Ethan Smith runs Graphite, the agency behind Webflow’s AEO program. His enterprise clients track 1,000 to 5,000 prompts. Asked how many Graphite tracks for itself: “Do we need to track 5,000 prompts for Graphite? Probably not. We probably only need to be looking at 10 to 15” (Smith on the Niklas Buschner podcast).
Sit with that gap. The practitioner who sells the 5,000-prompt program runs his own company on ten. Trackers price by prompt count, and the recommended prompt count keeps rising, and these two facts are related.
How many prompts does a small brand need?
Ten to fifteen. Smith’s words: for a small or new company, 10 to 15 prompts “is totally fine and a good way to get started” (Buschner interview). The 1,000-to-5,000 band exists for enterprises with hundreds of products and personas. He says the ideal there would be tens of thousands, same logic as SEO keyword tracking, and that it is “too expensive to actually do that.” The wish is a keyword-rank reflex imported from a channel where lookups cost nothing.
For a company with one product and three buyer types, a 5,000-prompt dashboard tells you, with precision, which 4,970 prompts nobody asks. The invoice does not shrink to match.
There is a structure for scaling when you earn it: take your top search topics, convert them to question form, and build a feature-by-persona matrix at around 10 prompts per page (Buschner interview). That matrix is what 1,000 prompts looks like when it is deserved. A list of 1,000 paraphrases is what it looks like when it is sold.
Why does sampling depth beat prompt breadth?
Because AI answers are a probability distribution, and one run per prompt samples that distribution once. Graphite’s own study found you need to ask the same prompt 7 to 10 times to get a decent view of the distribution for most commercial prompts (Buschner interview). Wider distributions need more. Smith is blunt about what the variance is not: “I have not seen a significant difference in answers based on the person.” The wobble is the model, not your cookies.
The budget math follows. 10 prompts at 8 runs each is 80 samples that give you a mention rate with error bars. 80 prompts at 1 run each is 80 coin flips filed under 80 different labels. Same spend. One of them is measurement. We wrote up the single-run problem, with the study that quantifies it, in there is no AI rank.
Where do the ten prompts come from?
From evidence of what buyers ask, and Smith’s recipe is concrete (Buschner interview). The head: your own search data, or a competitor’s top paid keywords exported to CSV and converted to question form with an LLM. Paid keywords are the ones a rival spends money on, which is a stronger demand signal than anything a brainstorm produces. The tail: conversation data you already own. Sales calls. Support tickets. Reddit threads in your category, G2 reviews, YouTube comments.
Notice what is absent from the recipe: volume. Question research beats question quantity because a tracked prompt nobody asks is a decoration, and a real buyer question you missed is a blind spot no prompt count fixes.
Where does our question count sit?
Inside the band, on purpose. Our free scan asks 5 buyer questions at one run per engine, and we label the result a snapshot because that is what a single run is. The paid audit runs about 30 questions at 7 runs per question per engine. Thirty is Smith’s small-brand band doubled, sampled at the depth his own study prescribes. We could quadruple the question count and the report would get longer without getting more true.
The honest limit runs the other way too. Five questions at one run is an on-ramp, a first look at whether you exist in the answers. It is not a mention rate. We say so on the report.
The practical step costs an afternoon. Pull your last ten sales calls and write down the first question each buyer asked before they found you. Convert your top paid keywords to questions and keep the ten that a real person would type into a chat box. That list beats any vendor’s 5,000. Then run the first five through our free scan and see which answers you are in.