without its evidence.
6. All forty questions.Every question, every assistant, runs out of three. Nothing aggregated away.
7. Method and limits.How each question was asked, what counts as named, and where the figure is an estimate rather than a fact.

This field is full of promises. Here are the ones we refuse to make.

Published, because hiding it is the oldest trick

One price covers both halves: the measurement, which runs itself, and the work to change what it finds, which is a person’s month. We don’t sell a dashboard and leave you to act on it.

For context, $1,500 a month is roughly a junior SEO retainer, or four hours of a consultant who will send you a slide deck. The difference is not the price. It is that at the end of the month there is a number, taken the same way it was taken before the work started, that says whether any of it mattered.

The free first report is five of your buyers’ questions across two assistants, asked three times each — thirty answers. It is a real measurement under the same rules, not a trailer, and it is not the $750 audit, which is forty questions across every assistant. Three rules keep it honest rather than endless: one per company, requested from an email on the domain being measured, and a limited number each month.

Ten is not a marketing device. One person does the fix work, and ten companies is what one person can do without the eleventh getting a worse version of it.

You will always talk to the same person

Satayu “Bon” Isariyaphorn — founder. AI systems do the measuring, the research and the first drafts. I review everything before it reaches you, I answer your email myself, and I am the one accountable for the work.

That is also why the roster stops at ten. It is not a growth strategy; it is the number of companies one person can do this properly for.

satayu@fortyquestions.io

Straight answers

Can you guarantee an assistant will recommend us?

No. Anyone who says yes is guessing. We show you the number, do the work known to influence it, and show you the number again under the same rules.

How long before the number moves?

We don’t know for your category until we’ve measured it. Assistants that search the live web can reflect a new page within weeks. Changes that depend on a model being retrained take longer. You get a report every month, so judge us after three.

Is the work done by AI?

Most of the measuring, research and drafting is, which is why this costs less than an agency. A person reviews every deliverable and is your single point of contact.

Which assistants do you measure?

In the number: ChatGPT, Claude, Perplexity and Google’s AI answers in search — four, collected automatically, each question in its own fresh session. Beside the number: Gemini, which runs without web search because its terms don’t permit the automated grounded route. Beside means reported next to your figure, never added into it. If your buyers use another one, tell us.

What if we test it ourselves and get a different answer?

You probably will, and the Evidence section above explains why — six extra words in the question changed a result from 3 of 5 to 5 of 5. Send us what you typed and we’ll tell you which rule it broke.