How to measure whether ChatGPT and Perplexity mention you

A marketer recently asked us why new enquiries were coming in with 'via ChatGPT' filled in as the source, while the analytics package showed zero referrals. That gap is the problem. An AI assistant delivers an answer, not a click. And anything without a click never shows up in your reporting.

GEO stands for generative engine optimization: making sure your brand appears in answers from ChatGPT, Perplexity, Gemini and Copilot. Plenty has been written about the optimising part. This article covers the step before it: measuring where you stand today. Without a baseline, you will not know later whether your work changed anything.

Why your current dashboard misses this

AI answers behave differently from search results. That has practical consequences for measurement.

• No click, no referrer. Someone reads your name, remembers it and types it in directly later. In your data that shows up as direct traffic.
• Answers are non-deterministic. The same question, two minutes later, can produce a different list. A single measurement tells you nothing.
• Personalisation interferes. Chat history, location and account settings all shape the answer you see.
• There is no position 1 through 10. You are mentioned or you are not. At best you can note whether you appear early or late in the list.
• The platforms differ. Perplexity shows sources prominently and therefore does send clicks. ChatGPT links sometimes, and often not to the party it names.

What you can actually measure

You cannot measure a ranking. You can record five things that together give a usable picture.

1. 1Mention rate. In what percentage of a fixed set of questions are you named? This is your headline number.
2. 2Position within the answer. Are you the first suggestion, or the sixth name in a list of eight?
3. 3Context and sentiment. Are you described as a specialist, as the cheap option, or as a footnote to a competitor?
4. 4Cited sources. Which URLs does the model use to support the answer? Often these are not your own pages, but comparison sites, trade media or forums.
5. 5Self-reported origin. Add one field to your form: 'How did you find us?' It is not clean data, but it is the only place where the AI assistant gives itself away.

Start with a prompt set

The quality of your measurement depends on your question list, not your tool. Build a set of thirty to fifty questions your audience would genuinely ask, then keep that set fixed. Change the questions every month and you are measuring noise.

Spread the questions across four types:

• Category questions: 'best [service] for [type of company] in the Netherlands'
• Problem questions: 'how do I solve [problem] with a small marketing team'
• Comparison questions: '[your brand] versus [competitor]' and 'alternatives to [competitor]'
• Brand questions: 'what does [your brand] do' and 'is [your brand] suitable for B2B'

That last category is usually skipped. It is the most revealing one. If a model describes your brand incorrectly, you know immediately where your content falls short.

—You do not decide who an AI model mentions. You do decide what there is to find about you.

How to do this in one hour a month

For a small team, manual measurement is perfectly workable. You need a spreadsheet and some discipline.

1. 1Create one tab per platform: ChatGPT, Perplexity, Gemini. Start with two if time is tight.
2. 2Use a temporary chat or a logged-out session. Otherwise you are measuring your own history.
3. 3Ask each question three times and log: mentioned yes/no, position in the list, competitors named.
4. 4Copy the cited sources into a separate column. Your action list lives there.
5. 5Do this on a fixed day each month. Trends are useful, one-off readings are not.
6. 6Report one number upward: the percentage of questions in which you appear, per platform.

Expect 45 to 75 minutes for thirty questions across two platforms. That is manageable. What does not work: doing it once, making a chart and never looking at it again.

Tools, log files and what they prove

A range of trackers now automates this, and the larger SEO suites have modules for it. They save time once you pass fifty questions or work in several languages. Below that threshold the added value is limited. Stay critical of their claims too: no tool has access to what users actually ask. They run the same prompts you do, just more often.

Harder data comes from your server logs. There you can see which AI crawlers fetch your site, and which pages. Look for OpenAI's GPTBot and OAI-SearchBot, PerplexityBot, ClaudeBot and Google-Extended. Check your robots.txt while you are at it: we regularly find sites blocking these bots without anyone having decided to. In that case, there is nothing left to measure.

From measurement to decision

The numbers are not the goal. They help you choose. Named in 10 percent of category questions but 60 percent of brand questions? Your content is fine, the market simply does not know you yet. That is a brand issue, not a content issue. Seeing the same three external sources in the citations every time? Then you know where you need presence, regardless of what you publish on your own site.

What we honestly do not know: how much revenue this produces in the short term. Volumes are still small compared with classic search, and the platforms change behaviour without notice. So we treat AI visibility as brand work with a measurable side, not as a replacement for your performance channel.

Start small. Thirty questions, two platforms, three months in a row. After that you have a trend line instead of a snapshot, plus a list of sources you can act on. At REX.digital we build this baseline into the foundation phase, because without a starting point you cannot demonstrate improvement later. Put the first measurement in your calendar this week; it costs you one hour.