How many prompts should you track for AI visibility?

Behind 'how many prompts' sit different questions. Am I named at all today? That is a diagnosis, and a handful of prompts answers it in an hour. Is my visibility moving? That is a measurement, which needs a fixed set too large for sampling noise to move on its own. Do I know where I stand across everything buyers ask? That is coverage, and it needs a set that spans the buying journey for every product you sell. The arguments between 5, 40 and 1,000 are mostly people answering different questions.

Five buyer-phrased questions in a temporary chat, scored as you go, tell you whether you are on the shelf at all, which is what the 30-minute audit does. Ten queries across three assistants gives thirty answers, enough spread to sort you into invisible, misdescribed or outranked, which is the one-hour audit. The free AI visibility check sits in the same range: it runs up to 10 buyer-style prompts about your brand through Claude with live web search, with no login, and shows the competitors ahead of you and your citation gap; paid plans also cover ChatGPT, Perplexity, Gemini, Copilot, Google AI Overviews and Google AI Mode.

None of these can tell you next month whether anything changed: ten prompts run once are a snapshot, and the same ten run again a month later are two snapshots with no way to say how much of the difference is noise.

The noise argument sets the floor. A single AI answer is one sample from a model that varies its output, so a weekly score built on two prompts is mostly that variation, and the trend line means little. Averaged over 25 prompts, the random component shrinks enough for a move of a few points to be worth a look, and for a move that holds across two or three consecutive weeks to be worth acting on. That is why 25 is Discoverable's core set on every plan: the free plan follows 25 prompts weekly on Claude and checks them across every surface once a month, and paid plans check the core 25 weekly on every surface.

Weekly, not daily, is deliberate. AI answers do not meaningfully change day to day; models and indexes shift over weeks, and a weekly reading keeps single-answer randomness from being mistaken for real movement. Four weekly readings give you a direction, and eight give you a trend you can put in front of a board. Two rules: freeze the set, because every prompt you swap breaks the line, and keep branded prompts under ten per cent of it, because asking about your own name measures description accuracy rather than recommendation. What is prompt coverage? sets out how to build and validate the set.

200 prompts is the size at which one product's buying journey is covered rather than sampled: research questions, category questions, a comparison against each serious rival, use-case questions, pricing and fit, and a few branded checks, each in the several phrasings buyers actually use. On Discoverable's Growth plan the core 25 are checked weekly on every surface and the remaining 175 monthly, which works out at about 2,000 AI answers analysed a month. That arithmetic is what keeps a weekly trend affordable.

1,000 prompts per brand is for companies with more than one thing to measure. A business with four product lines needs a prompt set per line, because a blended score hides the line that is invisible; Discoverable's Scale plan tracks up to 5 product lines per brand, each with its own prompts, competitors and score (product-line tracking). Markets multiply the set again, because answers differ by country and language. Engines expand each prompt into sub-questions, so the fan-out behind 50 head prompts can run to several hundred. At that size the core 25 still run weekly and the tail runs monthly, about 7,600 answers a month per brand.

As of 31 August 2026, 1,000 tracked prompts per brand was the largest self-serve prompt panel published by any GEO tool on our pricing comparison; the closest published self-serve volume was 500 prompts at $400 a month, checked daily. Prompt research is how you fill a set that size with questions buyers actually ask.

Guides on this site quote different prompt counts because they answer different questions. Read each against its purpose.

5 prompts: the 30-minute audit's first step, a checklist diagnosis scored 2, 1 or 0 per answer.

10 prompts: the one-hour audit, the Perplexity citation guide and the free check, a first diagnosis with enough spread to name the pattern.

10 to 20 prompts: the startup guide's first tracked set, measured from a zero baseline.

20 prompts: the manual AI visibility audit in the glossary, about the most a spreadsheet survives when every engine is checked by hand.

20 to 40 prompts: the share-of-voice guide's workable set for weekly manual measurement, and 20 to 50 the share-of-model entry's prompt panel, both limited by how much checking a person keeps up with.

40 prompts: the prompt-coverage entry's example of coverage across a buying journey, against two as an anecdote.

25, 200 and 1,000 prompts: the product's tiers, where a tool rather than a person does the checking, so the ceiling is set by coverage rather than by effort.

None of these is wrong. Below about 20, you are diagnosing. From 25 checked weekly, you are measuring. From 200, you are covering a journey, and from 1,000 you are covering a company. Pick the number for the question you are actually asking, and if the question is whether your visibility is moving, the answer is 25 fixed prompts, every week, on ChatGPT and every other engine your buyers use.