Pinterest now guesses your big projects to send better alerts

Pinterest now tries to work out what bigger project you are working on, then sends emails and push alerts to match. It replaced an older system with a fine-tuned language model, the kind of AI behind chatbots. In early tests, email clicks rose 1.1% and push notification opens rose 1.3%.

What changed

Pinterest calls a long-running goal a journey. Think of someone renovating a kitchen. They search for cabinet designs, save countertop ideas, then look at backsplash tiles and lighting. Those steps point to one project, not several unrelated interests.

The old system pulled keywords from activity, grouped them, named the groups and ranked them. It beat earlier interest-based alerts. But it could split one project into pieces, give vague labels, or mistake a stray click for real intent.

How the new system works

The new model reads up to 150 of a person's actions over 360 days, plus a short profile. It then writes a ranked list of journey names. Searches and saves count for more than incidental clicks. It is also told to weigh repeat activity, recency and specificity.

• It is told not to merge unrelated interests.
• It is told to ignore one-off activity and noisy automatic tags.
• It can return nothing when there is not enough evidence of a real journey.

Pinterest treats changes to these instructions as production launches, not casual edits.

How it was built

Pinterest did not have a large set of human-labeled examples. Instead, it used advanced AI models to create synthetic training examples from user activity histories. Those were used to fine-tune smaller open-weight Qwen3 models. The team tested sizes from 0.6B to 8B parameters, a rough measure of model size, and picked 4B as the best balance of quality and speed.

One big speed gain came from doing less. The system tracks when each user's journeys were last generated and skips anyone processed in the past couple of days.

What the results mean

The numbers come from early A/B tests in the US and Canada, compared against the old clustering system. Pinterest also reports statistically significant gains in engagement and retention, without giving figures. The gains are modest, and they come from the company's own reporting.

For anyone who markets through apps or email, the lesson is simple. Alerts tied to a person's ongoing goal can beat alerts based on scattered interests.