Are you mentioned by ChatGPT and Perplexity:how to measure it

Why you must measure this and cannot assume it

For classic findability there is Search Console. For AI answers no such thing exists. There is no report that shows how often a language model named you, against whom you lost out or which page it used. You only close that blind spot by hand.

Two pitfalls make most home-made tests worthless. The first is your own account. If you ask an assistant that knows your conversations of the past months, then it names you and that says nothing. The second is rephrasing. Anyone who adjusts the question until the answer pleases measures their own patience.

Chapter 6 makes the distinction between readiness and visibility. This chapter is about the second. You can score ten out of ten on readiness and appear in no answer at all.

The fifteen questions in three groups

Fifteen is no magic number. It is the number you manage in about an hour a month across four engines and that is big enough to see movement. The division is six, six and three.

Group A: the category questions, without your brand name. Six of them. This is the group that counts. Here someone who does not know you asks precisely the question to which you are the answer. Patterns that work:

• Who can help me choose a [system] for my SMB in [region]?
• I am looking for an independent [role] for [type of project], not a supplier. Who?
• Which agency does [your speciality] in [region]?
• I want [concrete result] made. Who can advise me on that?
• Who helps SMBs with [the task that precedes your assignment]?
• I am looking for help with [sub-problem] in [town].

For an installer of industrial cooling in Kortrijk that becomes for example: "Who installs industrial cooling systems for food companies in West Flanders?" Write them as a customer types them. Spoken language and spelling mistakes may be in there if that is the reality.

Group B: knowledge questions to which your own content gives the answer. Six of them. Here you measure whether your blog and your service pages get picked up. Patterns:

• What does [your type of project] really cost over five years?
• What is the difference between [A] and [B]?
• When do you choose [option 1] instead of [option 2]?
• Which cost items are usually forgotten with [your subject]?
• How do I prevent [the risk that keeps your customers awake]?
• Is [the situation of the customer] suitable for [your solution]?

One rule with group B: every question has to match a page you have already published. If you ask something you have written nothing about, then you measure the market and not yourself.

Group C: brand questions. Three of them. What is [your name] and what do they do. Is [your name] a [your type of company]. What does [your best known service] cost at [your name]. This is the control group: does the model know you at all.

Expect no pride here. In our baseline measurement every engine that gave an answer scored 3 out of 3 in this group. That is the floor and no achievement: as soon as your site is indexed, answering about your own name works by itself. Group A is the real test.

The six measuring rules

These are binding. Anyone who lets go of one of them compares with noise next month.

1. Fresh session without personalisation. New incognito window, logged out, no chat history. A logged-in account that knows you gives a false positive.
2. Fixed language and location. The language of your market and the region in which you work. With Google leave the search region set as it is. If you change this, you measure a different market.
3. Word for word the same prompts, in the same order. Do not rephrase, not even if the answer disappoints. If you want to replace a question, do that in January and note it.
4. One round a month, around the same date. AI answers differ from day to day and sometimes between two goes at the same question. Day-to-day comparison measures chance.
5. Note three things per prompt. Are you mentioned, yes or no. Is there a clickable source link to your domain, yes or no. In which place in the answer: at the front, in the middle, at the back or not at all. Mentioned means that your name stands word for word in the text. A description that resembles you without your name does not count.
6. Keep every round unchanged. One file a month, always the same format. Never update an old round with what you found out later. The value sits in the series and not in the last measurement.

The engines

One methodical point with Google that often goes wrong. Note the difference between "no AI Overview shown" and "overview shown without you". Those are two very different zeros. In our baseline measurement not one of the fifteen Dutch-language searches in Belgium produced an AI Overview. Anyone who notes that as an ordinary zero sees a rise next year that says nothing about their website and everything about the rollout by Google.

