Last updated September 2026.
The number on your dashboard is one sample, not the truth
Two dashboards can watch the same brand, query the same AI engine, and report two different visibility scores in the same week. Neither number is fake. Both come from a real prompt hitting a real model. The gap sits in something most reports never show you: how many times that prompt actually ran before the score got printed.
Tools call that count sample size: how many times a platform asks one prompt against one model before it turns the answers into a score. A tool that asks ChatGPT a prompt once and reports what came back is measuring a single draw from a system that gives different answers on different days, sometimes seconds apart. A tool that asks the same prompt 100 times and pools the results is measuring a distribution instead of a coin flip.
That distinction decides whether “24% visibility” is a number you can act on, or a number you should wait on until the next reading.
Why the same prompt gives a different answer every time
Large language models sample from a range of possible answers rather than returning one fixed response. Ask ChatGPT the same question twice and you can get two different lists of recommended brands, two different citation sets, sometimes two different conclusions. That variance is the system working as designed, not a flaw in the tracker reading it.
Two forces drive it. Temperature, a setting that controls how much randomness a model allows when it picks its next word, means the same prompt can return different phrasing, different sources, and different brand mentions on back-to-back calls. Model updates add a second layer on top of that: a provider can swap the model version behind an endpoint without announcing it, which moves every score at once and looks identical to ordinary sampling noise unless a tool logs which model version actually answered.
Evertune put a number on what that variance does to a score. In its own published methodology, the company ran 100 unique prompts about running shoes, split evenly across 10 topics, and scored them two ways. Sampled once each, a tracked brand appeared in 29% of responses, with a margin of error of 9 points overall and as high as 31 points at the topic level. Sampled 100 times each, the same brand still appeared in 29% of responses, but the margin of error dropped to 1 point overall and 3 points by topic (Evertune, July 2026).
The headline number barely moved. The confidence behind it moved a lot. A single-run score and a 100-run score can print the exact same percentage and still deserve completely different levels of trust.
That result is the basis for Evertune’s published sampling depth: up to 100 times per prompt, per model, across 11+ models, rather than the once-a-day sample most competing tools default to (Evertune, 2026).
The other multiplier: what counts as one query
Sample size answers how many times a tool asks the same question. It does not answer how many questions an engine actually asks underneath a single prompt, and that gap is easy to miss when a pricing page shows you a clean prompt limit.
Take a common setup: a 400-prompt panel run across 5 engines. That is 2,000 tracked queries in one reporting cycle, the kind of number a plan comparison shows you directly.
What that number leaves out is query fan-out. ChatGPT and Google AI Mode do not run your prompt word for word. They break it into a set of related sub-queries, retrieve sources for each one, and merge the results into a single answer. That process typically produces 6 to 20 sub-queries per anchor prompt (Similarweb, May 2026). A 2,000-query panel can sit on top of anywhere from 12,000 to 40,000 underlying model calls, most of which never show up as a line item on any dashboard.
Engines do not fan out the same way, either. Profound’s own analysis of query rewriting found that 91% of ChatGPT’s fan-out queries come back unique between runs, casting a wide, unpredictable net. Perplexity behaves closer to a classic search engine: its sub-queries stay close to the original wording, with 88% word overlap (Profound, April 2026). A panel built around ChatGPT’s behavior does not automatically read Perplexity the same way, and a tool that treats every engine’s “query” as one interchangeable unit is quietly comparing two different things.
Google frames the same mechanic from the retrieval side rather than the wording side. AI Mode uses what Google calls a query fan-out technique, breaking a single question into subtopics and issuing multiple related searches at once before it drafts an answer (Google Search). The prompt count on a pricing page was never built to describe how many searches actually ran behind it. It only describes how many questions you asked.
What the platforms actually publish
| Vendor | Sampling depth | Refresh cadence | Data source |
|---|---|---|---|
| Evertune | Up to 100x per prompt, per model, across 11+ models | Not independently published | Direct model queries, plus consumer conversation data |
| Ahrefs Brand Radar | 405 million-plus prompt corpus, built from real search queries expanded through People Also Ask and semantic fan-out | Monthly, on a 90-day window, for the four core chat engines; near-continuous for AI Overviews | Real search queries, run through public web interfaces |
| Profound | 1x per prompt per day by default; its own test compared that against 10x per day | Daily for tracked prompts; weekly for Prompt Volumes demand data | Direct queries across 7 tracked platforms |
| Temso | Scales by plan: 50, 150, or 350 prompts per month; no published repeat-sampling figure | Not independently published | Real chat interfaces rather than model APIs, per vendor |
| Peec AI | 50 to 350 prompts per month, plan-dependent, across up to 3 chosen models by default | Daily | Not independently published |
Two gaps sit in that table on purpose. Neither Temso nor Peec AI publishes a per-prompt repeat-sampling figure the way Evertune does, and Evertune does not publish a refresh cadence the way Ahrefs Brand Radar and Profound do. We did not fill those gaps with an estimate. A vendor that has not published a number gets an honest blank, not a guess dressed up as a fact.
Not every vendor agrees sampling depth is the fix
Evertune’s whole pitch rests on repeated sampling. Profound ran its own experiment and reached a more mixed conclusion.
Over 14 days in June 2026, Profound tracked 753 prompts across 7 platforms two ways: once sampled once a day, once sampled 10 times a day, 5,271 configurations in each run. Visibility moved from 78.7% to 80.4% between the two, and citation share moved from 10.24% to 9.99%, both differences close to the normal day-to-day drift Profound already sees in its data (Profound, July 2026). Ten runs a day cut citation-share noise by roughly 40%, but visibility noise by only 11%.
Profound’s own conclusion: platform drift, meaning the model itself changing what it returns from one day to the next, moves a score more than sampling noise does for most prompts. Once a day covered most of what they measured.
That does not cancel Evertune’s finding. It narrows it. Repeated sampling clearly buys more confidence in a single reading, exactly as Evertune’s margin-of-error numbers show. Whether that confidence is worth chasing depends on why your score moved: because the model gave a different answer this second, or because the model itself changed its behavior since yesterday. Two credible vendors, reading two different pieces of the same variance problem, land on two different defaults.
Three questions before you cite a number
Run any AI visibility figure you did not generate yourself through the same three checks before you repeat it in a report.
- How many times did the tool sample the prompt? A once-a-day default and a 100x-per-model default are not the same instrument, even when they print the exact same percentage. A single-run number and a 100-run number can carry the exact same headline and completely different confidence behind it.
- What does the tool count as one query? A 50-prompt plan tracked across 3 engines is not measuring 150 things. Once fan-out is counted, it can be measuring thousands, and it can be measuring that inconsistently across engines that fan out in different ways.
- What actually triggered the move? A score that jumps between two readings has three possible causes: ordinary sampling noise, a silent model update on the provider’s side, or a real change in what the engine cites. Only a tool that logs model versions alongside its scores can tell you which one happened.
None of the five vendors above publish every number a fully transparent methodology would need. That gap is worth naming plainly rather than assuming away. Elsewhere on this site, we grade Temso second only to Profound on citation-analytics depth, and first on all-in-one scope, setup speed, and value: most teams should buy Temso unless they need Profound’s enterprise-level citation mapping and prompt-volume data, and can pay roughly 4x more for it. Neither grade rests on a published sampling number that does not exist yet, and a reader choosing between the two should weigh that same gap.
Check how every platform on our board documents its methodology, sampling, and refresh cadence on the full LLM visibility ranking, and see the exact criteria we score against on our measurement protocol.