How to measure whether AI is citing you, without buying a tool
There is no rank tracker for a conversation. But a disciplined manual sample of twenty questions produces a defensible trend line within a quarter, and costs an hour a month.
Key takeaways
- Sample a frozen list of about twenty commercially important questions monthly across the assistants your market uses.
- Record named, cited, position in the answer, and which competitor appeared instead.
- Use signed-out sessions and run from the target market; a single response is an anecdote, the monthly trend is the finding.
- Support the sample with assistant referral traffic and branded search volume, which is where AI awareness usually surfaces.
You measure AI visibility by sampling. Choose a fixed set of commercially important questions, run them through the assistants your market uses on a fixed schedule, and record three outcomes each time: whether your brand was named, whether your site was cited as a source, and which competitor appeared when you did not. Twenty questions, once a month, is enough to produce a real trend line.
Choosing the questions
Use the phrasing a customer would use, not the phrasing a marketer would. The most valuable questions are the ones immediately preceding a purchase decision: how much does this cost, who does this in my city, what is the difference between these two options, is this provider any good. Include a few brand questions asking what your company does, because the answer reveals what the models currently believe about you and where the errors come from.
Freeze the list. The temptation to quietly add questions you have started winning is strong, and it destroys the comparison. If the list must change, start a second one and report both until the first has enough history to retire.
What to record
- Named or not. A binary. Whether the brand appears in the answer text at all.
- Cited or not. Whether your domain appears as a linked source. Naming without citation is still valuable, and tracking them separately shows which kind of visibility you are building.
- Position within the answer. Being the first option mentioned is materially different from being fourth in a list.
- Who appeared instead. Over three months this becomes a competitor map that no keyword tool will give you, because it reflects what the models actually believe rather than what ranks.
Controlling for noise
Assistant responses vary between runs, between accounts and between regions. Reduce the noise rather than pretending it does not exist: use a signed-out or fresh session so personalisation does not contaminate the result, run each question from the market you are measuring, and accept that a single run is an anecdote. The trend across months is the finding, not any individual answer.
Record the date and the model version where it is visible. Providers update models frequently and a sudden shift across every question in the same month is far more likely to be a model change than something you did.
Connecting it to the business
Two supporting measures make the sample credible. Referral traffic from assistant domains in your analytics, which is small but real and reportedly converts at several times the rate of ordinary organic traffic. And branded search volume, because the most common path is that someone reads about you in an answer and later searches your name, which appears as branded demand rather than as a click on the article.
One answer is an anecdote. Twenty questions repeated monthly is a measurement.
Sources