There’s a useful line buried in a Peec AI blog post from earlier this year. Ethan Smith, CEO of Graphite, contributing expert commentary to a piece on AI search measurement, writes: “10 is enough for a quick estimate for entity comparison prompts.” He goes on to recommend running multiple variations of the same prompt, tracking the percentage of time your brand appears, and using both automated tools and manual logged-in checks. 

It’s practical advice, and the “quick estimate” framing is the important part. Smith is naming a threshold for a specific job: getting a fast read on whether a brand is showing up at all. That raises a question most marketers haven’t thought through: what sample size do you actually need for the other jobs? Weekly tracking, competitor benchmarking, quarterly reporting, and budget decisions all require different levels of precision, and the prompt count is what determines whether the data can support the decision. 

What 10 prompts can tell you 

At 10 prompts per topic, a brand that appears in 5 out of 10 responses has a visibility reading of 50%. That sounds definitive. But the 95% confidence band around that reading stretches from roughly 24% to 76%. The true underlying number could be anywhere in that range. 

What does that mean in practice? If the brand shows up in 5 out of 10 prompts, you know it’s present in AI answers. That’s useful. You’ve answered the question “are we in the conversation at all?” and the answer is yes. That’s exactly the job Smith’s guidance frames 10 prompts for, and the math supports it. 

What 10 prompts can’t tell you is how much of the conversation you own. A 50% reading could mean the brand’s real visibility is 25% or 70%. A reading of 20% has a confidence band running from about 6% to 51%. The brand could be barely present or holding majority share; the sample can’t separate the two. 

For exploration, that’s fine. For anything that needs to distinguish one number from another: is the brand at 35% or 55%? did visibility move this week? It isn’t. 

What changes as the prompt count grows 

The math gets more useful as the sample gets larger. Here’s the intuition without the formulas: 

At 50 prompts per topic, the confidence band around a 50% reading narrows to roughly 36% to 64%. Still wide, but narrow enough that a brand sitting at 20% and a brand sitting at 50% are clearly separated. You can start to compare brands against each other and get a rough competitive picture. 

At 100 prompts, the band narrows further, roughly 40% to 60% around a 50% reading. Week-over-week changes need to be large to clear the band, but a 15-point swing starts to look like a real move rather than noise. 

At 200 or more prompts, the band is tight enough that smaller movements become distinguishable. This is the range where weekly tracking, competitive benchmarking, and trend analysis start to become defensible provided each prompt is also run multiple times to account for the variability AI engines introduce on every pass. 

The exact thresholds depend on the category, the baseline visibility, and how small a change matters to the decision. But the general shape is consistent: more prompts, narrower bands, more decisions the data can support. 

The question most marketers skip 

The sample size question usually comes up — if it comes up at all — during platform evaluation. How many prompts does this tier include? But the more important question is the one that follows: what decisions does that prompt count actually support? 

A platform that provides 50 prompts per topic with daily tracking gives the marketer a dashboard full of numbers that update every day. Whether those daily updates represent real movement or band fluctuation depends on the prompt count. At 50 prompts, a brand moving from 40% to 30% could be a genuine decline or it could be the confidence band doing what bands of that width do. 

The fix isn’t necessarily more prompts. It’s knowing which decisions your current prompt count supports and which ones it doesn’t. Exploration and presence checks need fewer prompts. Competitive benchmarking needs more. Weekly tracking needs more still. Budget decisions need the most. 

If the platform doesn’t surface confidence intervals alongside its metrics, the marketer has no way to make that judgment from the dashboard. The numbers look equally precise at every sample size, even when they aren’t. 

Matching the prompt count to the decision 

The practical framework is straightforward: 

“Are we showing up at all?” 10 to 20 prompts per topic can answer this. If the brand appears consistently across a small sample, it’s in the conversation. This is the exploratory tier Smith describes. 

“How do we compare to competitors?” This needs enough prompts that the confidence bands around two brands’ readings don’t overlap. At low sample sizes, two brands reading 35% and 50% may not be meaningfully different. At higher sample sizes, that gap becomes clear. Roughly 50 to 100 prompts per topic starts to support this, depending on how close the brands are. 

“Is our visibility changing week over week?” This is the most demanding use case because it requires the band to be narrow enough that a real week-over-week change is distinguishable from noise. It also requires multiple runs per prompt per measurement window, since AI engines produce different answers on every pass. This is where prompt counts in the hundreds, combined with multi-run measurement, become necessary. 

“Should we reallocate budget based on this data?” A budget decision deserves the tightest measurement. Confidence intervals on every metric, a stable baseline before the “before” reading, and enough prompts and runs that the reported change is clearly outside the noise floor. If the data can’t survive the question “how do we know this is real?” it isn’t ready for a budget conversation.