Start here

What this site is, what changed overnight, how to read a result, and how far to trust one. About ten minutes.

The ten-minute routeWatch the video
Stored batch
Loading

What this is, in one minute.

This is a personal model of Pinterest, built only from public figures. It holds a made-up population of simulated people, a set of journeys such as planning an outfit or finding a recipe, and a model of ads. Code runs that world one day at a time.

You pick one change, called a scenario card. The engine runs the same world twice, once with the change and once without, and reports the difference. It repeats that across many simulated worlds. Each world is one full set of assumptions, and it is kept only if its starting point matches the figures Pinterest has published.

The most useful output is usually a question, not a number: which assumption does this result lean on, and what would you measure to settle it?

What is new since yesterday.

Eight things changed on 12 and 13 October 2026. Figures on this list are read from the site's own files when the page loads.

One look on every page
Every page now uses one design. The home page is a walkthrough in eight steps, with a field of 640 dots that each stand for about one million monthly users in the model. Blue means a metric rose, amber means it fell, grey means no clear change.
Home
The briefing
Search and shopping at Pinterest from the public record: what the company says it is doing, what it has shipped, what users say, and where those two accounts pull apart. It ends with a first ninety days.
Briefing
What was tried
A record of documented launches, experiments and removals, by year and area. Use it to see whether a what-if has a real precedent.
What was tried
Tests against reality
The engine was given the nearest available change for a set of published results and comparisons. Direction mostly matched and size mostly did not. A replay of the 2021 search update could not match the reported decline. The page also dates real public events.
Reality
The engine change
A click out of Pinterest now leads somewhere that delivers or does not. A failed destination blocks the purchase or the finish from that click and costs satisfaction. There are three new cards, D01 to D03, and one old card (F09) now repairs delivery.
How it works
A counted audit of the live site
Searches on the live site were counted by script: how many result tiles were ads, how many linked out, and whether a sample of linked pages gave what the pin promised. One account, one evening.
In the briefing
What customers ask for
30 asks from public reviews and forums, ranked by strength of evidence. Two dominate: fewer ads, and an AI switch that really works. In the 500 newest US App Store reviews, 214 mention ads and 136 mention AI.
Customer asks
What the simulator says about them
Sixteen of the asks were run through the engine as 18 cards across 48 paired worlds, and 20,842 public Reddit posts were sorted by subject with a local model, and a second model checked a sample of 300. Under the stated assumptions two asks raised both decisions and revenue with the whole range above no change and beyond simulation noise (keep my place, search for the words I typed), and the page shows what happens when they ship together. In the posts, interface changes and ads are the subjects that grew over the last six months.
The comparisons

How to read a result.

One real stored result, labelled. Every result on the site is built from these six things.

  1. The sentenceRead this first. It gives the typical change, then where nine simulated worlds in ten landed.
  2. One dot per simulated worldEach dot is one full set of assumptions, run with and without the change. Hover a dot for its number.
  3. The nine-in-ten barIt runs from the 5th to the 95th percentile of the worlds. It shows how much the assumptions disagree. It is not a margin of error for reality.
  4. The large dotThe typical world: half came out above it and half below.
  5. The no-change lineZero. A bar that crosses it means the worlds do not agree on direction.
  6. ColourBlue rose, amber fell, grey moved too little to tell from simulation noise. A fall is not always bad and a rise is not always good.

Try it: when worlds agree, and when they do not.

Three real stored results. Switch between them, then drag the slider to use fewer of the stored worlds.

A ten-minute route.

Eight stops. Each has one thing to do and one thing to notice. The links open the right scenario.

  1. Briefing1 minute

    DoRead the one-screen version at the top.

    NoticeEach claim carries its source. The company's stated problem is turning clicks into revenue, not finding users.

  2. Scenarios: a clear result2 minutes

    DoRead the headline sentence, then hover a few dots.

    NoticeEvery dot sits on the same side of the no-change line. The worlds agree on direction and differ on size.

  3. Scenarios: an unclear one1 minute

    DoOpen this card and read its bet card.

    NoticeDots fall on both sides. The model does not know the direction. The bet card says what to measure first.

  4. Watch one world2 minutes

    DoPress Play. When it ends, press Run again with new luck.

    NoticeThe line wobbles between runs. That is luck: the same assumptions with different people and chance events.

  5. Map1 minute

    DoClick the United States, then California.

    NoticeThe note that says these figures are allocations. Pinterest reports three regional totals and nothing finer.

  6. People1 minute

    DoPick the first person and read the two diaries.

    NoticeThe chip that says Fiction. Diaries make a result easy to picture. They are not evidence.

  7. Customer asks1 minute

    DoOpen the top ask.

