What is retention rate?

How it is calculated

Take a cohort — everyone who first used the product on a given day — and measure what proportion return on day 1, day 7, day 30. Each checkpoint is computed against the original cohort size, not against whoever survived the previous checkpoint.

Day 1 tells you whether the first experience worked. Day 7 tells you whether a habit formed. Day 30 tells you whether the product has a place in the user's life. Products fail at different checkpoints for different reasons, which is why all three are tracked rather than a single number.

Why it governs everything downstream

Retention is the multiplier on every other metric. LTV is retention times monetisation, so a change in retention moves the amount you can afford to pay for a user more than almost any other lever.

It is also the honest judge of acquisition quality. Two sources at the same CPI with different day-7 retention are not comparable buys, and the install-stage metrics will never reveal it. This is the single most common blind spot in user acquisition — see OEM advertising and offerwall traffic for the clearest examples.

Reading it correctly

• Always by cohort. Blended retention across all users moves when acquisition volume changes, independently of the product.
• Always by source. An account-level average hides that one channel produces users who stay and another does not.
• Against your own baseline, not against a published benchmark from a different category.
• Alongside the curve shape, not just the checkpoint — retention that flattens has found a core audience; retention that keeps declining has not.

Retention and AI-acquired users

Early evidence suggests users arriving from AI assistant recommendations tend to retain better, and the reason is structural rather than surprising: they arrived having already stated their constraints, compared options in conversation, and chosen deliberately.

That is a selection effect rather than a property of the channel, and it is exactly the selection effect you want. It is also a reason to cohort AI-sourced users separately while volumes are small, rather than blending them into a paid average where the difference disappears.