Keyword research produced a list of query strings, each with a monthly volume. Prompt research produces a different asset: an inventory of the questions people ask AI products, written the way people actually phrase them, tagged by intent and persona. That inventory is the working file for everything downstream. It decides what pages you build across this Part, and it becomes the prompt panel you sample in Measuring AI Search Visibility.
What people actually ask
The best public data on real usage is a NBER working paper from OpenAI and Harvard economist David Deming's team, built on a representative sample of ChatGPT conversations.How People Use ChatGPT (Chatterji, Cunningham, Deming, Hitzig, Ong, Shan, Wadman; NBER Working Paper 34255), the first large-scale study of real ChatGPT conversation data. Nearly 80% of usage falls into three categories: practical guidance, seeking information, and writing. Non-work messages grew from 53% of all usage to more than 70%. The paper also classifies messages by what the user wants back, and its Asking category, seeking information to inform a decision, is where AI search visibility gets decided.
Notice what is missing from that picture. Nobody types "best crm small business" into ChatGPT. People ask in sentences, with their situation attached: "We're a six-person agency running everything on spreadsheets. What CRM should we move to, and what will the migration cost?" The situation changes the answer, and capturing situations is the part keyword research never had to do.
From keyword list to prompt inventory
Your existing keyword data still matters as a seed, because every head term implies a family of prompts. The conversion is mechanical:
- Take a commercial keyword and write the question forms a person would say out loud: "best X for Y", "X or Y, which should I pick", "is X worth it at this price".
- Attach real constraints: team size, budget, existing stack, location, use case. Constraints come from how customers talk, and customers rarely talk the way your positioning doc does.
- Mine the places customers already phrase the problem in full sentences: sales-call transcripts, support tickets, community threads in your category, reviews of your competitors. Copy their sentences into the inventory verbatim.
- Tag each prompt with an intent from the taxonomy below, a persona where personas matter, and a buying stage.
- Refresh quarterly, and after any pricing, product, or regulatory change in your category.
A workable first inventory is 50 to 150 prompts. Bigger is not automatically better: every prompt you add is one you will later pay to run repeatedly, across models, with search on and off.
The intent taxonomy
Five intents cover the prompts a brand should care about. Informational: "how does term life insurance work". Comparison: "Webflow vs WordPress for a marketing site". Recommendation: "what project management tool should a 10-person construction firm use". Validation: "is Freshbooks legit, any problems I should know about". Transactional: "find me the cheapest direct flight to Denver next Friday", the intent that agents increasingly execute rather than describe.
Recommendation prompts are the new money keywords. A ranked results page gave you ten visible slots and clicks distributed down the page; an AI answer names a shortlist, and you are on it or you are absent. Validation prompts matter more than their volume suggests, because they run late in a deal, when a specific buyer checks a specific name.
Reverse-engineering fan-out
When an AI product searches, it does not run your prompt as one query. It generates several related queries, a step Google calls query fan-out, and retrieves results for each. Google's own example: "how to fix a lawn that's full of weeds" fans out into "best herbicides for lawns", "remove weeds without chemicals", and "how to prevent weeds in lawn".Google documents both retrieval-augmented generation and query fan-out, with this example, in its guide to AI features on Google Search. Retrieval and Grounding: How an AI Answer Gets Assembled walks through the full pipeline.
You cannot read the actual fan-out queries from most products, so predict them. For each priority prompt, write the sub-questions a careful researcher would ask next: cost, alternatives, prerequisites, failure modes, "for whom does this not work". Those predicted queries go into the inventory alongside the prompt that spawned them, and your pages need sections that answer them; Content Architecture: Building for the Chunk covers how.
Do not spin each predicted query into its own page. Google states that generating pages for every query variation, primarily to influence rankings or AI responses, violates its scaled content abuse policy.The warning against creating separate content for every possible search variation, including fan-out queries, is in the same Google AI features guide. Sections within a strong page cover fan-out; page-per-variation is spam with extra steps.
Persona variants: find out whether your category is taste-driven
The same recommendation prompt asked as a college student and as a retiree can produce different brand lists, because the persona text changes both what the model retrieves and which prior associations it draws on. How much it changes depends on the category. In OppAlerts' persona analysis across 100 industries, taste-driven categories like apparel rewrote most of the brand list when the persona changed, while utility categories like insurance moved far less.The persona sensitivity analysis is part of the AI Search Visibility research; method and per-industry scores are in The Research Behind This Guide. Exact sensitivity scores for apparel and insurance: [number: verify].
Test your own category in an hour. Pick five recommendation prompts, run each plain and then with three personas drawn from your actual customer segments, and compare the brand lists. If the lists barely change, skip persona depth and spend the effort elsewhere. If they change heavily, persona variants become first-class rows in the inventory, each one tracked on its own, and the spread between personas becomes a finding in itself: it tells you which segments you are invisible to.