6.100The tyranny of the marginal user
A product that grows by counting users keeps reshaping itself for its next, least engaged user, and slowly gets worse for the people who use it most.
Ivan Vendrov, "The Tyranny of the Marginal User", on his Substack Nothing Human (14 September 2023), written after six years at Google.
- 1Read
Vendrov's argument starts with the business. Consumer apps want more users, even users who get little from the app, because low-value users can still be shown ads and because network effects make every extra user a little more valuable. So the target is a count, typically daily active users (DAU), the people who open the app in a 24-hour period. Since each user pays about the same, often nothing, and incentives work at the margin, a company with a billion users is not really designing for those billion. It is designing for the billion-and-first, the person closest to drifting away. His example is a small "see less like this" button under each item in a feed: it gives users more control, but it takes a few pixels from a headline or a picture, the least committed users scroll a little less, daily users dip in the A/B test (a randomised comparison of two versions), and the button never ships. Repeat that for years and settings, depth and user control wear away while feeds and defaults grow.
Each step is locally sensible, which is why it is hard to stop. The test measures the change in the average, and averages hide the spread: a feature that matters a lot to a few heavy users and costs a sliver of attention from many casual ones loses. Committed users are also slow to leave, so the damage to them shows up late; Vendrov notes that by the time loyal users go, the team that shipped the change has been promoted. It is Goodhart's law at product scale: the user count once tracked how much people valued the product, and optimising the count finds the cheapest users to keep. Google's own note on retiring Google Reader in 2013 shows the accounting in plain words: the product had a loyal following, but usage had declined, so it was closed.
There is a fair case on the other side. Simpler products really do help many people, and a tool only experts can use leaves most people out; designing for newcomers is often right. The essay is one insider's argument, not a measured study, and its evidence is personal experience. Read it as a warning about what you count, not as a case against simplicity. The remedies follow from the mechanism: charge users so that value rather than attention pays the bills, read test results by how engaged users were beforehand, track how the most active group is doing, and keep power features reachable but out of the newcomer's way, so the two paths do not fight over the same pixels.
- 2Note
Example
Vendrov opens the essay with online dating. When he first tried OKCupid, around 2016, you wrote a long essay about yourself, answered hundreds of questions about your personality and what you wanted in a partner, and saw a match score from 0 to 100 percent for people nearby; he found the high scores eerily good. By 2023, he writes, it had become another swipe app, see a face and swipe left or right, and he calls the 2016 version the best online dating would ever get. He describes watching the same force at Google, steadily cutting features that users loved.
- 3Apply
Where it applies
When an A/B test says a feature lowers daily users, split the result by how engaged people were before the test, because a loss among casual users and a gain among heavy ones is a trade-off to decide, not a verdict. In an artificial intelligence (AI) product, watch your heaviest users when you tune defaults for first-session retention: hiding the model choice or the custom instructions field may lift sign-ups and quietly lose the people who build on you.
- 4Think
Think about it
Who are your ten most committed users, and which change in the last year made the product worse for them to keep someone who barely uses it?
- 5Log
Applied it to something real? Log it.
- 6Check
Test yourselfAccording to Vendrov, why does a product with a billion users end up designed for its least engaged user?
Growth and ads reward the user count, each user pays about the same, so the gains come from keeping the person closest to leaving. Every change is tested against whether that person stays, and features for committed users lose.
- 7Go
Related models
- Goodhart's lawpage 6-36
When a measure becomes a target it stops being a good measure, because people optimise the number instead of the thing it stood for.
- Incentives drive behaviourpage 6-42
People do more of what a system rewards and less of what it punishes, so to predict behaviour, look at what gets paid for.
- Averages hide the spreadpage 6-9
A mean describes a group, not any member of it; without the spread you cannot tell whether anyone is near it.
- Network effectspage 6-52
A product becomes more valuable to each user as more people use it, which makes leaders grow faster and makes leaving hard.
- Second-order thinkingpage 6-71
Keep asking "and then what?" until you have looked at the consequences of the consequences.
- Goodhart's lawpage 6-36
- 8Source