The level you aggregate at decides the answer

Here is an idea that should give you pause: your blended retention rate can improve in the same period that retention falls in every customer segment you have.

That is not a data error. It is Simpson’s paradox—and in SaaS analytics, not a curiosity but a standing risk hiding inside every aggregate you report.

The number that went up while everything went down

Suppose you sell to two segments, SMB and enterprise. Compare last year to this year:

Both segments declined by a point, but the number on the board slide went up, from 81% to 83%.

Saying retention improved is not inaccurate—it did. The blended number genuinely rose, and if the shift toward enterprise is permanent and the segment rates hold, 83% will hold too.

What is wrong is the attribution. The gain did not come from your renewal team keeping customers better—every segment retained a point worse than last year. It came from acquiring more enterprise customers, who retain better. The aggregate says something true about the shape of your customer base and something misleading about how well you renew. Credit it to renewals and you will lean on a motion that is quietly slipping, then be blindsided when enterprise acquisition slows and the metric stalls.

The segments don’t even need to decline for the number to mislead. Had each simply held flat, the mix shift alone would have lifted the blend from 81% to about 84%—a larger “improvement” than the real case, and still entirely a product of who you acquired, not how well you renewed.

The culprit is the level of detail

Nothing in that table is false; the aggregate and the segments are all arithmetically correct. They disagree because they are computed at different levels of detail—different grains of aggregation—and the grain you pick determines the answer you get.

It is not only retention, and not only segments. The same trap lives in every metric and dimension you could roll up:

Customer versus customer-product. Take an account that moves from your Basic edition to your Pro edition with a 20% increase. At the customer-product grain, where each product edition is its own row, that is churn on Basic and a new sale on Pro. Basic’s gross retention drops and new-business ticks up. At the customer grain, the customer upsold. Both views are correct; they answer different questions.

Time. A monthly retention series and a quarterly one describe the same customers, but the quarterly view nets out anyone who churned and re-signed inside the quarter and smooths over a bad month that a monthly view would flag. Neither is wrong. The quarterly one is just answering a slower question.

What counts as a customer. A logo, a contract, a billing account, a parent organization? Upgrades, consolidations, and multi-entity deals all resolve differently depending on that definition, and it is a choice of grain most teams never make explicitly.

Nobody makes decisions on the raw data. They make them on the aggregate—the board metric, the KPI tile, the summary row—so whatever grain that summary was built at is the grain your strategy is quietly optimizing for. Choose it badly and the costs are concrete: you over-cut a segment that was fine, pour spend into “growth” that is really a mix shift, or assure a board that all is well while the base erodes underneath.

Finding the right level

The obvious fix—always go finer—is seductive, because more detail feels like more rigor. But two problems show up as you slice.

The cells multiply faster than you expect. Cross five segments by three products by six regions and you have ninety intersections—too many to track and certainly too many for a single slide. A metric nobody can hold in their head is a metric nobody acts on: the dashboard technically contains the answer, but it is buried across ninety cells.

Each cell shrinks, and small cells get noisy. Your customer base is finite. Split it ninety ways and each intersection holds a handful of accounts—and at that size, a single large customer renewing or churning swings the number wildly. What looks like a trend in the “enterprise / Pro / EMEA” cell is often just one account moving, or noise you have mistaken for signal. The finer you slice, the more your metrics drift from measurement toward anecdote.

So the right level is not the finest one available. It is usually the coarsest grain that still separates things that genuinely behave differently—and no finer. Finding that grain takes two things, and neither is enough on its own.

Understand the business, not just the data. The grain has to follow how the business actually works: how the products fit together, how customers operate, how the sales team sells. Much of that never shows up cleanly in a warehouse. A ten-minute conversation with a front-line sales rep will often surface a distinction that would take a week to tease out of the data.

Then slice the data. With that context, cut the data every way you can and watch which attributes actually separate behavior. Most will move together; a few split into genuinely different profiles. Those few are your entities; the rest are noise dressed up as dimensions.

Build for movement between levels

Suppose you have settled on the right grain. Are you finished—no reason to look at the data any other way again? Not for long.

A grain is a working decision, not a permanent one. It needs constant validation; the business will redraw it as products, segments, and sales motions change; and some metrics only resolve when viewed at more than one grain at once. Each of those points to the same conclusion: the model that lasts is the one that treats grain as flexible rather than fixed—the one that can absorb next month’s ad-hoc question without a rebuild.

A model that carries the underlying dimensions—customer, product, segment, cohort, period—lets you roll up and drill down on demand. It lets an analyst ask the same question at three grains in a minute instead of a 2-day sprint. The ability to move between levels is what turns “the number looks fine” into “the number looks fine, and I checked underneath.”

The strongest models go a step further: they do not just let you move between grains, they compute across them at once. Return to the example account that swapped Basic for Pro. Run churn at the product grain and the dropped Basic edition books as churn, full stop. A multi-grain metric knows the same customer picked up Pro in the same motion, and nets the two—the Basic loss resolves into a net upsell figure instead of landing in churn, and gross churn is left reporting only the customers who actually left. That only works because the model can see the customer grain and the product grain at the same time and classify the swap for what it is. It is where a surprising share of “why doesn’t churn tie to the renewal report” disputes actually begin—two teams computing the same event at two different grains.

That is the quiet thesis behind the name of this practice. Getting your analytics to the next level is rarely about a fancier tool or another dashboard. It is about getting to the right level—the grain that answers the question you are actually asking—and being able to move when the question changes. The number that improved while everything beneath it got worse is always sitting there. The only defense is to look underneath.