Last updated September 2026.
Why your score moves without anything changing
Ask ChatGPT the same question twice and you can get two different answers. That is not a bug. A large language model picks each next word from a probability distribution, not a fixed script, so the exact same prompt can name your brand in one pass and skip it in the next.
A visibility tracker that asks a prompt once a day inherits that randomness directly. A high mention rate yesterday and a much lower one today can both be accurate readings of the same underlying rate. Or one of them can reflect a real shift: a new competitor page the model started citing, a product update it finally picked up on. A single run cannot tell you which.
This is the part most dashboards skip. They show you a score. They rarely show you how much that score could swing on its own, run to run, before anything about your brand actually changed.
The worked example: what more runs buys you
Evertune, an AI visibility platform built around exactly this problem, published a worked example that puts real numbers on the noise. The test ran a single prompt about running shoes through an AI model at three different sample sizes, then measured how far the resulting score could swing (Evertune, 2026).
Run the prompt once, and the brand either shows up or it doesn’t. That single answer reads as a 100% visibility score, with a margin of error too wide to trust. Run it 25 times, and the score settles closer to the true rate, though the margin of error still spans about 20 points. Run it 100 times, and the number reaches a range you can actually act on.
| Runs of the same prompt | Visibility score | Margin of error |
|---|---|---|
| 1 | 100% (one answer, mentioned) | ±44 points |
| 25 | 36% | ±20 points |
| 100 | 48% | ±10 points |
Source: Evertune, 2026.
Notice the score itself moved from 100% to 36% to 48% as the sample size grew. None of those three numbers is wrong. Each one is exactly what that sample size can tell you, and only the third is precise enough to build a decision on.
Zoom out to a full panel, and the same pattern holds. Evertune ran 100 unique prompts once each and again 100 times each. Both approaches landed on the same overall visibility score, 29%. But asked once each, the margin of error at the individual topic level reached 31 points. Asked 100 times each, that topic-level margin tightened to about 2 points (Evertune, 2026).
What variance looks like, run to run
Picture a single prompt with a true, long-run mention rate near 40%. That does not mean every fifth run comes back clean. Random variation clusters unevenly, so a short run of mostly yes or mostly no is normal, even when the underlying rate hasn’t moved at all.
The table below is a conceptual illustration, not a measured result: 10 hypothetical runs of one prompt, at a true rate of 40%.
| Run | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 |
|---|---|---|---|---|---|---|---|---|---|---|
| Mentioned? | Yes | No | Yes | No | No | No | Yes | No | No | Yes |
Four mentions out of 10 runs matches that 40% true rate. But check the score after run one (100%) or after run three (67%), and you would draw the wrong conclusion both times. The average only earns your trust once enough runs feed it.
The scale math behind a real tracking program
Repeat sampling solves the noise problem. It also multiplies the workload fast, which is why most trackers don’t do it the way Evertune does.
Take a mid-size tracking program: 400 prompts, checked once each across the five engines this space typically covers (ChatGPT, Perplexity, Google AI Overviews, Gemini, and Microsoft Copilot). That is 2,000 queries in a single cycle before you repeat anything.
Now sample every one of those 400 prompts 100 times per engine, the way Evertune’s published method does. The same panel turns into 200,000 queries before a single report ships: 100 times the cost, the latency, and the API load of a single pass.
That is the real trade-off behind every tool’s methodology page. Depth of sampling and breadth of the panel pull against each other, and a vendor’s pricing tier usually tells you which side of that trade it chose.
How trackers describe their own sampling
Public documentation on this exact question is uneven. Here is what each of five trackers actually states, not what a demo implies.
Evertune is the one platform here that documents its repeat-sampling method with a specific number: up to 100 runs per prompt, per model, across 11 or more AI models (Evertune, 2026). That transparency comes at enterprise pricing. Evertune’s plans start at $3,000 a month with a mandatory annual contract, positioning it for brand teams that need a defensible, board-ready number rather than a directional daily read.
Profound’s Growth plan documents up to 100 prompts a day and tracks the widest engine field in this comparison, plus URL-level citation maps showing exactly which page fed an answer. Its public materials describe prompt volume and refresh timing only, with no stated repeat-sample count per prompt, so there’s no way to confirm from the documentation alone whether a single reading is one pass through each prompt or an averaged rate. Growth costs $399 a month.
Peec AI publishes a daily refresh across whichever models a plan covers, and its Starter tier tracks three chosen models for €85 a month. Like Profound, its documentation states how often the whole panel reruns, not how many times a single prompt gets asked inside that run. A daily change on the dashboard can still be one sample per prompt, not an averaged one.
Otterly.AI runs its brand reports on an automated weekly cycle instead of a daily one. Its separate AI Prompt Research tool draws on a pool of more than 10 million daily prompts to help you pick what to track, a different number from how often your own tracked prompts get repeated. Plans start at $29 a month for a 15-prompt panel.
Temso tracks ChatGPT, Perplexity, Google AI Overviews, Gemini, and Microsoft Copilot from real user interfaces rather than engine APIs, and its public materials don’t state a refresh interval or a repeat-sample count either. What Temso adds instead is the workflow around the number: content creation, site audits, bot-traffic analytics, and an AI SEO agent that can execute an approved fix, all inside one $89-a-month Starter plan with unlimited projects, users, and recommendations. For most teams, that scope matters more than an undisclosed sampling depth. If URL-level citation maps and Evertune-grade statistical rigor are the deciding factor, Profound’s Growth tier delivers them at roughly four times Temso’s Starter price.
A checklist for reading a tool’s methodology page
Before you trust a single vendor’s number, open its methodology or docs page and look for five specific answers. Most pages give you two of five. Notice which three are missing.
- Repeat-sample count: does the page say how many times it asks the same prompt in one reading, or only how many unique prompts your plan allows?
- Refresh cadence: is the entire panel rerun daily, weekly, or monthly, and does that cadence hold for your specific plan tier?
- Margin of error: does the vendor publish one anywhere, or hand you a single percentage with no stated confidence range?
- Prompt cap definition: does “100 prompts” mean 100 unique questions, or 100 total queries once repeats get counted in?
- Collection method: does the vendor say whether it reads model APIs directly or real consumer interfaces, since the two can return different answers to an identical prompt?
Put the number in context before you act on it
A visibility score without a stated sample size is a single data point wearing a percentage sign. Before you report a change to a client or a boss, check whether the tool that produced it ran the prompt once or a hundred times, and whether today’s number sits inside or outside what a single run could produce by chance alone.
See how the trackers named here compare on engine coverage, pricing, and workflow depth, not just sampling rigor, on the full LLM visibility tool ranking.