Ask a marketing team what Google says about them and you get a rank tracker, a Search Console export and a fairly confident answer. Ask what ChatGPT says about them and the room goes quiet.
That gap is the whole problem. A growing share of buying decisions now start with a generated answer, and almost nobody is measuring their position in one. Worse, the absence is invisible: there is no impression count for the shortlist you were left off.
Here is a method you can run yourself this week. It is not sophisticated. It is just written down, which is the part most teams skip.
Step one: write the prompts, not the keywords
Start with twenty to forty questions a real buyer would type. Not keywords — questions, in the words a person actually uses, including the messy ones.
Cover four intent bands. Category questions: who is the best supplier of X. Comparison questions: is X or Y better for a company like mine. Problem questions: how do I fix Z. Local questions: who does X near me, if geography matters to you.
The temptation is to write prompts you would win. Resist it. A prompt set built to flatter you produces a baseline that improves on paper and never in the pipeline. Write the questions your hardest prospect asks.
Step two: run them across four assistants
ChatGPT, Google Gemini, Perplexity and Microsoft Copilot cover the majority of assistant usage today. Add Claude if your buyers skew technical.
Run each prompt in a fresh session with no personalisation and no memory of your business. Logged in, with your own history, you will see yourself far more often than a stranger would — which is exactly the trap that makes internal AI visibility checks so misleading.
- Use a clean browser profile or incognito window
- Turn off any memory or personalisation features
- Run each prompt at least twice — responses vary between sessions
- Save the full response text, not a summary or a screenshot
Step three: record four things per prompt
For every prompt and every assistant, note: whether you were named at all, what position you appeared in, what reason the model gave, and which competitors appeared alongside you.
That last column is the one teams undervalue. The competitor set an assistant returns is a live map of who the model considers comparable to you — which is frequently not who your sales team thinks you compete with.
The reason column matters too. When a model names a competitor because they publish pricing and you do not, it has told you precisely what to fix.
Step four: accept that the number will move on its own
Assistant responses are not deterministic. Run the same prompt twice and you may get different names. That variance is real and it does not mean your measurement is broken.
Handle it the way any noisy metric is handled: multiple runs, report the pattern rather than a single result, and treat small movements as noise. If you appeared in one of six runs last month and three of six this month, that is a signal. If you moved from position two to position three in a single run, that is weather.
Anyone reporting AI visibility as a single precise percentage without describing their sampling method is either not measuring carefully or is hoping you will not ask.
What the baseline actually buys you
Three things, all of which are hard to get any other way.
It gives you an honest starting point, so improvement can be demonstrated rather than asserted. It gives you a prioritised work list, because the reasons models cite for choosing competitors are usually specific and fixable. And it gives you an early warning system for the most damaging failure mode of all — an assistant confidently repeating something about your business that is out of date or simply wrong.
The mistake almost everyone makes
Starting the work before recording the baseline.
Six months in, someone asks whether the programme is working. Without a saved record of what assistants said before anything changed, the honest answer is that nobody knows — and the dishonest answer is whatever the agency needs it to be.
The baseline takes an afternoon. It is the cheapest insurance in the entire discipline, and it is the step most consistently skipped.