For years, my customer segmentation work followed a depressingly reliable arc. I would spend the better part of a day building a gorgeous deck with twelve neatly-colored cohorts, present it to a conference room of nodding heads, and then watch absolutely nothing happen as a result. The problem, it took me embarrassingly long to admit, was not the analysis. The analysis was fine. The problem was that 'high-value users in the EMEA region who signed up in Q2' is a description, not a decision. Nobody in that room could actually DO anything with twelve overlapping slices, so they very reasonably did nothing with all of them and went back to their inboxes. I was producing segmentation as a performance of analytical rigor, not as a tool anyone could pick up and use.
The constraint that fixed how I think
The Segment a Customer List into Actionable Cohorts prompt has one rule baked into it that genuinely rewired my approach: every single cohort must map to exactly one concrete action, and the whole plan is capped at four to six cohorts total. That constraint is annoying in the best possible way. It forced me to ask, for every slice I wanted to keep, 'what would our team actually do about these specific people on Monday morning?' If the honest answer was 'nothing in particular,' the cohort did not make the cut, no matter how interesting it looked on a chart. My twelve vanity slices collapsed into five real cohorts that each arrived with a job already attached.
What I actually fed it
I pasted a representative sample of our user table, listed the available fields I could actually filter on (signup_date, plan, last_active_at, weekly_logins, feature_x_used, support_tickets), and set the segmentation goal to the blunt and specific 'reduce churn in the next 60 days.' Gemini 2.5 Pro returned five cohorts as a clean Markdown table, each one carrying the exact filter logic written against my real column names, a single-line reason it mattered specifically for churn, the one action it warranted, and an honest sample-based estimate of how much of our base it represented. Then, below the table, it handed me the precise WHERE clause for each cohort so I could pull the real lists immediately instead of translating descriptions back into SQL myself.
- 'Slipping Loyalists' — on a paid plan, formerly active every week, now zero logins for 14 days — action: a personal check-in email from their named account owner, not a marketing blast.
- 'Stuck at the Door' — signed up more than 7 days ago and still never used feature_x — action: a one-time activation nudge with a 90-second setup video and nothing else.
- 'Silent Sufferers' — 2 or more support tickets in 30 days AND a declining login trend — action: proactive human outreach before they quietly rage-churn and tell their network why.
- The Overlap and Priority section told me, unprompted, that if a user matched both 'Slipping Loyalists' and 'Silent Sufferers,' the support signal wins and they belong on the second list first — solving an assignment problem I would have argued about for a week.
The honest caveat about estimates
The percentages it gives are sample estimates, and crucially it labels them as estimates rather than passing them off as exact — which I came to appreciate the very first time I presented one and someone immediately asked 'is that exactly 18 percent?' I could honestly answer 'that's a sample estimate; the SQL it gave us will return the real number,' and then I did exactly that. I run each provided WHERE clause against the full table to get the true counts before anyone builds a campaign on them. The prompt is for the thinking and the structure; the database is for the ground truth. It also flatly refused to invent a 'lifetime_value' cohort because I had not listed that field as available, and instead told me to add the field if I wanted that cut — that restraint, that refusal to hallucinate a column, is exactly what I want from a tool whose output I am going to point a real retention budget at.
If your segmentation work consistently ends in a polite nod and a closed browser tab, the issue is almost certainly that your cohorts have no jobs to do. Grab the Segment a Customer List into Actionable Cohorts prompt on Prompt Dock, give it your real available fields and a single sharp goal, and let it refuse to make a cohort that nobody will ever action. And when you want to measure whether the action actually worked sixty days later, the Turn a Metrics Dump into a Skimmable Executive Summary prompt turns the follow-up numbers into a one-screen update for the same leadership team that nodded at the original deck.