AEO and GEO
How do you measure AI visibility across ChatGPT, Perplexity and Gemini?
Measuring AI visibility takes repeated checks on the same realistic customer questions, because an assistant often names a different business first when asked again. In our own checks across 16 Central Texas towns (800 total answers, 906 repeat pairs), asking the exact same question again put the same business first only 53% of the time. That is why 56 publishers recommend a fixed prompt panel and 29 recommend running each prompt multiple times.
Updated 19 November 2026
To measure AI visibility, ask a fixed set of customer questions again and again, and pair those answer checks with your own referral data and customer intake. In our corpus of 320 publishers, 56 back a fixed panel of prompts, 47 back finding which sites the assistants cite, 35 back analytics set up to see AI referrals, and 29 back running each prompt many times. A single spot check in your own browser tells you little, because the same question can return a different first pick the next time. Asking new customers how they found you (23 publishers back it) catches what analytics misses.
What sources say to measure
Each row below is one tactic, and each cell counts how many of the 320 publishers in our 435 saved sources took that stance. A publisher can take more than one stance, so the columns do not add up to a total.
| Step | Recommend it | Measured it working | Call it useless or warn | Measured no effect | Measured harm | Our label |
|---|---|---|---|---|---|---|
| A fixed panel of prompts run on a schedule; share of answers naming the business | 54 | 4 | 11 | 0 | 0 | Recommended and measured |
| Run each prompt many times, since answers vary | 25 | 6 | 5 | 1 | 0 | Recommended and measured |
| Find which sources the assistants cite for the target prompts, then get onto those | 41 | 10 | 4 | 0 | 0 | Recommended and measured |
| Find the real prompts customers use | 18 | 3 | 1 | 0 | 0 | Recommended and measured |
| Analytics set up to see AI referrals | 35 | 1 | 10 | 0 | 0 | Contested |
| Ask every new customer how they found the business | 19 | 4 | 2 | 0 | 1 | Recommended and measured |
| Track how AI visitors convert against other visitors | 26 | 8 | 5 | 1 | 0 | Recommended and measured |
How much answers vary when you ask again
Between 14 November 2026 and 17 November 2026, we ran 16 town checks in Central Texas for HVAC and plumbing (800 answers, comparing 906 repeat pairs of the exact same question on the same assistant). Across all three assistants, the same business came back in first place only 53% of the time: 74% on Perplexity, 46% on ChatGPT, and 35% on Gemini. All three assistants agreed on the same top business in just 1 of the 16 towns.
| Assistant | Same first pick when asked again |
|---|---|
| ChatGPT | 46% |
| Perplexity | 74% |
| Gemini | 35% |
| All three | 53% |
How tracking tools collect their answers
If you are choosing a tracking tool, ask how it gets its answers: from the logged-out app people actually use, or from a developer API that may search differently. Below are two trackers' own method notes, quoted from their sites; the stability figure in the first is the vendor's own number.
Chatgpt.com in a browser, logged out and set to your region.
The official Sonar API, with its sources.
Same mention result next day. Based on 1,034 comparisons, July 2026.
MentionFlow, Methodology, read 19 November 2026
Peec AI’s UI scraping simulates real user interactions, so the data it collects matches what the average user sees on these AI platforms.
Peec AI, Peec AI docs: how Peec AI collects data, read 19 November 2026
What this means for your business
- Never judge your visibility from a single prompt: run each question several times before deciding whether you gained or lost ground.
- Test the full questions customers actually ask, not short keyword strings copied from a rank tracker.
- Add "Did you ask an AI assistant?" to your intake form or first call, so you hear about the customers no analytics tool can see.
We saved 435 guides, studies, forum threads and platform docs from 320 publishers in full, recorded 1592 word-for-word method statements across 99 tactics, and logged 216 reported results with numbers. Every count is by publisher, not by article, so one prolific blog only gets one vote per stance, and every quote is checked word for word against our saved copy. A high count is not proof that a tactic works: many GEO writers sell GEO services or software, and a measured claim usually means one site's before-and-after numbers rather than a controlled test.
Common questions
Why do ChatGPT and Gemini give different answers to the same question?
The assistants generate each answer fresh, and the web searches behind them can return different pages from one run to the next. In our 906 repeat pairs across 16 town checks, the first-named business stayed the same only 53% of the time overall (46% on ChatGPT, 74% on Perplexity, 35% on Gemini).