Key takeaways
- Checking AI once shows if you appear today; measuring monthly reveals your share of voice and whether you are gaining or losing ground over time.
- The method is simple: a fixed set of real buying questions, run across ChatGPT, Perplexity, Gemini and Claude, logged as named, cited or absent.
- Two numbers carry the measurement: your mention rate across all prompts, and your share of voice against every rival named beside you.
- AI answers shift from one run to the next, so judge yourself on the monthly trend, never on a single day's result.
- The average brand is named in only 16.3% of AI answers, so a low first score is a starting line, not a verdict.
You can find out whether AI recommends your business in an afternoon. Working out whether that picture is improving or slipping, month after month, is a different job, and it is the one worth building a habit around.
To measure your AI search visibility, build a fixed set of real buying questions, put them to ChatGPT, Perplexity, Gemini and Claude on a set schedule, and log for each answer whether you are named, cited or absent. Turn those logs into two numbers, your mention rate and your share of voice, then re-run the whole thing every month so you are reading a trend rather than a single snapshot.
If you have never looked at all, start with the quick one-off version first. Our guide to checking whether AI recommends your brand walks through a single sitting that tells you where you stand today. This article is the ongoing cousin of that check. Same raw method, but built to run again and again, scored properly, so you can watch your visibility climb or slip over the months instead of guessing.
Why measure it, not just check it once?
Because AI answers are a moving target. The same question can return a different set of names from one week to the next, so a single check tells you about one moment, while measurement tells you about the direction of travel.
These tools are not consistent in the way a Google ranking is. In one study, researchers put the identical question to a range of large language models on five separate occasions and found some changed their answer up to 40% of the time (Journal of General Internal Medicine, December 2025). Run your own test twice in a fortnight and you will see the same wobble: named third one day, missing the next. That is not a fault in your method; it is the nature of what you are measuring, and exactly why one reading is never enough.
It is worth the effort because the audience is real and large. Around 900 million people now use ChatGPT every week (TechCrunch, February 2026), many of them asking it the questions they used to type into a search box. Being named in those answers is the new shortlist; being absent is quiet lost ground you never see.
Most businesses are absent more often than they would like. One 2026 analysis found the average brand is named in just 16.3% of the AI responses where it could appear, while category leaders manage 56.5%, roughly three and a half times as often (AthenaHQ, September 2026). A low first score is normal. What you do about it over the following months is the point.
How do you measure AI search visibility, step by step?
Fix a prompt set, run it the same way across the main engines, log the outcome of every answer, then repeat on a monthly schedule. Scoring turns those logs into trackable numbers, which we come to next. The discipline is in keeping every part of it identical each time.
-
Build a fixed prompt set from real buying questions. Aim for 10 to 20, a little more than you would use for a quick check. Mix four kinds: category questions (“best [service] company”), location questions (“[service] near me”), comparison questions (“most reliable [service] in Leicester”), and problem-shaped questions, because that is how people really ask (“my boiler keeps losing pressure, who can fix it in Leicester”). The word “fixed” matters most. Once the set is written, you do not change it, because the whole value of measurement is comparing like with like month after month.
-
Run the set across the main engines, signed out. ChatGPT, Perplexity, Gemini and Claude, each with web search switched on, in a private window so your own history does not flatter the results, and a fresh chat per question so the last answer does not colour the next. Run it the same way every time: same wording, same order.
-
Log the outcome of every answer. For each question, on each engine, record four things: whether you are named (yes or no), whether you are cited (your own site linked as a source), your position in the list (first, mid-list, or a footnote), and which rivals are named instead. Named and cited are not the same: one is the recommendation, the other is the evidence behind it, and they often need different work to fix.
-
Repeat monthly, same prompts, same method. One run is a single data point, not a trend. Put the test in the diary for the same date each month and treat it as a standing appointment rather than a when-I-remember job.
One note on what this measures: whether the engines name you, not how much traffic those answers send to your site. That visits side lives in your analytics, and our guide to tracking AI traffic in GA4 covers the referral data. This method measures the answer itself, upstream of any click.
What should you actually score?
Two headline numbers do most of the work: your mention rate, the share of all your prompts where you are named, and your share of voice, your mentions as a proportion of every brand named across those same answers. Track citation rate and average position alongside them.
| Metric | What it tells you | How to work it out |
|---|---|---|
| Mention rate | How visible you are overall | Answers that name you, divided by total answers, as a percentage |
| Share of voice | How visible you are versus rivals | Your mentions, divided by all brand mentions across the same answers |
| Citation rate | Whether AI trusts your own site as a source | Answers that link your site, divided by total answers |
| Average position | Whether you lead the answer or trail it | Where you appear when named (first, mid-list, footnote), averaged out |
| Monthly trend | Whether you are gaining or losing ground | The direction of all of the above, read across several months |
Mention rate is your absolute visibility: out of everything you asked, how often did your name come up. Share of voice is the competitive version, and the more honest one, because being named twice while a rival is named nine times is very different from being named twice in a quiet category. The same 2026 analysis that found a 16.3% average mention rate also found a brand’s own domain goes uncited in 84% of answers, which is why citation rate is worth logging on its own: being talked about and being used as a source are two different wins.
How often should you measure, and what counts as progress?
Monthly, with the prompt set and method held completely still, and progress is the trend across several runs rather than any single good day.
Because the answers wobble, one strong month proves very little on its own, and one weak month is not a crisis. Read the line, not the dot. A few months in, you will see whether your mention rate is genuinely climbing, whether your share of voice is gaining on the names that keep beating you, and whether you are cited more often as well as named.
Your first run is simply your baseline. Do not be alarmed by a low opening figure; be interested in it. The number to watch is next month’s, measured against this one.
The one rule that cannot bend is consistency. Change the wording of a question, test on a different kind of day, or let personalisation creep back in, and you have quietly broken your own baseline. The trend only means something if the method never moves. If you would rather see this sitting next to your other numbers, it is the sort of tracking we build into our measurement work, alongside rankings and traffic, so AI visibility is not a separate spreadsheet nobody opens.
What do you do when the score is low?
Treat every gap the measurement exposes as a to-do list. In practice that means answer-shaped content on your own site, consistent facts about you everywhere the engines look, and a presence in the sources they cite.
Measurement on its own changes nothing; it just tells you where to aim. The questions where you are never named are your content gaps. The rivals who are always named are worth studying, not because they are bigger, but because they are better represented in the places the machines trust. The sources that keep appearing behind the answers are your target list.
The work of turning those gaps into mentions is its own discipline, and we set it out in full in our answer engine optimisation playbook. Nobody can promise you a fixed spot in an AI answer, and anyone who does is worth avoiding. What you can do is give the engines a clearer, better-evidenced picture of your business and stack the odds in your favour.
Where doing it yourself gets hard
The method is free and genuinely works. The hard part is running it the same way every month, for months, without letting it slip or drift.
Three things tend to break the do-it-yourself version. Time, because an hour a month becomes the task that always slides. Consistency, because it is easy to change a word or a day without noticing you have moved the goalposts. And coverage, because holding a clear cross-engine picture in your head, or spotting a two-point drift over six months, is harder than it sounds. Most owners run it once, learn something, and never run it again.
That is why we built a tracked version. Our free growth review runs your real buying questions across the same engines, scores them the way described here, and hands you a plain-English read on where you are named, where you are not, and who is winning the answers you want. It is free, there is no lock-in, and no obligation to carry on afterwards. If you would rather talk it through first, get in touch and we will point you in the right direction.