How often two identical prompts name the same brands, measured over three runs of one prompt set on one morning
Same 22 prompts, same two engines, three runs inside six minutes on 2026-10-22. A brand that appeared at all appeared in all three runs 53.0% of the time on ChatGPT and 26.1% on Gemini. 132 readings, with the prompt set and the method printed.
Published by AI Knows Us · Updated 22 October 2026
We sent the same 22 buying questions to ChatGPT and to Gemini three times each, inside a six minute window on 2026-10-22, and scored which brands each answer named. On ChatGPT, a brand that appeared at all appeared in all three runs 53.0% of the time (35 of 66 brand and prompt pairs, 95% interval 41.2% to 64.6%). On Gemini it was 26.1% (30 of 115, 95% interval 18.9% to 34.8%).
Every other number published about AI visibility rests on this one, and we could not find anybody publishing it. If one reading of one prompt is a coin flip on which brands come back, then a single run figure, including the ones we publish, needs a repeat count attached before it means anything.
What was measured
| Item | Value |
|---|---|
| Date | 2026-10-22, three runs between 05:10 and 05:15 UTC |
| Prompts | 22, frozen before the first run |
| Runs per prompt per engine | 3 |
| Engines | ChatGPT and Gemini |
| Readings in total | 132 |
| Market | United States, fixed in the prompt text |
| Readings that ran a live search | ChatGPT 66 of 66, Gemini 59 of 66 |
| Brand roster | 26 names, fixed in advance |
The four cuts of the same data
One consistency figure would hide the shape of this, so here are four, each with its denominator.
| Measure | ChatGPT | Gemini |
|---|---|---|
| Brand and prompt pairs seen at least once | 66 | 115 |
| Of those, present in all three runs | 35 (53.0%) | 30 (26.1%) |
| Of those, present in exactly one run | 18 | 53 |
| Prompts where all three runs named an identical set | 7 of 22 | 1 of 22 |
| Prompts where the first brand named was the same all three times | 5 of 22 | 4 of 22 |
| Mean overlap between two runs of one prompt | 0.584 | 0.424 |
| Prompts that named nobody in all three runs | 4 | 0 |
The overlap row is the one to put in a report. It is the average share of names two runs of the same prompt agree on: 0.584 on ChatGPT over 54 prompt pairs and 0.424 on Gemini over 66. Pairs where both runs named nobody are left out of it, because two empty answers agree trivially and would lift the figure for no reason.
What the numbers mean in practice
Take the hardest row first. Of 22 prompts on Gemini, 1 came back with the same set of names three times. On ChatGPT it was 7 of 22. So the normal case, on both engines, is that two runs of one prompt return partly different shortlists.
Then take the row that decides whether a dashboard means anything. On Gemini, 53 of 115 brand and prompt pairs showed up in exactly one of three runs. A brand tracker that ran that prompt once, on that morning, had close to even odds of recording a mention that two other runs did not produce, or of missing one that they did.
This is why a single number with no repeat count behind it is unreadable, and why a month on month chart built from one reading per prompt moves on its own. Our own guide to choosing the questions worth tracking has recommended three runs per prompt as the working minimum for a while. This page is the measurement behind that advice rather than the advice on its own.
Which brands held up and which did not
Counted once per reading in which the name appears, out of 66 readings per engine. A brand low in this table is not a small brand. It is a brand whose appearance on these prompts, on this morning, depended on which run you looked at.
ChatGPT
| Brand | Readings naming it |
|---|---|
| Profound | 32 of 66 |
| Otterly | 31 of 66 |
| Peec | 28 of 66 |
| Semrush | 26 of 66 |
| Ahrefs | 12 of 66 |
| Athena | 9 of 66 |
| Scrunch | 5 of 66 |
| HubSpot | 3 of 66 |
| Conductor | 1 of 66 |
| Goodie | 1 of 66 |
| Rankscale | 1 of 66 |
Gemini
| Brand | Readings naming it |
|---|---|
| Otterly | 46 of 66 |
| Profound | 44 of 66 |
| Semrush | 35 of 66 |
| Peec | 27 of 66 |
| Ahrefs | 12 of 66 |
| SE Ranking | 11 of 66 |
| Athena | 5 of 66 |
| Frase | 4 of 66 |
| HubSpot | 4 of 66 |
| Writesonic | 4 of 66 |
| Surfer | 3 of 66 |
| Rankscale | 3 of 66 |
| Scrunch | 3 of 66 |
| Conductor | 3 of 66 |
| AgencyAnalytics | 2 of 66 |
| Evertune | 1 of 66 |
The two engines do not agree with each other either, which is a separate finding with a separate use. A shortlist built from one engine is a shortlist for that engine. The same brands counted across the full United States bank are published in which brands these assistants put forward for an American buyer, and the hosts behind those answers in which websites they cite. We put a worked example of the same effect on a single domain in one question won and one question lost, same day, same two apps.
Our own row in this run
Across the 132 readings these three runs produced, our own company was named in 0 of them and our domain appeared in 0 of the 3,981 citation URLs recorded. Cited and named are two findings here as everywhere else on this site, and they are reported apart rather than added. The same pair of zeros across a much larger bank, before and after we published 507 more pages, is written up in whether 507 new pages changed what these engines say about us.
Why this measurement had to be made rather than looked up
The plan that commissioned this page said our pipeline already repeated readings, so the raw material existed. It did not. The audit file holds 636 readings and 636 distinct engine and question pairs, which is one reading of each question, and the runner refuses to write a second reading of a question into the same file by design, because resuming a failed run is what that rule exists for. So the repeats were run fresh, under three separate run names, which is the only way the same code path produces three readings of one prompt.
