The question arrives from clients now, usually phrased as a worry: do we show up when someone asks the AI? It is answerable. It is just answerable badly in about four different ways, and the bad answers all look like data.
1. Ask the question a buyer would ask
The single most common mistake is typing the brand name into the prompt. “What do you think of Acme Consulting?” will produce a paragraph about Acme Consulting. That is not visibility, it is an echo.
The question worth measuring is the one asked before anyone knows you exist: best tax advisor in Vienna, which analytics tool for a small agency, who should I use for GDPR-compliant reporting. If your brand shows up in the answer to that, you have visibility. If it only shows up when you name it, you have a search result.
2. Ask the model a real person is actually served
Answer engines route different users to different model tiers, and the cheap tiers give noticeably different answers from the ones behind a consumer app. Measuring a tier nobody is served produces a number that is precise and irrelevant.
The same applies to web access. A model answering from training data alone is describing the world as of its cutoff; the same model with browsing enabled will cite pages published last week. Those are two different questions and they deserve two different answers — pick the one that matches how your buyers actually reach the model, and write down which you chose.
3. Repeat, because one answer is not a measurement
Ask an engine the same question twice and you will often get two different lists. That is not a bug; generation is probabilistic. The consequence for reporting is strict: a single check cannot tell you anything moved. You need the same prompt set, on the same cadence, across the same engines, for long enough to have a baseline.
Weekly is the sweet spot — frequent enough to catch a real change within a reporting cycle, infrequent enough to stay affordable when you are doing it for a dozen clients across five engines.
4. Record more than yes or no
“Mentioned” is the crudest possible reading of an answer. Two things change the meaning of a mention:
- Position inside the answer. Named first, in the opening sentence, is a different outcome from being the fourth item in a list at the bottom.
- What the answer says about you.A model can name you and then recommend someone else, or attach a caveat. “Mentioned in eight answers, three of which steer buyers elsewhere” is a truer sentence than “mentioned in eight answers”, and it points at different work.
Keep the full answer text, not just the verdict. When a client asks in November why the number fell in September, the archived answers are the only thing that can tell you.
The one thing nobody can promise
No tool, agency or consultant can guarantee that a language model will recommend a brand. There is no bidding surface, no submission form and no support queue. What can be promised is an honest measurement, a trend line, and the page-level work that plausibly influences it. Anyone selling the guarantee is selling something they cannot deliver.
Running it without building it
The method above works entirely by hand — open five engines, ask your prompt set, write the results in a spreadsheet, repeat weekly. It costs nothing but time, and the time is the reason most people stop after two weeks.
If you want the first answer without the spreadsheet, our free AI-visibility check runs one unbranded prompt across five engines and tells you which of them name you. No account, no personal data — it is the same first question, just faster.