Tutorial

Cohort analysis, from zero to confident

This guide teaches cohort analysis itself, then every setting of the Cohort Matrix visual — with a small sample dataset you can download and follow along with.

1. What a cohort matrix tells you

A cohort is a group of entities that started in the same period — customers whose first purchase was in January, users who signed up in week 12, SKUs launched in Q2. The matrix puts one cohort per row and asks the same question across each column: how many of them were still active 1, 2, 3… periods after they started?

Three reading directions, three insights:

• ●Across a row — the survival curve of one generation: how fast does a January customer fade?
• ●Down a column — quality over time: are newer cohorts stickier at month 3 than older ones? This is where product and marketing changes show up.
• ●The diagonal — calendar effects: a bad month (outage, price change) cuts diagonally through all cohorts at once.

2. Quick start

1. Add the visual to a report page.
   Once it is on AppSource: Get more visuals → search “Cohort Matrix”.
2. Drop your entity (Customer, User ID, SKU…) into Entity.
3. Drop the activity date into Event date.
   A date column, or a text key like 2024-09. Don’t use a date hierarchy — the visual will detect it and tell you to use the plain column.
4. Optionally drop a value (Sales, Quantity) into Value.
   Even if you only analyze counts, adding a measure keeps Power BI from cross-joining dimension tables and hitting row limits.
5. Read the matrix. Rows are cohorts by first activity; the Size column counts new entities per cohort; each cell shows retention for that period offset.

3. The concepts that make or break the analysis

Classic vs unbounded retention

Classic counts an entity in month 3 only if it was active in exactly month 3. Unbounded ("in or after") counts it if it was active in month 3 or any later month — so a customer who skips a month but comes back still counts as retained. Classic shows activity rhythm; unbounded shows true survival. Unbounded is always ≥ classic, and it can never rise as you move right.

Incomplete periods

The bottom-right of every cohort matrix is a staircase of cells that haven’t happened yet — and one diagonal of cells that are in progress. Cohort Matrix leaves future cells blank (never a fake 0%) and hatches the in-progress diagonal with an asterisk. The averages row ignores both. If partial data bothers your audience, switch on Complete periods only.

Why the first cohort lies

Filter a report to “last 24 months” and the first month’s cohort inflates: everyone active that month looks “new” because their earlier history is filtered away. Cohort Matrix keeps the row but excludes it from color scaling by default (gray, italic, with an explaining tooltip) so it can’t distort the heatmap. Toggle: Exclude first cohort.

4. Settings reference

Setup

Display

Heatmap

5. Advanced: relaunch cohorts (cohort date override)

Normally the cohort is the entity’s first event. Sometimes you want a different anchor: retention of SKUs since their relaunch, customers since a re-onboarding, users since a campaign. Drop that anchor date into the Cohort date well:

• ●Cohorts group by the override date, not first activity.
• ●Events before the anchor are excluded — pre-relaunch sales don’t pollute the analysis.
• ●If an entity has several anchor dates, the latest wins.
• ●The caption appends “(cohort = your field)” so nobody misreads the chart.

Try it with the relaunch sample: three SKUs with launch dates; S1 relaunched in March 2026 shows offsets 0, 1 and 3 — and its pre-launch billing is correctly dropped.

6. Performance checklist

• Use month-grain date keys (a Year-Month column) instead of full timestamps when analyzing monthly cohorts — far fewer distinct rows.
• Always include a measure in Value, even for count analysis — it prevents Power BI from cross-joining dimensions into millions of rows.
• Above 300,000 rows the visual warns you rather than silently truncating. Pre-aggregate or switch to pre-computed mode.