Back to Insights
AEO

How Many Prompts Should an AEO Agency Track? Our Number Is 95

Author

Tanuj Sarva

Published

November 10, 2026

Read Time

6 min read

How Many Prompts Should an AEO Agency Track? Our Number Is 95

Key takeaways

  • We track 95 prompts, chosen because each one is a question a real buyer would plausibly ask and each maps to a specific page that should answer it.
  • The number is less important than its derivation. A set of 95 chosen from buyer behaviour beats a set of 500 generated from keyword permutations.
  • A very large prompt set is usually a tooling artefact. Generating a thousand variations is cheap, monitoring them is automated, and almost none of them corresponds to anything a person would type.
  • A very small set is usually underscoped. Under about twenty prompts you cannot distinguish a real movement from the natural variance in assistant answers.
  • Assistant answers vary between runs for the same prompt, so a single check is not a measurement. Any honest tracking regime samples repeatedly and reports a rate rather than a state.
  • Ask an agency how its prompt set was built and what each prompt maps to. An answer about volume rather than derivation tells you the set is generated.

Nobody asks this question and it is the best one available. Prompt set size and derivation reveal more about an AEO practice than any case study, because they cannot be answered impressively without actually having a method.

We track 95. That number is published in our own content plan, along with which article answers each prompt and which are still unwritten.

What follows is why 95 rather than 500 or 20, how a set should be derived, and the measurement problem that makes the whole exercise harder than it looks.

How many prompts should an AEO agency track?

Short answer: Enough that each one corresponds to a question a real buyer would ask, and no more. For most businesses that is somewhere between twenty and a couple of hundred, and the derivation matters far more than where in that range you land.

We use 95 for this site. That is not a recommended number for anyone else; it is the number of distinct buyer questions we could identify and map to a page.

What different prompt set sizes usually indicate
Set sizeUsually meansThe problem
Under 20Underscoped or just startingCannot separate movement from variance
20 to 200Derived from buyer questionsRequires real work to build
200 to 1,000Partly generatedMost prompts nobody would type
Over 1,000Tooling artefactMonitoring noise, reported as coverage

The middle row is the only one that requires effort, which is why it is the least common. Generating permutations is instant and sounds larger.

How should a prompt set be built?

Short answer: From how buyers actually phrase things, which means sales calls, support tickets, the questions you answer on the phone, and your own Search Console query data. Keyword tools are a poor source because people prompt an assistant in sentences, not in keywords.

  1. Start with the questions you already answer out loud. Sales and support hear the real phrasing every day and almost nobody writes it down.
  2. Add the question-shaped queries from your own Search Console data. Those are real sentences real people typed.
  3. Add the comparison and shortlisting questions, because that is where a buying decision actually happens.
  4. Map each prompt to the page that should answer it. A prompt with no page is a content brief, which is the most useful output of the whole exercise.
  5. Discard anything you cannot imagine a person saying. This step removes most of a generated set and is the step that gets skipped.

Step four is where the value is. A prompt set is not primarily a measurement instrument, it is a map of what you have not written yet.

Our own plan works exactly this way. Each of the 95 prompts names the article that answers it, and the ones with no article are the backlog.

Why is bigger not better?

Short answer: Because monitoring cost scales with set size while useful information does not. A thousand generated prompts produce a thousand results, almost all of them about phrasings that nobody would ever use, and the noise from those swamps whatever signal the handful of real ones carry.

There is also a reporting failure that follows from size. A large set gets summarised into an aggregate visibility score, and an aggregate across mostly irrelevant prompts is a number with no decision attached.

Smaller sets can be read individually. If you track ninety-five prompts you can look at all ninety-five and notice that three moved, which is actionable.

The question that exposes a generated set: ask the agency to read you ten prompts at random from it. A derived set sounds like things people say.

A generated set sounds like keyword strings with question words bolted on.

Why is a single check not a measurement?

Short answer: Because assistant answers vary between runs of the very same prompt. Model updates, sampling behaviour and personalisation all change the output, so checking once tells you what happened on that one occasion rather than what typically happens when somebody asks.

This is the part of AEO measurement most often misrepresented. A screenshot of your brand in an answer is evidence that it can happen, not that it does.

The honest regime samples the same prompt repeatedly over time and reports a rate: named in some proportion of runs, not named in the rest. A rate can be tracked; a state cannot.

It also means small changes are usually not changes. A brand appearing in six of ten runs this month and five of ten last month has probably not moved.

Any tool or agency reporting a precise visibility percentage without describing its sampling is reporting a single draw as if it were a distribution.

What should a prompt report tell you?

Short answer: Which prompts you are named in and at what rate, which competitors are appearing alongside you, which prompts have no page behind them at all, and what has changed since the last report. That fourth column is the only one that implies an action.

The competitor column is more useful than the ranking column. Knowing who is being named instead of you tells you what the assistant currently considers authoritative.

The uncovered-prompt column is the content backlog, and it is where the report earns its fee. A prompt nobody has written a page for is a specific, prioritised brief.

What a report should not contain is a single composite score. It compresses away every decision you could have made from it.

Frequently asked questions

What is a prompt in this context?

A full question as a person would type or say it to an assistant, such as which agencies handle answer engine optimization for professional services. It is a sentence rather than a keyword, which is why keyword tools are a poor source for building a set.

Should the prompt set include competitor names?

Yes, because buyers prompt with names during a shortlist. Comparison prompts are where decisions happen, and they are the prompts most agencies leave out because the answers are uncomfortable.

How often should prompts be checked?

Often enough to establish a rate rather than a state, which in practice means several times per prompt per reporting period. Monthly single checks produce numbers that move for reasons unrelated to anything you did.

Do I need a tool for this?

A tool helps at scale and is not required to start. A spreadsheet, the assistants themselves and a disciplined sampling schedule will give you a defensible baseline for a set of ninety-odd prompts.

Does prompt tracking tell me about revenue?

Not directly, and nobody can honestly claim otherwise, because a citation inside an answer sends no referrer. It tells you whether you are in the consideration set, which is upstream of revenue and worth knowing on its own terms.

Sources and further reading

Ask us to read you ten at random

Our 95 prompts are published alongside which article answers each one and which are still unwritten. Ask for the list, and ask what is still missing from it.