The explainer, in four minutes

Transcript

The narration, scene by scene. Every figure in it is one the video shows on screen, checked against this project's files.

1. What if Pinterest changed one thing? 0:00

   What if Pinterest changed one thing? This is a simulator built entirely from public data. Here is how it works, and how to use it well.

2. A made-up Pinterest, built from public numbers. 0:12

   It is a made-up Pinterest. Three hundred thousand simulated people in each world, and a hundred and ninety-two worlds, each with different assumptions. Sixty-seven what-ifs, from new features to rivals.

3. Same world. One change. Compare. 0:30

   It starts from the public record. Every unknown becomes a range. The engine runs people through searches, saves and clicks. Then it runs the same people again with one change switched on, and compares. Code runs the world. A language model only nudges reactions, inside fixed limits.

4. Read the range, not the dot. 0:51

   So read the range, not the dot. Each dot is one assumption world. Here, Google adds a visual moodboard: visits fall about two percent in a typical world, and between about one and a half and three percent in nine worlds out of ten. It is not a forecast.

5. Is any of it grounded in reality? 1:10

   Is any of it grounded in reality? We sorted a hundred and three public figures honestly. Most don't match the engine's definitions closely enough to test. Two activity checks miss. That is the honest starting point.

6. We audited 360 linked Pinterest results. 1:34

   So we went and checked. We audited three hundred and sixty real Pinterest results. About two in three instruction pins gave usable steps. In default shopping search it was lower. With the Products filter, nearly nine in ten led to something you could buy.

7. Three bets, worked through. 1:57

   Here is the useful part: three bets, worked through. Keep the shopper's choices: it helps, unless choices already carry over most of the time. Match shopping to the task: the revenue cost comes from cutting ads, not from matching. Destinations that deliver: a screen rarely beats better relevance, and only when it keeps almost all the useful results. Each brief ends with the next measurement.

8. Two activity checks missed. We raised activity to their reference levels. 2:30

   Then we raised the engine's activity to the public reference levels and ran it all again. All three bet checks held.

9. Listening beyond r/Pinterest. 2:45

   We listened beyond Reddit, in twenty-four places. In Japan, twenty-two of fifty App Store reviews were about search breaking Japanese typing. Shoppers ask where a pin is from. Advertisers ask where the sales went.

10. 1,014 features and changes 3:04

    And we catalogued over a thousand Pinterest launches and changes since twenty ten, with what people said about them where we could find it.

11. Six rules. 3:19

    Six rules. Read the range. Ask which method. Find where it turns. Check the precedent before the size. Trust nothing near zero. Then go measure.

12. Pinterest Simulator 3:38

    Pinterest Simulator. Public data only. Ranges, not forecasts. Start with the three bets.

Credits

Narration voiced with ElevenLabs text to speech. Music synthesised for this video. Illustrations are AI-generated decoration, not Pinterest content or evidence. Charts and examples come from this project's public-data research and stored simulations.