How to measure your AI search visibility

What does a working measurement setup look like?

Here is ours, as it ran in the 30 days to 4 October 2026. Attensira tracks its own site, attensira.com, with the same product customers use, and we read it over the Attensira MCP.

The questions. 61 buyer questions: category questions such as "best AI search visibility tools", comparison questions such as "Peec AI alternatives", how-to questions such as "how to get cited by ChatGPT", and three branded questions kept in their own group.

The engines. ChatGPT and Perplexity, read from their consumer products the way a buyer sees them, three samples per question per day. In the month that came to 2,274 ChatGPT answers and 2,100 Perplexity answers.

The three things recorded per answer. Whether Attensira was named. Whether a page on attensira.com was linked. Which other brands were named, in what order.

The views we actually read. Three of them, and in this order.

That is the whole system: a fixed list of questions, read often enough to average, with three facts recorded per answer and three views on top. Everything else in this chapter is how to build it and how to read it without fooling yourself.

How do you choose what to measure?

Start from questions, not keywords, and write them down before you look at any results.

• 20 to 50 questions is a workable range for one brand. More than you can read deeply is worse than fewer read well.
• Phrase them as buyers would, with the situation that makes them ask: "best invoicing tool for a 20-person agency", not "invoicing software".
• Leave your brand name out, except in a separate branded group. Unbranded questions test whether engines reach for you; branded ones test whether they describe you correctly.
• Cover the journey: problem questions, shortlist questions, comparisons, and how-to questions about using a product like yours.

The GEO measurement guide's prompt-set chapter covers this in depth, including our own set's mistakes.

How do you measure by hand?

You can get a first, honest picture in an afternoon a week, with no tool.

1. Take ten of your questions.
2. Ask each on each engine you care about, in a fresh session with no history, logged out where possible.
3. Repeat on five different days over two weeks. Engines vary from answer to answer; one ask proves nothing.
4. For each answer, record in a spreadsheet: date, engine, question, whether you were named, whether your site was linked, and every other brand named.
5. At the end, count. "Named in 3 of 25 answers on ChatGPT" is a finding. "Named sometimes" is not.

The limits are real. Fifty answers per engine catch big differences, not small ones; a manual read cannot run every day; and a logged-in session can personalise answers. But it will show you who wins your questions, which is usually the most useful single fact you will learn.

The free audit does a version of this for you once: it simulates the four engines Starter reads, puts eight questions about your category to them, and shows who they name instead of you, which pages they lean on, and what the crew would do first.

What should you track continuously?

When you need to see change, you need daily readings and enough samples to average. A measurement tool should record, for every answer:

• the engine and whether it was read from the consumer product or an API;
• the country it was asked for;
• the full answer text, stored, so any number can be checked;
• every brand named and the order they appear in;
• every link attached, with the sentence of the answer it supported.

From that, the numbers follow: share of voice (answers naming you over all answers), citation rate (answers linking you over all answers), competitors' rates on the same denominator, and which domains supply the citations. Attensira shows each rate with its answer count, switches the headline to the count, such as 3 of 12, below 20 answers, and shows an engine it did not read as a dash, never as zero.

How do you read the numbers without fooling yourself?

Six rules. Most bad decisions in this field break one of them.

1. Read the count before the rate. 1 of 3 and 300 of 900 are the same percentage and very different evidence.
2. Report branded and unbranded apart. In our own month, every mention we had came from our three branded questions. Pooled, that looked like modest visibility; split, it is absence from our category.
3. Treat a dash as a gap, not a zero. An engine you do not read has no rate.
4. Judge per engine first. ChatGPT and Perplexity named different leaders on many of our questions. An average across them is a headline, not a plan.
5. Judge changes over weeks, past a test. A rate moving between two weeks is mostly noise unless it clears a significance test; Attensira uses a two-proportion test at 95% and otherwise says "nothing we can prove yet".
6. Log every change to the question set. Adding questions changes the average with nothing in the world changing.

Who should own the measurement?

One person, with a fixed routine, and not the person whose work is being measured. That is not a judgement on anyone's honesty; it is how the prompt set stays fixed and how a flat month gets reported as flat.

The routine is short. Each week, read the leaders on your ten most important questions and anything flagged as a real change. Each month, write the one-page report below. After each change ships, record the questions it should move and read the result at one, two and four weeks.

If the person doing the work also chooses the questions and writes the report, every month will look like progress. Separate the roles, even in a team of two.

What do Google's own tools show?

Less than you would like. Google counts appearances and clicks from AI Overviews and AI Mode inside the overall Web search traffic in Search Console, not as a separate report. So Search Console shows the combined effect on your clicks and impressions, and cannot isolate AI features. Google also says AI Overviews and AI Mode may use different models, so the links they show will vary.

Use Search Console for what it does well: which pages earn clicks from Google, and for which queries. Read AI answers directly for whether you are named.

Why does one page show up as two rows?

Because engines cite the address they found, and they find more than one. Our own Pages view shows it for every page we have: the home page, below, and our pricing page, which engines attached far more often on www.attensira.com than on the bare domain, both counted. Count only one row and you undercount the page; rank pages row by row and the order can change.

Two lessons. First, add the www and bare-domain rows for each page together before you rank pages or compare months, and redirect one host to the other so new citations settle on one address. Second, the Pages view is the most direct evidence you have of which of your pages engines trust. Read it before you rewrite or remove anything: the page you think is minor may be the one carrying your citations.

How do you know whether a change worked?

Compare the same questions before and after the change went live, on each engine, with enough answers on both sides.

• Attach questions to the change before it ships. Which answers should this page, fix or placement move? Write them down.
• Confirm it is live. A merged change is not a deployed page. Check the URL.
• Wait, then compare. Attensira re-reads the attached questions at 7, 14 and 30 days after a change goes live, comparing a week before with a week around each date, and states the answer count with every verdict.
• Accept "nothing we can prove yet". It is the most common honest answer on thin samples, and it is better than a story.

A movement past the test still does not prove the change caused it; other things change in the same weeks. What it gives you is a dated, checkable record, which is more than most AI visibility reporting offers.

What should you report each month?

One page, five lines:

1. The question set, engines and days read, with any changes to the set dated.
2. Share of voice on unbranded questions, per engine, with answer counts.
3. The competitors named most on your top questions.
4. Your most-cited pages, and any that started or stopped being cited.
5. What shipped, and what the re-reads say about it, including "nothing we can prove yet".

The defensible report chapter of the GEO measurement guide builds a full example from our data. The next chapter here, on what does not work, covers the tactics that measurement most often exposes.