Straight answers to the questions buyers ask us most. Each one starts with the short answer; the reasoning and the numbers follow. Disclosure: we sell this kind of measurement (services and prices), and the method below is the one we use — including on ourselves.
How do you measure brand visibility in ChatGPT accurately?
Ask a fixed set of real buyer questions many times, with web search on, and report the share of answers that mention you with a confidence interval — not a single score. One screenshot or one run tells you very little: the same question returns a different set of brands from run to run.
Fix the questions before you look at results. Use 20–30 questions buyers actually ask, in their words. The question set defines the market you are measuring, so publish it next to the numbers.
Turn web search on and record who answered. Log the model, the date and the channel (the official API or another programmatic route). Programmatic answers are a proxy for the logged-in ChatGPT app, which adds memory, personalization, location and model routing.
Repeat. At least 7 runs per question per day, pooled over two to four weeks before you read a trend (the next question explains why).
Count three things separately: whether the answer names you, whether it cites one of your pages, and whether it recommends you. A mention can be a warning.
Compute intervals by question. Runs of the same question are correlated, so resample questions first and runs within each question second (a prompt-clustered bootstrap).
Compare yourself with competitors inside the same answers. The difference measured on the same answers cancels much of the shared noise and is far more sensitive than comparing two separate intervals.
Keep every answer verbatim with its timestamp, so any number can be recomputed and audited later.
How many times should you run a prompt to track AI search visibility?
At least 7 runs per question per day, pooled over two to four weeks — and beyond that, add questions rather than runs. Below 7 runs per question we label results insufficient; with fewer than 14 days of data we call them a snapshot, which shows structure but not change.
Runs of the same question are not independent: they share whatever that question pulls up. The intraclass correlation (ICC) measures how much. With m runs of one question, those runs are worth about m / (1 + (m − 1) × ICC) independent answers. In our own baseline measurement — an OpenAI model with web search, ten buyer questions in our own category, 7 runs each, on 9 October 2026 — the median ICC of brand mentions across the brands we tracked was 0.23. At that level:
Runs of one question
Worth about this many independent answers
1
1.0
3
2.1
7
2.9
14
3.5
30
3.9
any number
at most 4.3
The first few runs remove most of the run-to-run noise. After that, the uncertainty comes from which questions were asked, and only more questions reduce it. ICC differs by category and engine, so measure it rather than assume it; we report it with every result.
Is AI visibility tracking accurate?
It can be accurate about what it measures — if it shows its uncertainty and says exactly what it measured. Five things decide whether a visibility number means anything:
Sampling noise. Answers change from run to run, so a rate without an interval and a sample size cannot be judged. Intervals that treat every run as independent are too narrow: in our simulations they covered the true value only 51–72% of the time instead of 95%.
Which questions. A different question set is a different market. Ask to see the list.
Automated sampling versus the app. Trackers sample through APIs or automated sessions; a buyer's logged-in app adds memory, personalization, location and model routing. We have not measured that gap yet and have not seen a published measurement of it, so treat it as unknown, not zero.
Mentioned versus recommended. In our sandbox experiment, a new brand was mentioned in over a quarter of baseline answers and recommended in none of them; several answers told people to wait for reviews.
Brand detection. Most tools match names and aliases as text. That misses paraphrases and can be fooled by common words, so someone should read a sample of the answers.
Questions worth asking any vendor, us included: How many runs per question per day? Are intervals computed by question? Which model and which channel answered? Can I get the answer-level data? Are mentioned, cited and recommended reported separately?
How do you prove GEO results to your CEO?
Show, against a baseline taken before anything changed, that specific changes were picked up by AI engines — and call a change in share real only when it is bigger than the noise. Screenshots and single scores do not survive a skeptical CEO; dated, reproducible evidence does.
Baseline first. Fix the question set and measure before any change goes live.
Log every change with its URL and the date it went live. Without that log, nothing can be attributed.
First month: pickup. For each URL, report whether AI engines retrieved it, cited it, named you in the same answer and recommended you, with the first date each happened. We count a page as picked up when it is cited in at least 3 runs within 30 days of going live.
Corrections. Check what the AI says about you against a fact sheet you sign, before and after.
Around day 90: share. Compare yourself with competitors inside the same answers. A change is real only if its 95% interval excludes zero. If every tracked brand moved the same way, it was probably the platform, not your work. “Too early to tell” is an honest verdict.
Keep mentioned, cited and recommended apart in the report. A rising mention rate can be rising criticism.
What not to promise: a ranking or a share target. Nobody controls AI answers. You control your own pages, and you can earn coverage on the third-party sites the AI already cites.