Every prompt set is chosen. What has to be real is the demand.

Questions this guide answers

• Are real prompts better than synthetic prompts for AEO tracking?
• Does any AEO tool track real user prompts?
• How should an AEO vendor choose which prompts to track?
• How can I check whether a prompt set represents my buyers?
• Is AI prompt panel data worth paying for?

Direct answer

Every AEO visibility tracker runs a chosen set of prompts against the engines. That includes vendors with access to panel data on what people type into AI assistants: to track visibility, they still pick a subset and rerun it. So the real difference between methods is how the set is chosen, not whether its sentences were once typed by someone.

The test of a prompt set is whether it represents the intents and personas that carry demand for your buyers. That is where "real" belongs: in the demand evidence behind each intent, not in the wording of each sentence. And which way of choosing works better isn't settled by assertion. You can test it with the real phrasings you already have.

Why I'm writing this

Buyers evaluating AEO tools increasingly hear that only "real prompts" count and that everything else is synthetic. It's a strong pitch, because "real" sounds like evidence and "synthetic" sounds like guessing. Part of it is true: a prompt set with nothing behind it is guessing. The rest skips a step, and the step it skips is the one that decides whether your visibility numbers mean anything.

In Prompts vary. Intent and persona don't. I argued that buyers word the same need in endless ways while the intent and the persona asking stay stable. This essay follows that argument to the question every vendor gets asked: where do your prompts come from, and are they real?

Everyone selects

Suppose you had every prompt anyone in your market typed into an AI assistant last month. You still couldn't track all of them, and most would be one-offs worded for one person's situation. To measure visibility, you would pick a set, run it against the engines, and rerun it so the numbers can be compared over time.

The moment a sentence is picked and rerun, it changes role. It is no longer one buyer's moment; it is a probe that stands in for an intent. The engine answering it doesn't know where it came from, and on Google's AI surfaces it doesn't answer only the exact words: Google's guide to its generative AI features describes query fan-out as "a set of concurrent, related queries generated by the model" to address the user's query (Google Search Central). Reading answers straight from a panel selects too: you see what that panel's members asked. So "real versus synthetic" describes where a sentence started, not what the measurement does with it. Every method selects. The question is how well it selects.

The one criterion

A prompt set is good if it represents the intents and personas that carry demand for your buyers. Not for the internet in general, and not for whoever happens to be in a panel. Two checks test it, and a third keeps it measurable over time.

• Coverage. Every intent and persona that matters to your pipeline has a prompt, across every buying stage, including the low-volume decision questions where shortlists are made.
• Demand. Each intent is backed by evidence that buyers have it: search demand for the underlying question, how often it comes up in public discussion, or how often your own buyers raise it.
• Stability. The set is frozen and rerun the same way, so a change in the numbers reflects a change in the answers, not in the questions.

Two ways to meet it

There are 2 broad ways to choose a set, and both can meet the criterion or miss it. Sampling starts from prompts observed in a panel or a log. Designing starts from evidence of demand and writes 1 representative prompt per intent and persona.

Why “synthetic” is the wrong word

"Synthetic" frames the question as real versus fake. A prompt designed from demand evidence isn't fake any more than a survey sample is fake for being drawn on purpose. Survey researchers don't ask whether a sample is real. They ask whether it represents the population it's used to describe, and they put every sample, however it was drawn, to that test. Prompt sets deserve the same.

So I'd retire the word. The useful labels describe the method: sampled or designed, and grounded in demand or not.

Make it testable

Whether a designed set represents your buyers as well as real phrasings do is an empirical question, and you can run the test yourself with phrasings you already own. It's the test we're putting our own prompt sets through.

• Collect real phrasings. Long, question-form queries in Search Console (set aside automated probes), questions from sales calls and support tickets (how to find them), and answers to a "which AI did you use, and what did you ask?" field on your signup form (how to ask).
• Map each to an intent and persona in the vendor's prompt set. Any phrasing that maps to nothing is a coverage gap.
• Run both on the same engines, repeatedly. Compare how often the brand is named and recommended for the designed prompt and for the real phrasings of the same intent (how many runs a number needs).
• Decide what gap matters, then check it. Before you run anything, set the smallest difference in mention or recommendation rate you would act on. Run enough times that the uncertainty around the measured gap is narrower than that margin. If the gap, uncertainty included, stays inside the margin, the designed prompt represents that intent. If it doesn't, add phrasings for that intent.

What to ask any vendor

These questions apply to every method, sampled or designed, and a good vendor can answer all of them in writing.

• How is the set chosen, and what evidence of demand stands behind each intent?
• Which intents and personas does it cover, and which buying stages?
• Is it the same set every run, and how are changes dated?
• Can I add my own buyers' questions alongside it?
• How would you check the set against real phrasings of the same intents, and will you run that check with me?

How SolCrys fits

SolCrys designs each workspace's Golden Prompt Set from search demand, the questions people ask around it, and community discussion where available. Every prompt carries the intent it tests, as its prompt type, and the persona it's written for, and prompts carry a demand band so you can see the evidence behind them. You can add your own buyers' questions alongside the set.

The set is frozen and rerun on ChatGPT, plus Gemini, Google AI Overviews, Perplexity, and Claude on paid plans. That's the SolCrys Loop (Measure → Diagnose → Execute → Verify): a change in the numbers reflects a change in the answers, and after a fix ships, the same prompts show whether the answer moved and held across runs.

About the author

Eason Wang is Co-Founder & CPO of SolCrys. He holds a PhD in Machine Learning and has spent 18+ years building enterprise products, starting at Microsoft Research Asia. His current focus is agentic AI workflows and the infrastructure required to run them in production. Connect on LinkedIn.

Written October 2026. This essay builds on Prompts vary. Intent and persona don't.

Sources

• Google Search Central: Google's guide to optimizing for generative AI features on Google Search

FAQ

Related guides