    NoticeThe label for how strong the evidence is. A loud complaint is not the same as a common one.

  8. Reality1 minute

    DoScroll to "Tests against what really happened".

    NoticeDirection matched far more often than size. That is the single most useful thing to remember about this engine.

Or let the site walk you through it. The tour visits the real pages, rings one thing at a time, and can be ended with Escape.

Five questions this can help with, and five it cannot.

It can help with

  1. What has to be true for this idea to work?Each card states its assumptions. The page shows which one the result moves with most.
  2. What should we measure first?The bet card names the smallest measurement that would settle the result.
  3. Is the direction steady, or could it be luck?Count the simulated worlds on each side of the no-change line, then run one world with new luck.
  4. Could a fix backfire?The new delivery cards trade better destinations against fewer results to choose from. Either side can win.
  5. Has Pinterest tried something like this?What was tried and the briefing give the public precedent, with sources.

It cannot answer

  1. How much will this feature lift users or revenue?Nobody publishes the rates that decide it. No feature effect here has been tested against real people.
  2. What is the chance this happens?A share of simulated worlds is not a probability. How often the range covers a real outcome is unknown.
  3. What will next quarter's reported users be?On past quarters a simple seasonal rule did better than the engine.
  4. What did a rival's launch do to Pinterest?That cannot be estimated from quarterly totals. Credible studies use weekly data with clean controls.
  5. What will happen in one country or state?Those figures are the model's own split of three reported regional totals.

What to trust, and what not to.

The engine was tested against what really happened. These are the results, read from the test files.

Reality tests
Loading
Direction
Some support
In the published results and comparisons checked, the engine usually moved the same way.
Size
Do not rely on it
Size was usually wrong by more than a factor of three. Read a range as "small", "medium" or "large" at most.
User counts
Do not rely on it
Run from past quarters, the engine loses to a simple seasonal rule.
History
Failed
A replay of the November 2021 search update could not be made to fit: none of 120 draws matched the published growth, and without the event the model put US monthly users at 90.54 million when Pinterest reported 86 million.
Feature effects
Untested
No feature effect has had a fair test against real people.
The baseline
Partly by construction
The starting world matches Pinterest's published totals because worlds that do not match are thrown away, and regional users are scaled to fit. That constrains the baseline. It says nothing about any feature.
The public record
Check the source
The briefing, What was tried and Customer asks quote public sources and say how each was read. Complaints are hand-picked accounts, not a sample of users.

Words used on this site.

Simulated world
One full set of assumptions, run with and without a change. Never a real place, and never a chance of a real outcome.
Range
Where nine simulated worlds in ten landed: from the 5th to the 95th percentile. It is the spread of the assumptions.
Baseline
The same world without the change. Every result is a difference against it.
Scenario card
One what-if: a feature Pinterest could ship, something that could happen to it, or several at once. Each has a code, such as E03.
Journey
A kind of project a person brings to Pinterest, such as an outfit for an event, a room refresh or a weeknight dinner.
Stage
How far along a project is: exploring, comparing or deciding.
Mechanism
One of two ways to model a feature. The feature changes how a visit plays out, and the engine works out what follows.
Research prior
The other way. The feature reaches a share of visits and lifts the chance of progress by an amount taken from research ranges. Those ranges are assumptions.
Luck against assumptions
Luck is who the simulated people are and what chance events they meet. Assumptions are the rates the world runs on. "New luck" changes the first. "Fresh worlds" changes the second.
Allocation
A country or state figure. Pinterest reports three regional totals. Anything finer is the model's own split, not a measurement.
Docking
Giving the engine the nearest available version of a published result or comparison and comparing direction and size with what Pinterest reported.
Delivery
Whether the page behind a click gives what the pin promised: real instructions, a working source, a product that is there to buy.

How to run and refresh it.

Run these from the project folder. The site reads stored results, so nothing on a page changes until a new batch is made.

./target/release/pinsim serve          # the site, at http://127.0.0.1:7878

cargo build --release                  # rebuild the engine after a code change
python3 tools/build_world.py           # rebuild the world from the research files
./target/release/pinsim --worlds 192   # the full stored batch, two hours or more with 67 cards
python3 -I tools/explain.py            # rewrite the plain-language explanations, about ten minutes
python3 -I tools/history.py            # refresh the dated events on the Reality page
cargo test --release                   # the engine's own checks

python3 -I tools/voices.py audit       # second-model check of the coded posts (all 20,842 are coded)
python3 -I tools/voices.py report      # then rebuild the post figures on Customer asks

If a page shows a banner about older stored results, run the batch line again. The full list of commands is at the top of PLAN.md.