That is worth saying plainly on the page. A measurement nobody can produce on request is an assertion, and this category is full of consistency claims with no run count attached.
The method, written so a stranger can repeat it
The prompt set is the 22 questions in three batches of our United States bank: tools and software, agencies and services, and competitor named alternatives. Those three were chosen by subject before any run, because naming a brand is the answer to that kind of question. Choosing prompts after seeing which ones produced brands would have picked the result.
Each question was sent as a single turn with a live web search requested, with a persona line and a market lock line, exactly as in the larger audit it came from. No question names our own company.
Scoring is once per reading in which a name appears, never once per occurrence, and the sources block is removed before the answer prose is read, so a name under a Sources heading counts as a citation rather than as a recommendation. Link text is kept, because a name a reader sees and clicks is a recommendation whatever the markup does. The roster is 26 names, fixed before the runs, and it is lifted out of the one scoring file this programme uses rather than written a second time: two hand written patterns in two scripts once produced two different figures for the same quantity and the wrong one went live.
The interval on each headline figure is a 95% score interval on the count and its denominator, so the precision on the page is the precision 3 runs of 22 prompts actually supports.
What this measurement does not cover
- One morning. Three runs inside six minutes on 2026-10-22. Drift over a day, a week or a model update is a different measurement and this is not it.
- Two engines. ChatGPT and Gemini. Our product reads 5 surfaces and three of them are outside this run.
- 22 prompts in one category. These are questions about AI visibility tools and agencies. A different category can be steadier or noisier and nothing here says which.
- Three runs, not thirty. 3 runs give the intervals printed above and no more precision than that. More runs narrow them.
- A fixed roster of 26 names. A brand outside that list is invisible to this scoring, so the totals are a floor rather than a census.
- Gemini searched 59 of 66 times. The readings that did not search describe training data rather than the live web, and they are inside the figures above.
- We sell the remedy. We charge money to measure and improve brand visibility in these answers, and this page says that one reading of one prompt is close to a coin flip. Repeat it from the prompt set below rather than take our word for it.
Who should not buy the product behind this page
Anybody who wants a single AI visibility score, delivered once a month from one reading per prompt, should not buy it from us, because this measurement is the argument against that product. Anybody whose procurement needs single sign on, SAML, SOC 2, ISO certification, a data processing addendum or a white label should read our pricing page first: none of those words render on it, checked by word boundary on 2026-10-22. Profound's own pricing page, read 2026-10-22 at tryprofound.com/pricing, publishes SSO, SAML and SOC 2 on Enterprise and up to 9 answer engines tracked, with the price given as custom.
On plan choice, SEO Start at $149 renders 0 questions, 0 credits, 0 engines, so it is the wrong plan for this work. Growth at $299 is the cheapest plan that reads all 5 surfaces.
Questions about run to run consistency
How many runs per prompt should a measurement use
Three is the working minimum on this evidence, because on 22 Gemini prompts only 1 returned an identical set of names three times. More runs narrow the interval, and the interval is what tells you whether a change between two months is real.
Does this mean AI visibility numbers are meaningless
It means a single reading per prompt measures a tendency, and the way to read a tendency is with a repeat count and a range beside it. The same data gave stable answers too: 35 of 66 ChatGPT brand and prompt pairs held across all three runs.
Why is Gemini less consistent than ChatGPT here
On these 22 prompts Gemini named more brands per answer, which gives more room for a name to appear in one run and not another, and it ran a live search in 59 of 66 readings against ChatGPT's 66 of 66. Both are visible in the table above and neither is a claim about the engines in general.
Could the prompt set have been chosen to produce this result
The three batches were fixed by subject before the first run and every question in them is printed below. Picking prompts after reading the answers is the one choice that would decide a number like this, so it was closed off in advance.
What would make this figure move
A wider prompt set across several categories, more runs per prompt, and the same measurement repeated after a model update. Each of those is a separate run against the same frozen set, and the set is published for exactly that reason.
The 22 prompts, printed so this can be repeated
- What is the best tool to track whether AI assistants mention my brand?
- How do I measure my share of voice in AI answers?
- What are the best AI visibility tracking tools right now?
- What software monitors a brand across ChatGPT, Claude, Perplexity and Gemini?
- What is the best generative engine optimisation agency?
- What tools track AI search rankings and citations?
- Is there a free tool to check whether ChatGPT recommends my business?
- How much does generative engine optimisation cost?
- Which agencies do answer engine optimisation well?
- Should I hire a GEO agency or build the capability in house?
- What does a generative engine optimisation agency actually do, week to week?
- How do I choose an AI SEO agency and what should I ask them?
- Is Profound worth the price for AI search monitoring?
- What is the cheapest AI brand monitoring tool?
- Which AI search tracking tool covers the most engines?
- Is there an open source alternative for tracking AI mentions?
- What are the best alternatives to Profound for AI visibility tracking?
- Which AI visibility platform has the best free tier?
- What is the difference between an AI visibility tool and an SEO tool?
- How does Otterly compare with Peec AI?
- What do users complain about with AI visibility tools?
- Which AI visibility tools integrate with Google Analytics?
Each one was sent with a persona line, a United States market lock line, and one instruction to run a live web search and name the specific tools, platforms, agencies or websites the engine would genuinely recommend and in what order, with the real source URLs it retrieved. The wider dataset these prompts come from is published as 597 readings across ChatGPT and Gemini in three markets.