The three key figures and how you calculate them

Citation Frequency. The number of prompts in which you are mentioned, divided by fifteen, per engine. Our baseline measurement of 2 August 2026 gave for ChatGPT with search 5 out of 15, so 0.33. For Perplexity 3 out of 15, so 0.20. Always calculate this per engine and never as one average across all engines together, because precisely the difference between the engines is the interesting part.

Share of Voice. Your share in the company you keep. In group A count all the parties named per answer and count how often you stand among them. Suppose the six answers together name 22 companies and that you stand there twice, then your share is 2 out of 22. That figure is strict and useful for exactly that reason: it moves also when your own score stays the same and a competitor disappears.

Brand Visibility. Prominence plus source link. Our own operationalisation: at the front of the answer counts 3, in the middle 2, at the back 1 and not mentioned 0. Take the average over the fifteen prompts. Note separately besides that which part of your mentions had a clickable source link, because a mention without a link yields authority and no visitor.

The full baseline measurement table with the four engines is in chapter 6. Use it as calibration and not as a norm.

The score sheet

One file a month, for example geo-citaties/2026-08.md. In it two tables and two sentences.

The first table is the score sheet. One line per prompt, one column per engine and two columns for the source link and the position.

The second table is the most important one and is missing in most examples: per question from group A you note all the companies that did get named, per engine. Two lines of observation underneath. What stood out, what changed against last month.

Count on an hour to an hour and a half per round with four engines. Always have it done by the same person, because the boundary between the middle and the back is a judgement and different people put that boundary elsewhere.

Why one measurement says nothing

A single round is a photograph. The same question can produce a different answer tomorrow without anything having changed on your site. Anyone who builds conclusions on one measurement builds on noise.

What you are looking for is a change of direction around a moment you made yourself: the publication of a series of articles, the indexing of new service pages, a mention in the trade press. Without a time series you cannot see whether something worked and so you also cannot stop what does not work.

Zero is a fine starting point. Group A stands at zero for almost everybody at the start, also at companies that rank excellently in the classic way.

For a new site, zero is the baseline

If you launch a new website, take the measurement straight away and not once there is something to celebrate. You will score zero. Every new site scores zero, because no system knows you yet. That is not a failure but your starting point, and without that point you cannot show later that anything worked.

So measure before you start writing and building, not after. It is the cheapest measurement you will ever take and the only one you cannot make up for later.

What you may reasonably expect after that, based on how these series behave:

If that does not happen, it is not working. Then you stop or change approach instead of pushing on because it has already cost so much time. That is the whole point of measuring.

What the first measurement really yields

Not your own score. The list of competitors.

In our baseline measurement the engines largely named different competitors for the same question. There is therefore no fixed list you have to get onto. And the round produced four parties we did not know and that position themselves almost identically to ourselves. That is market research you get thrown in for free. For most SMBs it is worth more in the first year than the score sheet itself.

Another find from that same round: with the question what the company is, two of the four engines built their answer on the LinkedIn company page and not on the homepage. Anyone who perfects their own website and lets their profiles lapse optimises the wrong source. That is chapter 10.

What you can do yourself and what you cannot

To do yourself, without tools:

• Draw up the fifteen questions. That is half a day of thinking and it is the only step you really have to do well, because you are stuck with it for years afterwards.
• Run the round in an incognito window. An hour a month.
• Fill in the score sheet and work out the three key figures. That can be done in a text file or a spreadsheet. No subscription is needed for it.
• Keep the list of named competitors up to date and really look at it twice a year.

What you need help with or what you had better leave:

• Statistical certainty. Fifteen questions per engine is no sample on which you can calculate significance. It is a thermometer and not a laboratory.
• Automating through an API. That looks attractive. Only the API is a different product from the chat interface: other settings and often no search function. You then measure something other than what your customer sees.
• Paid services that promise to place you in AI answers. There is no placement for sale. What is for sale is measuring work you can do yourself.

Would you rather not do this yourself: SEMANU carries out the full SEO and GEO audit for 1,450 euro, report and working session included. What it covers is on the guide page.