Harvest: where real questions live

  • Your own funnel: sales-call notes, support tickets, demo questions, lost-deal reasons.
  • Public question surfaces: Reddit and niche forums, Quora, "People Also Ask", competitor FAQ pages, review-site complaint sections.
  • The engines themselves: ask ChatGPT/Perplexity "what should I consider when choosing [category]" and harvest the sub-questions they surface — that is the fan-out, partially visible.

    • Cluster: intents, not strings

    Group phrasings by the decision they serve, not the words they share. "HEPA vs electrostatic", "is ionization safe" and "which filter technology lasts longest" are one cluster (technology choice) even though they share no keywords. Embedding-based clustering makes this mechanical at scale; for a first pass, a spreadsheet and an afternoon work fine. Name each cluster by its decision: technology choice, compliance, total cost, vendor trust.

    Map and score

    1. For each cluster: which page — yours or third-party — answers it with you present?
    2. Score coverage: clusters where engines name you / total clusters.
    3. Rank gaps by buyer value × winnability. A high-value cluster where engines currently cite thin sources is your first target.

    From map to roadmap

    The map converts GEO from "publish more" into sequencing: gap clusters become pages, weak clusters get answer-block reformats, owned clusters get defended with freshness and corroboration. Re-run quarterly — buyer language drifts, new clusters appear (any new regulation, standard or competitor creates one), and the map keeps the program honest.

    Frequently asked questions

    How many clusters should a category map have?
    B2B categories typically resolve to 40–80 clusters; consumer categories 25–50. Below 20 you are probably grouping too coarsely; above 150 you are splitting intents that engines treat as one.
    How is this different from keyword research?
    Keyword research inventories strings people type into a search box. Coverage mapping inventories decisions buyers need to make, in their own language, including questions they only ask an AI because a search box never handled them well.