This section is desk research and structural analysis of a product I use, not user interviews. What it establishes is the size and shape of the opportunity, how pairs form in practice, and what can't be known from outside.
Figures are from Duolingo's product blog and from 2024 and Q2 2026 shareholder reporting.
The decision usually happens outside the app. Two people talk, agree, and come to Duolingo to act on it. That shapes the whole entry side of the feature: the invite flow serves someone who has already decided, not someone being persuaded, and the invitee may open the app days after the conversation, so an invite has to be findable rather than persuasive.
It is also what makes the lapsed case work without an in-app hook. Two friends who both stopped can decide together and both return, and neither had to be reachable.
Anything the app does to promote the feature is there to circulate the idea, not to convert. Its job is to make sure the conversation can happen at all.
The computable segment is users with a lapsed or low-progress course who also have at least one connection. Both halves are queryable.
The 57% is a floor, not a ceiling. Invites can reach people you don't follow: sending one creates the follow, accepting creates the return follow. The friendship usually already exists; it just wasn't in Duolingo yet. Course Buddies creates connections it doesn't require.
What can't be computed is the conversion through that segment. That is the pairing rate, and it is the key unknown.
Any funnel here would rest on inputs I don't have, and every invented figure becomes a target that can't be defended. The question that matters is whether the feature plausibly moves the metric or is a nice thing that makes the app nicer, and that is answered by the mechanism rather than by arithmetic: the quiz exchange is the only daily-cadence element in the feature, Duolingo's own comparable delivers 22% on daily completion, 57% of users have someone to pair with, and pairing rate is the number V1 measures first.
Pairing rate. The largest unknown. The claim is not that everyone pairs. It is that a meaningful share do, and that a pair is worth many times a solo starter. If the rate can't clear a bar, that is the kill signal.
Self-selection. Engaged users are naturally more social, so a naive comparison of paired against unpaired learners would overstate the effect. Measurement has to be designed around this, and the reactivation read in particular needs a holdout. This is named here because it is the most likely way the results could mislead.