Most agencies start GEA with a keyword export because that is what exists. That is fine as raw material, but a keyword list is a list of phrasings, and you need a list of situations. The translation work is a clustering exercise: group phrasings that share a buying circumstance, then name the circumstance rather than the phrase.
A useful cluster has four parts: who the person is, what triggered the search now, the constraints they carry, and the outcome they want. "Cheap CRM" is not a cluster. "A founder-led sales team of ten, triggered by a renewal quote that doubled, constrained by no technical admin, wanting a working pipeline in two weeks" is a cluster, and it is immediately obvious what evidence would win it.
Aim for somewhere between five and twelve clusters per product line. Fewer than five and you are describing a category, not a set of situations. More than twelve and you cannot produce distinct evidence for each, so the clusters collapse into each other and your hints start repeating.
Prioritise clusters on three axes: how often the situation occurs, how well your product actually fits it, and how weakly incumbents cover it. The third axis is the one teams skip, and it is where most early wins come from, a cluster where the obvious competitor's material is silent on a key constraint is a cluster you can take with a single well-evidenced page.
Write each cluster down in one sentence, in the buyer's own words, and store it with the prompts you expect it to produce. Those prompts become your monitoring set: they are what you will test against engines weekly to see whether you are named, hedged, or absent.
Finally, keep clusters honest. A cluster your product genuinely fits badly will not convert even if you win the citation, and repeated recommendations that disappoint reduce your standing over time. Deliberately de-prioritising a poor-fit cluster is a legitimate, and often profitable, planning decision.