How many prompts do you need to track AI visibility?

Nobody has published a sample-size study for AI visibility tracking. The evidence base behind this library holds 140 checked claims, and none of them sets a prompt count. So the honest answer is a method, not a number.

A prompt set is a sample. It stands in for the questions real buyers ask. Its size should follow from four things:

• Coverage. How many distinct buyer questions decide whether you get shortlisted.
• Mix. How many prompts name your brand, and how many don't.
• Repeats. How many times you run each prompt, because answers vary between runs.
• Cost. What each answer costs across every engine you check.

The working rule: start small enough that you can read every answer yourself. Grow the set only where you see a gap in coverage or an answer that changes from run to run. Ten well-chosen prompts that map to real decisions tell you more than two hundred generated variants of the same question.

Coverage of buyer questions

Each prompt should map to a question a buyer asks at a specific stage: discovery ("best payroll software for a 20-person company"), comparison, pricing, and fit for a use case. If two prompts would get the same answer, you only need one. Adding near-duplicates inflates the count without adding information.

Branded and unbranded mix

The two types answer different questions. Unbranded prompts test whether an engine names you at all. Branded prompts test what it says once it does, including whether it gets your price or product names wrong. Most sets need both, and the balance depends on whether your problem is being absent or being described wrongly. The trade-offs are covered in branded vs unbranded prompts.

Repeat runs

Retrieval is per query and discretionary. As of October 2026, ChatGPT search decides per prompt whether to search, and Google says AI Overviews often do not trigger at all. So the same prompt can return a grounded answer on one run and a parametric one on the next. Running each prompt more than once tells you whether a mention is stable or a coin flip.

Cost

Total answers equal prompts times engines times runs. Say you track 20 prompts on 5 engines with 3 runs each: that is 300 answers per check. Some answers carry a search fee on top of tokens. Anthropic's web search tool documentation lists $10 per 1,000 searches plus token costs. Decide your check cadence at the same time, because weekly runs multiply every number above.

Prompts are a proxy for an input nobody outside the engine can see. OpenAI states that ChatGPT search "typically rewrites your query into one or more targeted queries" before sending them to search providers. The Gemini API documentation says grounding automatically generates one or more search queries. Google says AI Overviews and AI Mode "may use different models and techniques." So your prompt is not the query that retrieves pages, and more prompts do not close that gap.

Engine-reported data shows the limit of any sample. In Otterly's 2,000 AI blogs experiment, Microsoft's AI Performance report credited one test site with 42,817 citations, against 3,934 logged by Otterly's prompt-based tracker. The time windows may not match, so it is not an exact ratio. It does show that a prompt sample sees a fraction of what happens, and why Bing's counts differ from tracker counts.

The 2026 critical survey of 45 GEO studies found no technique with a stable, longitudinal, cross-platform effect on discoverability. Nothing in that literature measures how many prompts make a reliable visibility number. Treat any score as your tool's model of visibility, not a measurement of it.

• Small or new brands. If no engine names you on unbranded prompts, a bigger set mostly confirms a zero. Fix access first: check that retrieval crawlers can reach your pages and that key facts are in the HTML before JavaScript runs.
• Many products or markets. Each product line, country or language is its own set of buyer questions. Coverage, not a global total, sets the count.
• Changing the set mid-trend. Adding prompts changes what you measure. A sound score normalizes by answers returned, so the same error rate scores the same at 20 answers or 400. If your score drops when you add prompts, read whether adding prompts should change your score.
• Logged-in testing. OpenAI notes that ChatGPT may use saved memories when rewriting a search query. Run checks logged out with memory off, or the results won't repeat.

1. Start with three to five questions per buyer stage. This is a working rule, not a research finding. Write them in the buyer's words. The free checker on this site runs up to three buyer questions as a one-time snapshot, which is a quick way to see a first baseline.
2. Run each prompt more than once. Note which answers stay the same and which flip between mentioning you and not.
3. Add prompts where coverage is thin. If a stage or use case has no prompt, write one. If two prompts always get the same answer, drop one.
4. Add repeats where answers are unstable. Instability is a reason to run more, not a reason to write more variants.
5. Check your set against real queries. Bing Webmaster Tools exposes grounding queries, the phrases the AI actually used. Compare them with your prompt set and add prompts for any pattern you missed.
6. Freeze the set before trending. Change it on a schedule and note the date, so a jump in the line has a known cause.

For the full build process, see how to build a prompt set.

Is there an industry standard number of prompts for AI visibility? No. No published study sets one, and the evidence base behind this library contains no sample-size finding for prompt tracking. Numbers you see in tool plans are product limits, not research. Size your set by counting the distinct buyer questions that matter to you, then adding repeat runs where answers vary.

Should I add more prompts or run the same prompts more often? It depends on the problem. If a buyer stage or use case has no prompt, add one. If a prompt gives a different answer each run, repeat it instead of writing variants. Retrieval fires per query at the engine's discretion, so repeats measure stability, while new prompts measure coverage.

Do more prompts make my AI visibility score more accurate? Not by themselves. Prompts are a proxy, because engines like ChatGPT rewrite your prompt into their own search queries. More prompts give a bigger sample of that proxy, not a view of the real queries. Better coverage of real buyer questions and checking against Bing's grounding queries improve the sample more than raw volume.

Should every prompt run on every AI engine? Usually yes, and report each engine separately. Engines draw on different indexes and behave differently. In Otterly's experiment, Microsoft Copilot supplied most of one site's tracked citations while ChatGPT, Gemini and Google's AI surfaces almost never cited it. A blended number would have hidden that split.