Prompts vary. Intent and persona don't.

Questions this guide answers

• Do you need real prompts to measure AEO?
• Why track intent and persona instead of exact prompts in AI search?
• How do you choose prompts for AI visibility tracking?
• Does SolCrys use real data to build prompt sets?
• What is the right unit of AEO measurement?

Direct answer

Track intent and persona, not exact prompts. The wording of a buyer's prompt changes every time it's typed, but the intent behind it (the decision they're making) and the persona asking (their role, company size, and constraints) are stable, and they change only when the market does. Answer engines work from that need: Google says AI Overviews and AI Mode may use a "query fan-out" technique that issues multiple related searches across subtopics and data sources (Google Search Central), so the answer draws on the need behind the question, not only on its exact words.

A prompt set that covers the intents and personas carrying demand in your category, each phrased the way that buyer asks, measures what real buyers see. Because intents and personas hold still while wording churns, the same set can be rerun over time, and a change in your numbers reflects a change in the answers rather than a change in the questions.

Why I'm writing this

Teams evaluating AEO tools increasingly ask whether a vendor's prompts are "real." It's a fair question with a hidden assumption: that the literal prompt a buyer typed is the thing to measure. I think that assumption is wrong, and that acting on it leads teams to track wording instead of demand.

Our Golden Prompt Set is built from real demand signals, and every prompt in it carries the persona it stands for. This essay explains why we built it around intent and persona, and why that is the stable foundation measurement needs.

One intent, many prompts

Here is one buyer need written 5 ways. A real buyer could type any of them, and none is more "the real prompt" than the others. What they share is the decision being made (choose a CRM) and who is making it (a VP of sales at a 50-person B2B SaaS company).

Why intent and persona are the stable unit

Wording is the noisiest part of a prompt. Intent and persona are the parts that carry the decision.

• Wording churns; intent doesn't. Prompts differ in length, word order, named brands, and context, and the same buyer rewords as the conversation goes. When its memory settings are on, ChatGPT also draws on saved memories and past conversations (OpenAI Help Center), so the literal string was never the whole input. The decision behind it changes only when the buyer's need changes.
• Persona changes what a good answer is. A 10-rep startup and a 500-rep enterprise asking the same question need different shortlists, so the persona belongs in the prompt. The exact wording doesn't need to match anyone's.
• Engines expand the question. Google's AI surfaces may fan a question out into related searches across subtopics before answering, and answers vary from run to run. A literal prompt is one sample of a need the engine is already generalizing.
• Intents move with the market, and slowly. New intents appear when a category changes: a new entrant, a pricing shift, a regulation. That's when a prompt set should change too. Wording changes every time someone types.

Stages, not steps

The buying journey isn't a straight line, and an intent-based set doesn't assume one. Gartner's research describes B2B buying as non-linear: buyers often revisit buying tasks rather than finishing them in order (Gartner). In an AI assistant those loops get shorter. One conversation can move from a category question to a comparison to a risk check in a few turns, as the follow-up in the table above illustrates.

So a buying stage belongs to each question, not to a position in a sequence. A buyer can enter at a comparison, step back to a category question when a new option appears, and ask about risk first or last.

• Cover every stage for each persona. Buyers enter anywhere, so a set that tracks only discovery questions misses the comparisons and risk checks where shortlists are decided.
• Don't read the stages as a funnel. Visibility on comparison questions doesn't wait on discovery; measure each stage on its own.
• Make the decisive turn the prompt. If the shortlist is decided on a follow-up ("which of those handles long enterprise deals?"), write that intent's one prompt as a standalone question that carries the context the buyer gave earlier.

What this means for measurement

Measure coverage of intent and persona, and measure it the same way every time.

• One prompt per intent and persona that matters, not 10 paraphrases. Paraphrases of the same need remeasure the same decision; a missing intent or persona is a blind spot.
• Write each prompt the way that persona asks. Include the role, size, and constraint a real buyer would give, so the engine answers the question that persona has.
• Freeze the set, then rerun it. Because intents hold still, a frozen set stays representative, and a change in your numbers reflects a change in the answers. Judge any move across several runs and engines (how many runs a visibility number needs).
• Read results by intent and persona. "Named in 3 of 4 discovery answers for mid-market sales leaders, absent for startups" tells you what to fix; an average across unrelated strings doesn't.

How the Golden Prompt Set is built around intent and persona

Each SolCrys workspace tracks a Golden Prompt Set built from real demand signals for its category: search demand, the questions people ask around it, and community discussion where available. Those signals show which intents carry demand and how the buyers behind them phrase their questions. On Google's AI surfaces, search demand is direct evidence: AI Overviews and AI Mode build on Google's Search index and core ranking systems (Google's May 2026 guide), so AEO for Google is an operating layer on top of SEO foundations.

The SolCrys Loop (Measure → Diagnose → Execute → Verify) runs on that set across ChatGPT, Google AI Overviews, Gemini, Perplexity, and Claude, depending on plan, and reruns the same frozen prompts after a fix ships, so you can see whether the answer moved for that intent and persona and whether the move held across runs.

• Intent, as prompt type. Every prompt carries a type that names the visibility question it tests: Discovery (does the engine name you for the category?), Feature Research (for a specific use case or constraint?), Comparison (how are you framed head-to-head?), and Brand Accuracy (are the facts right?). Each set is built to a fixed type mix, not improvised per customer.
• Persona, on every prompt. Every prompt carries the buyer role it's written for, such as a VP of marketing or a data center architect (or Generic when no single role applies), and persona is a filter in your Prompts table.
• Demand, as a band. Prompts carry a High, Medium, or Low demand band, with the keyword behind them in the band's popover, so you can prioritize the intents with the most buyer interest.
• Stable by design. The set is reviewed with you and changes when the category does. Prompts you keep retain their trend lines, and new ones start their own history.
• Your buyers' own words. Questions from your sales calls, support tickets, and demos slot in alongside the GPS.

Where real prompt data fits

Real prompts are evidence for intents and personas, and the most useful ones are already yours. A list of literal strings is a snapshot of how some buyers worded something once; mapped to intents and personas, it keeps its value as the wording moves on.

What would change this

If AI platforms start reporting prompt-level data to site owners, the way Search Console reports queries, intent discovery gets faster. The first step with that data will be the same: map each prompt to the intent and persona behind it, and check that your set covers them.

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. For the practical method of finding candidate prompts, see How to Find the AI Prompts Your Brand Shows Up For.

Sources

• Google Search Central: AI features and your website
• Google Search Central: generative AI optimization guide (May 2026)
• OpenAI Help Center: Memory FAQ
• Gartner: the traditional linear sales process no longer matches how B2B buyers buy (Nov. 13, 2019)

FAQ

Related guides