What the simulator says so far
Loading the stored batch.
Three Search & Shopping decisions, worked through with turning points and the next measurement
Five things worth knowing
Worked out from the stored batch each time this page loads. Every figure is a simulated range under stated assumptions, not a forecast.
How a result is made
Five steps, the same for every card. How it works has the pictures.
- Draw a worldEvery unknown (how often people return, how often a search ends in a decision, how much ads annoy) is picked once from its plausible range.
- Check it against PinterestThe world runs for six simulated months. It is kept only if it reproduces the figures Pinterest publishes. About one draw in five passes.
- Run it twiceSame people, same luck. Once as things are, once with a card switched on: a feature, a rival's move, a change in ads.
- CompareThe gap between the two runs is that card's effect in that world, as a percent change over 150 days.
- RepeatAcross many kept worlds the effects form a range. The range is the finding.
How to read a result
One real row, labelled. Every picture on this page is built from the same four things.
- Typical change
- The median: half the simulated worlds came out above it, half below.
- Range
- Where nine simulated worlds in ten landed. A wide range means the answer depends on things nobody outside Pinterest knows.
- Worlds up, unclear, down
- How many worlds moved by more than twice the simulation noise. "Rose in 190 of 192" means the direction held across nearly every sampled set of assumptions. It does not mean the size is right, or that assumptions outside the sampled ranges would agree.
- Decisions reached on Pinterest
- The headline for features. A project (an outfit, a room, a gift) that reached a choice during a Pinterest visit. It is a simulated quantity with no public counterpart.
Features, ranked by change in decisions reached on Pinterest
One row per feature, all on one scale. The dot is the typical change and the bar covers nine simulated worlds in ten. Features modelled two ways show both. Pick a name to open it in the simulator.
Where the two methods agree, and where they do not
Several features are modelled both ways, reaching the same visits. Agreement is weak comfort, since both could be wrong together. Disagreement is useful: it marks the features whose result depends most on how they are modelled.
If you could ship three
Each world ranks the features by change in decisions, all under research priors so they are on one footing. A feature that makes the top three in most worlds is a steadier bet than one with a high typical change and a wide range. Build cost is not in this.
Outside events, ranked by change in visits
Things that happen to Pinterest. These are judged on visits, since they act on whether people come at all. The scale here is not the scale of the feature picture.
Bundles
Several cards at once, to see whether effects add up, and whether a feature bundle can offset a rival's move.
Ad load: what more or fewer ads does in this model
The same worlds with ad load set to multiples of today's. More ads earn more per visit at once and cost visits slowly. How fast they cost visits is tuned to studies of other platforms, so this shows the shape of the trade and cannot find Pinterest's best ad load.
The three revenue columns
Price per ad fixed: revenue follows the number of ads shown. Budgets fixed: advertisers spend the same whatever the inventory, so revenue cannot change by construction. Follows click-outs: budgets track the clicks advertisers receive.
One year on: what the people you already have are worth
Takes the people who were monthly users on day 0 and follows only them for a year: the revenue they bring and how many are still monthly users. This is the nearest thing here to value per existing user. It is not lifetime value: going past a year would need a rule for when people leave for good, and public data cannot pin one down.
What to measure first
For each feature, the assumption its result is most tied to, and the typical change when that assumption is low, middling or high. A big spread between low and high means measuring that one thing would narrow the answer most.
How far to trust this table
These are associations inside a selected set of worlds, not causes. "Narrows the range by" is how much tighter the outcomes are inside thirds of worlds split by that assumption. Splitting on anything narrows it about 3% by chance, and these inputs were picked because they correlate in this same batch, so the figure flatters them. Read it as a rough order.
Full scorecard
One line per card. Features are graded on decisions and events on visits, so do not compare the two.
What the columns mean
Size: small is a typical change under 0.5%, medium up to 3%, large above. Revenue per monthly user: the change in revenue divided by the change in monthly users, with the price of an ad held fixed. Starting users still active: the change in the share of people who were monthly users on day 0 and still are on the last day. Revenue from starting users: the nearest thing here to value per existing user. One 150-day window cannot say how long people stay over years, so no lifetime value is shown.