PM case study · Unofficial concept

My Inspiration Preferences

Making existing generative-AI content controls easier to find, easier to understand, and more effective in the feed.

All metrics on this page are simulated examples created to illustrate how the experiment would be evaluated. They are not real Pinterest data or measured results. This prototype is not affiliated with Pinterest.

Problem

Unwanted AI-generated content can make inspiration discovery feel less relevant or authentic.

People come for ideas they can actually make, wear, cook or visit. Realistic AI images blur that line and erode trust in what they save.

Hypothesis

Giving users better control over AI-generated content will improve discovery satisfaction and meaningful saves.

If controls are visible where frustration happens, more people will use them, and the feed will better match intent.

What changes

Not a new filter — a better experience around the controls that exist.

Discoverable. “Tune my feed” lives in the feed, not deep in settings
Transparent. Subtle AI labels, “why you’re seeing this”, and label reporting
Effective. Preferences refill the feed with relevant human-made Pins

Pinterest already offers some generative-AI labels and controls. This concept does not claim to invent AI-content filtering, and acknowledges that AI-content identification is imperfect.

North Star for the experiment

Simulated

Meaningful Pins saved per active discovery session

A save counts as meaningful if, within 7 days, it is revisited, clicked through, or organised onto a board. This rewards saves people come back to, not saves that just inflate engagement.

Control

1.42

Treatment

1.58

+11.3% illustrative lift · 95% CI ±3.8% (simulated)

A/B test design

Control

Existing Pinterest discovery experience.

Treatment

Prominent feed controls, clearer labels, and improved application of AI-content preferences.

Unit: User, 50 / 50
Surface: Home & Explore
Duration: 4 weeks (2 full cycles)
Decision: Ship if North Star ↑ and guardrails hold

Secondary metrics

Simulated

Metric	Control	Treatment	Change
Users engaging with content preferences Opened Tune my feed or changed a setting	1.9%	7.4%	+5.5 pts
Pin save rate Saves ÷ Pin impressions	3.10%	3.36%	+8.4%
“Not interested” actions per session Lower suggests fewer unwanted Pins	0.41	0.33	−19.5%
Seven-day returning-user rate Returned to discovery within 7 days	46.2%	47.5%	+1.3 pts

Guardrail

Simulated

Advertising revenue per user

−0.4%

Within threshold (No worse than −1.0%)

Promoted slots unchanged; AI-creative ads deprioritised for opted-in users

Guardrail

Simulated

Feed latency (p95)

+6 ms

Within threshold (≤ +20 ms)

Preference filter applied at re-rank stage

Guardrail

Simulated

Diversity of discovered creators

+2.4%

Within threshold (≥ −2% vs control)

Unique creators per 100 impressions

In this prototype

What “Prefer human-created content” does to the feed

Computed live from the prototype's sample dataset of 69 Pins — mechanics, not outcomes.

Top of feed	Before	After
Organic Pins shown	26	26
AI-generated Pins shownDetection misses some — they still appear	12	2
Promoted PinsAd load unchanged	4	4
Unique creators	20	21

Constraints we kept

Ads stay. Promoted slots are fixed; opted-in users see non-AI creatives first.
Imperfect detection. Labels can be wrong, so every label is reportable and copy never promises accuracy.
Relevance first. Removed Pins are replaced with the next most relevant human-made ideas.
Creators using AI. Content is labeled, not banned, and still shown to people who want it.

This demo session

Counted in this browser only

0
Saved Pins: 0
Saves: 0
Not interested: 0
Preference changes: 0
Label reports: 0
Unsaves

Demo script

1Explore in Before mode — scroll the current experience.
2Turn on AI X-ray to reveal how many Pins are AI-generated.
3Switch to After and open “Tune my feed”.
4Choose “Prefer human-created content” and watch the feed refill.
5Open a Pin: why it’s recommended, its label, and reporting.
6Toggle Before / After, then walk through this page.