Retention is a UA metric

UA teams get paid to buy installs at a price, and how long the player stays decides the price they may pay. That second half of the sentence is the part most organisations assign to a different department. Retention sits with product, LTV sits with analytics, and the UA team receives a target CPI derived from both and gets on with hitting it. The result is a growth function where the people spending the money have no formal responsibility for the number that decides whether the money was well spent.

The argument here is that retention is a UA metric in the same way CPI is: measured by source and reviewed weekly, with the person who chose the traffic owning it. Not because product should stop caring, but because the acquisition channel is the single largest controllable input into what retention looks like, and no one else can act on that.

The mechanism: retention is mostly decided before the install

A cohort's retention curve is the product of two things: how good the game is, and who was sent to it. Product owns the first. UA owns the second entirely.

The same game will show very different D7 retention across sources and across creatives, and again across bidding targets; the differences aren't noise. A creative that promises a mechanic the game doesn't deliver recruits players who leave at the first session. Tune a bidding target to install volume and it finds the users cheapest to convert, who are disproportionately the ones who install anything. A rewarded placement recruits players motivated by the reward, not the game. The UA team made each of those choices, and each shapes the retention curve before the player has seen the tutorial.

This is the point Adjust's retention report made in March in aggregate terms: sessions rose across gaming genres while installs softened, which is another way of saying the market shifted from finding more players to keeping the ones it found. We covered the buyer's reading in "ATT opt-in is 39%. Adjust's retention report and the paid-to-organic ratio nobody budgets for." The structural version of that observation is that when installs are flat, the only remaining lever on revenue is the quality of who you buy, and retention is how you measure quality.

The incentive problem, and why it persists

If this is obvious, why does the org chart resist it? Because retention is a slow metric and CPI is a fast one, and teams answer for whatever shows up this week.

A UA manager who hits CPI target is visibly succeeding on Friday. Nobody knows whether the cohort retains until D7 at the earliest, D30 for anything that matters to payback, and by then the campaign that produced it has gone through three more rounds of optimisation. The feedback loop is too long to attach blame or credit, so the organisation attaches it to the fast metric and lets the slow one drift to a department that reports monthly.

The automated networks have absorbed this incentive exactly. Their models optimise toward whatever event you send back, and most teams send back installs and an early purchase event because those arrive fast enough to train on. The model is not wrong to optimise for what it is told. The team is wrong to tell it something that does not predict value.

Here is the trade-off. When UA does take ownership of retention by source, the first thing it discovers is that its cheapest channels are usually its worst retainers, and the first instinct is to cut them. That is frequently a mistake.

Cheap, low-retaining traffic is not automatically bad traffic. It's traffic you're measuring on the wrong horizon, or the wrong metric. Rewarded and ad-supported sources produce players who churn faster but may monetise through ad impressions on the sessions they do have. In an ad-supported title, which Sensor Tower's May data put at 84 percent of downloads, that can be perfectly good economics. The failure is not buying low-retention traffic. It is buying it while measuring it against an IAP-shaped LTV model.

The second trade-off is with product. Once UA is accountable for retention, it gains a legitimate interest in the tutorial and the first-session pacing, and beyond that in the day-two return hook. That interest is healthy, but it creates a new negotiation. Product will reasonably say that UA is sending the wrong players; UA will say, just as reasonably, that the first session loses the right ones. Both are usually true. The organisation needs a forum where the two settle it on cohort data rather than seniority.

A working framework: retention by source, on a UA cadence

The structure we recommend has three parts, and you can run it from an MMP export and a spreadsheet.

Weekly: D1 and D3 retention by source and by top creative concept, compared against a trailing eight-week baseline for that same source. The question isn't "is D1 good" but "did D1 move on this source since last week", because a move usually means a creative or a bidding change altered who you're recruiting.

Monthly: D7 and D30 retention by source, set against the CPI paid and converted into a retained-user cost. Divide CPI by D7 retention to get cost per D7-retained user. This single number reorders most channel rankings; it's the one to put in front of finance.

Quarterly: the shape of the retention curve by source, not just the point values, since two sources with equal D7 can have very different D30 if one recruits players who plateau and the other recruits players who keep leaking. Curve shape is what payback actually depends on.

A worked illustrative example

Consider an illustrative puzzle title buying on two networks. Network A delivers installs at 1.80 with D7 retention of 12 percent. Network B delivers at 2.60 with D7 of 22 percent. On CPI, A wins by a wide margin and the team's weekly review has been rewarding it for a quarter.

Cost per D7-retained user on A is 15.00. On B it is 11.82. B is cheaper by the metric that predicts revenue, and by roughly a fifth. If the title is IAP-led, the team has been scaling the wrong channel for three months. If the title is ad-supported, it needs a further step: ad revenue per retained user on each source over the same window, because A's churners may still be watching rewarded video on day two.

The decision rule that falls out of this: never rank a channel on CPI alone once it has ninety days of history. Rank on cost per retained user at the horizon that matches your monetisation model, and only then look at CPI to understand why the ranking is what it is.

Who owns the number

The practical answer is that the UA lead owns retention by source and product owns retention by cohort age, and the two meet monthly on the same table. The UA lead's brief changes from "hit CPI" to "hit cost per retained user", which is a harder job and a slower one, and the only version of the job that a flat-install market will keep paying for.

Related archive reading

These articles provide related context and remain subject to their stated review status.

• UA and product teams optimise against each other
• The platform tax and the true cost of an install