Most AI search visibility numbers are a single screenshot of a system that gives a different answer each time you ask. A study published in January 2026, built on 2,961 prompt runs, found that ChatGPT and Google’s AI features almost never returned the same brand list twice. That makes measurement design, not tool choice, the first decision a business evaluating GEO (Generative Engine Optimization) services has to get right.
- SparkToro and Gumshoe.ai found that ChatGPT and Google’s AI answers returned the same brand list in fewer than 1% of repeated runs, so one-off checks mislead.
- Visibility rate across many repeated prompts is a sturdier metric than any single answer or ranking position.
- Separate brand mentions from linked citations, and report each AI platform on its own rather than as a blended score.
- Before paying for any tracking tool or GEO service, ask how many runs sit behind each number.
Why a single check tells you almost nothing
AI answer engines are probabilistic, so asking the same question twice does not reliably produce the same brands. In a study published in January 2026, SparkToro co-founder Rand Fishkin and Patrick O’Donnell of Gumshoe.ai analysed 2,961 prompt runs across ChatGPT, Claude, and Google’s AI Overviews, with Google’s AI Mode used when Overviews did not appear. Each of 12 brand-recommendation prompts was run 60 to 100 times per platform. ChatGPT and Google’s AI returned the same brand list in fewer than 1% of repeated runs, and the same list in the same order in fewer than 0.1%.
The implication is practical. If a vendor shows you a screenshot where your brand appears, or does not, that image is one draw from a distribution. Decisions about budget should rest on the distribution, which means many runs of the same prompt.
What to track instead
Visibility across many runs is more stable than position within any one answer. The same SparkToro and Gumshoe.ai study reports that in its headphone prompts, Bose, Sony, Sennheiser, and Apple each appeared in 55% to 77% of 994 responses, even though the exact lists and their order kept changing. Fishkin and O’Donnell concluded that ranking position inside an AI answer is largely noise, while the share of responses that mention a brand carries real signal.
For a business, that points to a simple definition: Visibility rate: the share of repeated runs of a buyer-style prompt in which your brand is named. Track it per prompt and per platform, then watch the trend over weeks instead of reacting to any single result.
Four metrics, compared
These definitions are our recommended working set, not an industry standard. No single accepted standard for AI visibility reporting exists yet, so agree on definitions before you compare any two reports.
| Metric | What it counts | Best used for | Main limitation |
|---|---|---|---|
| Visibility rate | Share of runs that name your brand | Headline trend across a prompt set | Says nothing about how favourably you are described |
| Citation rate | Share of runs that link one of your URLs as a source | Judging whether content is being used as evidence | Some platforms rarely show links |
| Share of voice | Your mentions divided by all brand mentions in the category | Comparing against named competitors | Depends heavily on which competitors you include |
| Sentiment and framing | Whether answers recommend, caveat, or dismiss you | Spotting reputational problems | Needs human review of sampled answers |
A repeatable measurement routine
A routine that holds up under the variance above is a controlled experiment, repeated on a schedule. Keep every setting constant except the thing you changed, so that movement in the numbers can be linked to something you actually did.
- Write 15 to 30 prompts in the language a buyer would use, covering problem-stage, comparison-stage, and vendor-stage questions.
- Fix the platforms you will test, and note whether each run is logged in or out and which location it uses.
- Run every prompt at least 10 times per platform, and save the raw answers rather than only a score.
- Calculate visibility rate and citation rate per platform, and keep them separate.
- Repeat monthly, and annotate the chart with content or technical changes you shipped in between.
Ten runs is a pragmatic floor for a small business, well below the 60 to 100 used in the research study, so treat small movements as noise.
Questions to put to a tracking vendor or GEO provider
Fishkin’s recommendation in the study’s coverage was that providers selling AI visibility tracking should explain their methodology and show their math. That standard applies equally to agencies that report GEO progress. Their reports are only as credible as the sampling behind them.
Ask how many runs sit behind each percentage, whether prompts were fixed or rewritten between periods, and whether answers were gathered through consumer interfaces or APIs. Ask also how mentions and linked citations are separated, and whether raw outputs can be exported for your own audit. A provider that cannot answer these is reporting an impression, not a measurement. The same scepticism we apply to agentic SEO vendor claims belongs here.
What the evidence doesn’t support
Several limits are worth stating plainly. The SparkToro and Gumshoe.ai research was conducted in partnership with Gumshoe.ai, which sells AI tracking tools, and it is not peer reviewed; its authors called for larger follow-up studies. It used controlled prompts, and real users phrase questions far more variably, so real-world variation is likely higher still. It also tested a small set of 12 prompts, so category-level conclusions should be treated as indicative. Finally, many vendor blogs publish per-platform citation percentages without disclosing sample sizes, and we have left those figures out rather than repeat numbers we could not trace to a method.
If you are working out where GEO fits in the broader picture of AI search and SEO, measurement is the part that keeps the rest honest.
Frequently Asked Questions
How many times should a prompt be run to trust the result?
There is no agreed threshold. The SparkToro and Gumshoe.ai study used 60 to 100 runs per prompt and platform. For a small business, at least 10 runs per prompt gives a rough visibility rate, and more runs make it steadier.
Is being mentioned by an AI tool the same as being cited?
No. A mention names your brand in the answer text, while a citation links one of your pages as a source. Some platforms show links rarely, so the two should be tracked separately.
Can a business measure AI search visibility without paid tools?
Yes, at small scale. A fixed prompt list, repeated manual runs, and a spreadsheet of saved answers produce a usable baseline. Paid tools mainly save time and extend the number of runs.
Measuring AI visibility properly is the step that turns GEO from a vague promise into something a team can manage. At Grow with Hafsa we build this kind of measurement into AI marketing and performance marketing engagements, so decisions about content and agent workflows rest on repeated evidence and not on a lucky screenshot.
Two related checks sit upstream of measurement: whether search crawlers can reach your pages at all, covered in which AI bots to allow in robots.txt, and whether special files help, covered in what the evidence shows on llms.txt.