Who Comes Back? Reading Retention Cohorts Without Guessing

Traffic tells you how many people showed up. Retention tells you how many of them found a reason to return. For anything people are meant to use more than once, such as a tool, a newsletter, a community, a store with repeat buyers, that second visit is the one that matters.

This post explains how a retention table is built, how to read one, what common shapes mean, and where the numbers mislead.

The idea in one paragraph

Take everyone who first visited on a given day. That group is a cohort. Now check how many of them came back one day later, two days later, a week later. Do the same for the cohort that started the next day, and the next. Stack the results and you have a retention table: one row per cohort, one column per period since the first visit.

OnyxMetric's retention view works this way. Each row is a group of users by first-visit date. The first column is the starting point and is always 100 percent. The later columns show what share of that group returned after that many days or weeks. Lighter cells mean more people stayed, darker cells mean more dropped away.

Why cohorts instead of one average number

A single "returning visitor rate" mixes people who arrived yesterday with people who arrived a year ago. It can rise because your product is stickier or because you stopped acquiring new users. You can't tell which.

Cohorts remove that confusion. Each row only holds people who started at the same time, so you can compare January's starters to February's like for like. If February's row is lighter than January's at week two, something changed between those cohorts: the audience, the onboarding, a release.

The three points worth watching

You do not need to study every cell. Three give you most of the picture.

• Day 1. Did people find enough value to return the next day? A weak number here points at the first experience: confusing landing, unclear next step, or a promise the first visit did not keep.
• Day 7. Did the product slot into a weekly routine? For tools used at work, this is often the honest signal.
• Day 30. Have they formed a habit? This is the number that correlates with long-term value.

What counts as good depends heavily on what you run. A recipe site, a SaaS dashboard and a news site will have very different curves, and comparing yourself to a figure from another industry is mostly noise. Compare yourself to your own past.

Shapes and what they suggest

The cliff. A steep fall in the first few days and then a long flat tail. This is common and usually tells you about onboarding. People arrive, do not understand what to do, and leave. Fix the first-run experience before anything else.

The slow slide. Retention falls a little every week with no plateau. Novelty fades and nothing is pulling people back. The remedy is a reason to return: new content, a notification, a workflow that builds on past use.

The plateau. The curve drops and then levels off at some non-zero value. That level is your core audience, the people for whom the product works. Your job is to raise the plateau and get more people onto it.

The smile. Rare, but real: retention dips and then rises. This happens with seasonal or event-driven products such as tax tools or sports sites, where users return when the season does.

Make it useful: ask what separates the returners

A retention table on its own describes. To act, compare. Filter the same table by:

• Acquisition source. Do visitors from search come back more than visitors from social? If so, that channel is worth more than its raw traffic suggests.
• Landing page or entry point. Users who start on documentation might return at triple the rate of users who start on a blog post.
• Device. A mobile experience that is rough will show up as lower retention on phones.
• A key action. Using goals and events, look at people who did something early, such as creating a project, against those who did not. If the first group returns far more, getting more people to do that action is your growth lever. Be careful: correlation is not proof, but it tells you where to run an experiment.

OnyxMetric lets you apply the same filters used across reports, so these comparisons don't need a separate export.

Things that make retention lie

Identity resets

Retention depends on recognizing the same person across days. OnyxMetric is privacy-friendly and cookieless. With user ID salting enabled, identifiers reset daily, so retention beyond 24 hours becomes less accurate. That is the trade you are making for stronger privacy. If long-term retention is central to your business, identify logged-in users so returns can be matched reliably, and do it with the consent your region requires.

Different devices

Someone who visits on a phone and returns on a laptop looks like two people unless you identify them. Anonymous retention will undercount, and more so for products used across devices.

Cohorts too small to read

A cohort of 15 people can swing 20 points on a single return. Widen the cohort (weekly instead of daily) until the rows stabilize.

Incomplete periods

The most recent rows haven't had time to reach the later columns. Don't read blank or low recent cells as a decline.

Bots and one-off spikes

A viral post brings a big, low-intent cohort that drags the row down. Hold that row aside when you judge the product. Bot traffic can do the same, so make sure your bot filtering is on before you trust a dip.

A review routine

Once a month, in ten minutes:

1. Look at Day 1, Day 7 and Day 30 for the last four cohorts. Note the direction, not decimals.
2. If a number moved more than a few points, list what shipped or changed that month.
3. Split the weakest cohort by source and device and see whether the problem is concentrated.
4. Choose one thing to try. Write it down with the metric you expect it to move and the date you will check.

Retention changes slowly. A change you make this week may not appear in the Day 30 column for a month. That delay is why writing down the prediction matters: otherwise you will forget what you were testing.

The short version

Retention answers one question: did anyone care enough to return. Read cohorts, not averages. Watch Day 1, Day 7 and Day 30. Compare your own cohorts over time. Split by source and entry point to find who sticks. And remember that identity limits set the ceiling on how accurate any of it can be.