How Pinterest catches advertisers before they leave

Like any ad-supported business, Pinterest lives on advertisers who keep spending. The trouble, as its engineers put it in a 2023 write-up, is that churn was handled reactively: the sales team only reached out after an advertiser had already stopped, and by then it is incredibly difficult to bring a customer back. In An ML based approach to proactive advertiser churn prevention, they describe how they moved from reacting to predicting, and, importantly, how they proved it worked.

Defining churn simply

Half the battle in a churn model is a clean definition. Pinterest kept it blunt: an advertiser is active if they spent more than zero dollars in the trailing 7 days, and churned if they spent zero. The model looks only at currently-active advertisers and predicts whether each will churn in the next 14 days. That is it. A crisp, measurable label beats a clever, fuzzy one.

How they built it

The model is a gradient-boosted decision tree working from a snapshot of each advertiser at a point in time. It reads more than 200 features across six groups: Performance (impressions, clicks, conversions, spend, CPM, CPC, CTR), Goal (how close they are to their target), Budget (budget and how much of it they use), Ads Manager activity (creating, editing, and archiving campaigns, pulling reports), Property (sales channel, country, industry, tenure, size), and Campaign configuration (targeting, bid strategy, objective, end date). Most features come in as minimum, average, and maximum over the past week and month, plus week-over-week and month-over-month change, so the model sees not just the level but the direction of travel.

They chose a tree model deliberately. It performs well on this kind of tabular data, it makes feature importance easy, and it works cleanly with SHAP, a method that attributes each individual prediction to the features that drove it. That last point is what turns a score into something a salesperson can act on.

Turning a score into an action

A churn probability on its own does not help an account manager. So Pinterest wrapped it in something usable:

• Every active advertiser is scored daily and bucketed into High, Medium, or Low churn risk, with the thresholds tuned to the precision and recall the sales team actually wanted.
• Account managers see a Churn Information Widget: the risk tier plus a SHAP-derived "churn reason" telling them why this account looks shaky, so they can prioritize who to call and what to talk about.

Then they did the thing many teams skip: they ran a proper randomized experiment. North American small-and-medium advertisers were split into a treatment group (account managers see the churn info) and a control group (they do not). The targets were sensible: recall above 70 percent across the high and medium tiers, precision around 70 percent in the high tier, and the model held up in production, with its live ranking and precision-recall scores landing within a few percent of the offline numbers. The result: a 24 percent reduction in churn rate for the high-risk accounts, a statistically significant win for the group whose managers could see the predictions.

What has changed since

Pinterest flagged sequential models (LSTMs, transformers) as future work to capture behavior over time, but has not publicly confirmed shipping them; its later advertiser-AI announcements are about creative and campaign automation, a different problem. The broader retention field, though, has moved in two directions worth knowing:

• From "who will churn" to "who can I save." Predicting churn is not the same as knowing where to spend a retention offer. Uplift (or causal) modeling targets the "persuadables," the customers an intervention will actually change, instead of wasting discounts on people who would have stayed anyway or those already lost. A 2025 insurance program using this approach reported cutting voluntary churn by 2.3 percentage points while using smaller discounts.
• Language models for the "why" and the "what next." Newer systems use LLMs to explain the likely churn reason and suggest the next best action, going a step past SHAP's feature attribution.

Why it matters for your business

Churn is not just an ad-platform problem. It is every subscription, every retainer, every repeat-buyer relationship. The transferable lessons:

• Predict, then act early. A customer showing warning signs today is far cheaper to keep than one already gone.
• Define churn crisply. A simple, measurable label (no activity in N days) beats an elaborate one.
• Ship the reason, not just the score. A sales or account team needs to know why an account is at risk to do anything about it. SHAP, or a language model, turns a number into a conversation.
• Prove it with an experiment. A holdout group is how you know the model changed outcomes, not just that it made predictions.

How we would build it today

For an Indonesian agency, SaaS, or dealership CRM, we would start with a gradient-boosted model predicting which clients go quiet in the next few weeks, from your own usage and spend history, with a clear "no activity in N days" label. We would surface each at-risk client to the account owner with the top reasons attached, inside the tools they already use. Then we would add the step Pinterest's own roadmap points to: not just who is at risk, but who an intervention can actually save, and let a language model draft the outreach and the likely reason. And we would keep a holdout group, so we can prove the retention lift in rupiah, not just in dashboards.