Measure your AI visibility: is it you in the answer, or your competitor?

Measuring whether you're visible to AI assistants can't be done with a classic ranking tool — it looks for positions in a results list that simply doesn't exist in a generated answer. What does work: your own, repeatable sample. Draw up a fixed set of questions a potential customer would realistically ask, put them to ChatGPT, Perplexity and Gemini, and note per question whether your business gets mentioned, with which source, and which competitor shows up instead of you.

That's not a one-off exercise. AI answers shift with every model update and every crawl, so last month's result says little about today. The value is in repeating it: the same questions, every month, in one simple overview — that's how you see a trend emerge instead of a single snapshot.

Why your old SEO tools don't see this

Rank trackers measure positions on a results page. An AI answer has no positions — it has a selection. The model picks a handful of sources, weaves them into running text, and (sometimes) states where the information came from. If you're not in that selection, you're nowhere — there's no "position 11" to grow from. That makes the line between being cited and not being cited much sharper than in classic SEO, and it makes standalone Google Search Console numbers unfit to argue AI visibility from.

On top of that, every AI assistant uses its own source mix. Perplexity cites almost always explicitly, ChatGPT does so inconsistently depending on the mode, and Gemini weaves in results from the Google index. Measuring visibility therefore means checking each assistant separately, not keeping one average.

Three ways to measure it yourself

1. Manual sampling. Gather 10 to 15 questions a buyer in your market would ask — not your brand name, but the problem ("reliable AI for legal documents in the Netherlands"). Ask them in a fresh, non-personalized session with each assistant and log the result in a spreadsheet: mentioned (yes/no), with source URL, and which competitor shows up instead. It costs an hour a month and gives the most honest picture, because you see exactly what's said, not just a score.

2. A technical scan. Separate from what a model actually answers, there's the underlying question: can an AI system technically read your site well in the first place? Our own Agentic Search Optimizer runs a site through 15+ checks — crawler access, schema.org, llms.txt, citable structure — and delivers a score with concrete fixes. That's not a replacement for the manual sample above (a technically perfect site can still get skipped), but it closes the easiest leak: a model that simply can't read you, also can't cite you.

3. Tracking mentions off your own site. AI assistants don't only cite your own pages — reviews, forum posts and trade publications count just as much as a source. A periodic search for your brand name on trade forums, LinkedIn and news sites shows whether external material is emerging that a model can cite when your own site doesn't cover the answer.

What counts as a good result?

There's no universal benchmark — "80% of questions" means nothing without the context of your market and competition. The useful comparison is relative: how often do you show up compared to your two or three direct competitors, on the same set of questions? If you appear in 2 of 12 questions and a competitor in 9, that's the concrete signal to act on — not an absolute number, but a ratio you can track month over month.

Watch for false negatives too: some assistants cite a source without naming the brand literally in the running text, but do list it as a link at the bottom. So check not only whether your name appears in the text, but also whether your URL shows up in the source list.

Setting up a simple measurement rhythm

Practically: fix the question set once, add a monthly reminder, and store the results in the same document so trends become visible instead of isolated measurements. After every substantive change to your site — a new knowledge base article, updated schema.org markup — add the next measurement as a separate data point, so you can see whether a change actually shows up in what a model cites. Without that repetition, every GEO effort stays guesswork.

Conclusion

You don't measure AI visibility with a ranking — you measure it with a repeated, honest sample: the same questions, every month, per assistant separately, compared against your competition. You check the technical foundation — can a model read your site at all — separately with the Agentic Search Optimizer. How that technical foundation and the editorial side together form a complete GEO picture is laid out in our pillar article on GEO.