NomaNoma
Chapter 3 of 6

Choose prompts that measure something

Five kinds of prompt and what each one tells you, how to read them against each other, and how to write a set that stays comparable.

The AEO guide · 4 min read

A prompt set is an instrument

Tracking prompts is not collecting interesting questions. It is building an instrument, and an instrument is only useful if it measures the same thing each time. A set chosen from the questions where you already look good does not measure your visibility. It measures your optimism.

Cover each claim several ways

Small changes in wording can change an answer a great deal. A single prompt is an anecdote. Several prompts testing the same claim from different angles, read together, are a measurement.

So build the set claim by claim. For each claim you chose in chapter 2, write a handful of prompts that a real buyer might use, varying the phrasing, the situation and the level of detail.

Five kinds of prompt, five jobs

KindWhat it asksWhat it tells youExample
CategoryWhich tools fit a needWhether you make the shortlist"What scheduling software suits a small physio clinic?"
ComparisonWhich of two options fits a situationWho wins head to head, and why"Is Brand A or Brand B better for a two-person clinic?"
ProblemHow to fix a symptom, no product namedWhether the engine connects the problem to you"How do clinics cut down on missed appointments?"
DirectWhether you do a specific thingWhether the engine understands you"Does Brand A send text reminders?"
Fact checkA plain fact about youWhether what it says is true"How much does Brand A cost per practitioner?"

Category and comparison prompts carry most of the signal, because they are where one brand can be swapped for another. Direct and fact-check prompts are diagnostics: they explain the result rather than being the result.

Read direct prompts against category prompts

The most useful thing a prompt set can do is tell you why you are missing, not only that you are. Put the direct prompts next to the category prompts for the same claim:

What you seeWhat it usually meansWhere the fix is
Named in category prompts, described correctlyThe claim is holdingKeep it maintained
Described correctly when asked directly, missing from category promptsThe engine knows you but does not trust you enough to recommendThird-party sources that back the claim
Described vaguely or wrongly when asked directlyThe engine does not understand this part of youYour own pages first

Write prompts a buyer would type

  • Use the buyer's words, not your positioning copy. Nobody asks for a "unified engagement platform".
  • Include the constraint that makes the question real: size, budget, industry, the tool already in place.
  • Spell out acronyms and ambiguous terms. A tracked prompt cannot answer a clarifying question.
  • Run each new prompt once by hand and read the answer before adding it to the set.
  • Check the answer has the shape the kind needs. A category prompt that does not produce a list of options is not measuring a shortlist.

How many, and how often to change them

You need enough prompts per claim that one answer changing does not swing the total. Keep branded prompts apart from the rest: you will be named in nearly all of them, and blended in they flatter every number.

Change the set as rarely as you can, and in dated batches when you do. Every change breaks comparison with the history before it. Never remove a prompt because its answer is unflattering; that is exactly the prompt that is working.

Where to run it

You can start by hand: a spreadsheet, a fixed list, the same questions asked of each engine on a schedule. It is slow, and it teaches you a lot about how the engines answer. Scripts give you control, at the cost of maintaining them as models, interfaces and rate limits change, and of making sure each run starts from a clean session. A tool takes that upkeep off your hands.

Key takeaways

  • Build the prompt set claim by claim, several prompts per claim, in a buyer's words.
  • Category and comparison prompts are the result; direct and fact-check prompts explain it.
  • Known but not recommended points to third-party sources; not understood points to your own pages.
  • Keep branded prompts separate, change the set rarely, and never drop a prompt for being unflattering.