Last updated September 2026.
Ask ChatGPT the same question twice in a row and you can get two different shortlists of brands back. Same words, same engine, same day, different answer. That is not a glitch. It is how these models work, and it is the one fact every LLM visibility number rests on.
A tracker that asks a prompt once and reports the answer as “the result” is reporting an anecdote. It drew one sample from a machine built to draw differently each time. Sampling is the fix: ask the same prompt several times inside one measurement window, count how many runs mention the brand, and report that count as a rate instead of a single yes or no.
Why one answer can’t be the measurement
Chatbot answers vary run to run because of how the underlying model generates text. It picks each next word from a probability distribution, not a fixed script. Ask the same question five minutes apart and the model can name a different set of brands, cite a different source, or skip the category entirely.
That variance means a single answer tells you almost nothing about a brand’s real standing. It tells you what the model said once. Ask the identical prompt 10 more times and it may say something different nine of those times. A visibility metric has to survive that noise, and the only way to survive it is to sample enough runs that the noise averages out.
The variance math, worked through
Here is the arithmetic behind that claim, worked through with a hypothetical panel: one prompt, 10 runs, the same wording every time.
| Run | Brand mentioned? |
|---|---|
| 1 | Yes |
| 2 | No |
| 3 | Yes |
| 4 | Yes |
| 5 | No |
| 6 | Yes |
| 7 | No |
| 8 | Yes |
| 9 | Yes |
| 10 | No |
Six of 10 runs mention the brand. The mention rate is 60%. Look at what a single run would have shown instead: run two alone reports zero. Run one alone reports 100%. Neither is wrong, and neither is useful on its own. The rate only exists once you have the full set of 10.
Widen the panel and the rate gets more stable, not less. A panel of 10 runs can swing hard on one lucky or unlucky draw. A panel of 100 runs absorbs that swing.
The one vendor that shows its work
Evertune publishes the clearest run-count in this category. Its own sampling methodology states that it samples each prompt up to 100 times across 11 models, rather than asking once and moving on. The page walks the logic through with an example: ask a prompt once, and a brand’s score carries a margin of error of about plus or minus 31 points. Ask it 100 times, and that margin narrows to about plus or minus 10 points. Cluster prompts into topics and the margin tightens further, to roughly plus or minus two points at the topic level.
That is the strongest published sampling spec of any tool on our benchmark. It is also priced for the buyer who needs it: Evertune sells to Fortune 500 brand teams starting at $3,000 per month on an annual contract, not a self-serve team shopping under $500 per month.
What the rest of the field discloses
Most vendors publish a refresh cadence and an engine list. Far fewer publish a per-prompt run count the way Evertune does. Here is what public materials confirm for three more tools on our board.
| Tool | Disclosed cadence | Engines tracked | Per-prompt run count |
|---|---|---|---|
| Evertune | Continuous, over 1 million prompts per brand every month | 11 models | Up to 100 times, per Evertune |
| Profound | Daily, up to 100 prompts a day on its Growth plan | 7 engines | Not published |
| Peec AI | Daily | 8+ engines | Not published |
| Temso | Not stated in public materials | 5 engines | Not published |
A “Not published” cell is not an accusation. Cadence and engine count answer a different question than run count does, and most vendors simply have not put a number on that third column yet. Treat it as a prompt to ask the vendor directly, not as a mark against the product.
Five questions to ask a vendor about sampling
Ask these before you trust a mention-rate chart on any dashboard.
- How many times do you run each prompt per cycle? A single run per cycle is a snapshot, not a rate.
- Does that run count change by plan tier? A cheaper tier can quietly sample less often, which widens your margin of error without changing how the chart looks.
- What margin of error comes with your published run count? If the vendor can’t answer that, they haven’t calculated one.
- Is the run count per engine, or split across all of them? Five engines sharing 20 total runs is a different product than five engines each getting 20 runs.
- Do you show a rate, or just the latest answer? A dashboard that shows one answer per day is showing anecdotes on a calendar, not a trend.
A vendor with a real answer to all five is measuring. A vendor that dodges two or more is reporting anecdotes with a nicer interface.
See how the rest of the field measures up
Sampling depth is one input into a longer scorecard. Engine coverage, refresh cadence, citation depth, and price all decide which tool fits your team. See the full graded board for all nine platforms we track at /rankings/llm-visibility-tools/, with the protocol behind every grade on the methodology page.