Can you do this yourself?

The measurement is genuinely a do-it-yourself job. Write down thirty questions your buyers ask, put them to the six systems in a logged-out session, and record which companies get named. That costs an afternoon and no money, and it tells you whether you have a problem worth spending on.

What is hard is everything after it, and the difficulty is not knowledge. It is consistency. The same questions have to be asked the same way every month or the comparison means nothing. Pages have to follow a map, not a content calendar. Corroboration has to be earned outside your own website. And all of it has to survive four to six months before the numbers move, carried by someone with a day job.

Most in-house attempts manage the first month. The real question is not whether you can, but whether it is still happening in March.

Start with the part that costs nothing

Do this before you talk to anyone, us included. It is the only step that is both cheap and decisive.

Write thirty questions your buyers ask before they know who to call. Not keywords — questions, in their words. The best source is not a keyword tool; it is the first ten minutes of your last twenty sales calls, your quote requests, and whatever your support inbox gets asked repeatedly. "Which shrink wrapper suits a low-volume bakery line?" is a buyer question. "Shrink wrapper manufacturers" is a keyword.

Put each question to ChatGPT, Google AI Overviews, Google AI Mode, Gemini, Perplexity and Claude. Record three things for every answer: which companies were named, in what order, and what sources the answer cited.

Then count how many of those answers include you. That number is your starting point, and it is more than any dashboard you own will ever tell you.

Why the session has to be logged out

Because signed in, you are partly testing your own history rather than what a stranger sees.

You have visited your own website hundreds of times. If your account history informs what comes back, you will read a more flattering answer than your buyer gets, conclude the problem is smaller than it is, and act on it. Use a private window, signed out, and where a system offers a location setting, set it to the market you actually sell into rather than the one you are sitting in.

Test what your buyer sees, not what you have taught the machine to show you.

Where in-house attempts usually stop

Four places, in the order they tend to arrive.

The wording drifts. Somebody rephrases a question in month three because the new version reads better. The comparison is now between two different tests, and the movement you are looking at is your own editing. A fixed set of questions, re-run unchanged, is the whole basis of the measurement — we treat it as a named thing for exactly that reason, and it is defined on Track.

Pages get written to a calendar. Three posts a month feels like progress and is the most common way this work fails, because two of them end up competing for the same question. One question, one page, decided before anything is written. Without that rule you are building a pile rather than a body of knowledge.

Corroboration needs asking, not publishing. Correcting a five-year-old directory entry, getting a description fixed on an association listing, producing something a trade publication actually wants — none of that is writing, and none of it fits in the gap between other jobs. Correcting an out-of-date entry is the most tractable version of it, and that is set out in when AI describes your company wrongly.

It meets a quarter end. The first three months usually show nothing. That is normal and we say so before anyone engages us, but it is a hard thing to defend internally when the person doing the work is also the person who should be shipping something visible.

What is genuinely harder without help

Three things, and none of them is secret knowledge.

The first is scale. Measuring yourself is straightforward; measuring yourself against five named competitors across forty questions and six systems, every month, is a different quantity of work.

The second is the technical layer — consistent structured data, one identifier for your company, descriptions that agree with each other everywhere they appear. It is learnable, and it is the kind of thing that is either right or quietly wrong for a year. We have written about the part of it that is most often oversold in does schema markup help you get cited by AI.

The third is being wrong about your own copy. Everyone believes their product pages explain what they do. Buyers and machines frequently disagree, and the person who wrote the page is the worst-placed person in the building to notice.

When doing it yourself is the right answer

When you sell one category, in one market, in one language, and you have somebody technical who can write and can protect a day a month for six months. That is a real situation and it is not rare. In it, the in-house version works, and paying an outside firm buys you speed and objectivity rather than capability you lack.

It is also the right answer when you simply want to know whether the problem exists. Nobody should buy a diagnosis of a condition they have not confirmed they have.

When it usually is not

When you sell into several countries or languages, because each one is a separate set of questions rather than a translation of the first. When nobody owns it, and it becomes the thing everyone agrees is important and nobody has time for. And when the figure has to be defensible to a board, because a number produced by the person whose work it measures is a number somebody will eventually question.

What this article cannot tell you

It cannot tell you whether your people have the time, and that is usually the deciding factor rather than any of the reasoning above.

It cannot promise the in-house route produces the same outcome as the outside one. It might. It depends almost entirely on whether it is still being done in month five.

And whichever route you take, AI systems are probabilistic and change continuously. A single reading is a snapshot. What tells you anything is the same questions, asked the same way, over time — which is the discipline this article is really about, and the reason it is harder than it sounds.

What to do this week

Run the afternoon test. Thirty questions, six systems, logged out, written down. Do it whether or not you ever speak to anyone about this, because the alternative is deciding in the dark.

If what comes back is uncomfortable and you would rather have it done properly — agreed questions, named competitors measured alongside you, every source recorded, the whole thing in writing and yours to keep — that is what the AI Visibility Audit is for. If what comes back is fine, you have saved yourself a purchase, and we would rather you knew that.