simulations

Nudges, at scale

deepdive

The system behind every tooltip, push and "come back": targeting, delivery, templating, measurement — derived from first principles, every approach followed to where it breaks, from one user to a hundred million.

Nudges, at scale

topic

One system, five hard problems, and a defining tension: you are spending someone else's attention, on credit.

derive itaudiencerenderedeventslearnWhat a nudge isthe irreducible parts＋Targetingwho gets it＋Deliveryhow it reaches them＋Contentwhat it says, on screen＋Measurementdid it work＋The tradewhat you never get all of＋

A nudge is a message the user did not ask for, delivered at a moment they did not choose, hoping to change what they do next. Push notifications, in-app tooltips, banners, emails, badges — different costumes, one system. And that system is deceptively wide: it is a database problem (who are the users), a distributed-systems problem (reach ten million of them in minutes), a UI problem (draw over someone else's screen without breaking it), and a statistics problem (know whether any of it worked) — all wearing one product name.

The defining tension never goes away: every nudge spends attention the product has not earned yet. Send too little and the system is dead weight; send too much and users disable the channel — an opt-out is close to irreversible, which makes over-sending one of the few genuinely unrecoverable mistakes in product engineering. Every part of the architecture below is partly a control on that spend: caps, arbiters, holdouts, fatigue models.

This is a deepdive: each part follows every credible approach to each sub-problem — including the naive ones — far enough to see where it breaks, with arithmetic where scale decides and labs where this site's own engine can run the claim. The map below is the system's block diagram; the closing article is the trade-off you cannot escape.

a ＋ goes one level deeper — click the box

go deeper

references

Plotline — the ultimate guide to in-app nudgesThe vendor view of the in-app half of this system: nudge types, placement, and what no-code deployment implies about the architecture underneath.
LinkedIn Engineering — ConcourseNear-real-time personalised notification generation at LinkedIn scale: what the candidate-generation half of this topic looks like when it is real.