If you read one line: Simile just raised $200M at a $2B valuation to sell simulated users you survey instead of real ones — and the smart way to use them is at the top of your research loop, to diverge and screen cheaply, never as the final call on a pricing or positioning decision.
The raise#
On July 30, Simile closed a $200M Series B at a $2B valuation, led by Greenoaks, with Index, Bain Capital Ventures, A*, and CVS Health Ventures in the round — just five months after a $100M Series A led by Index Ventures. Revenue is up 5x since the company launched in February 2026, and headcount has gone from a handful of researchers to 50+ (FINSMES).
The credibility signal that matters: founder Joon Sung Park is the Stanford researcher behind the 2023 "Generative Agents" paper — the "Smallville" experiment where 25 AI characters lived simulated lives, remembered each other, and threw a party nobody scripted. That work is why "synthetic users" reads as a category and not a pitch deck. Customers now include CVS Health, Deloitte, Gallup, and Wealthfront.
What a synthetic user is — and what it isn't#
A synthetic user is an LLM-driven simulation of a person: modeled demographics, preferences, and history you can interview or survey at scale, without recruiting a human panel. Point a hundred of them at your landing page and ask which headline they'd click. You get answers in minutes, for the price of tokens.
The temptation is to treat that output as data. It isn't — not the way a real interview is. A synthetic user reflects the model's priors about a population, not the population. It will confidently role-play "a 34-year-old SaaS founder in Austin," but it can't surprise you the way a real one will, because the surprising, out-of-distribution reaction is precisely the thing a language model smooths away. And that surprise is usually the entire reason you did the research.
Synthetic users are a divergence engine, not a verdict. Model the audience to explore cheaply; test the survivors on real humans.
Where they actually belong in your loop#
The mistake isn't using synthetic users. It's using them for the wrong step. Here's the split that keeps you honest:
- Top of the loop — use them freely. Generating message and positioning hypotheses. Screening a dozen value props down to two. Catching an obviously broken pitch before you spend a dollar recruiting anyone. This is where fast-and-cheap beats slow-and-rigorous, and where being directionally right is enough.
- The decision — keep humans on it. Your pricing call. Your final positioning. Anything safety- or trust-sensitive. These are confirmatory, and confirming against the model's own assumptions is circular. Spend your real-research budget here, on the one or two ideas that survived the synthetic pass.
Notice this is the same discipline that just got a compliance startup funded down the hall: put the model on the fuzzy, exploratory, high-variance work, and keep the consequential decision on something you actually trust — a real human here, a deterministic rule there. The model is the cheap, abundant part. The judgment is not.
The concrete founder math#
For a solopreneur, the value is a line item. A proper qualitative user study runs roughly $3–8k and a couple of weeks once you account for recruiting, incentives, and scheduling — a real tax when you're trying to move this week. A synthetic pass gives you a same-day read for pocket change.
So the play is not "synthetic or real." It's synthetic to widen and narrow, real to decide. Run twelve landing-page angles past a synthetic audience on Monday, kill the eight duds, and put your two survivors in front of five actual humans by Friday. You've spent your research budget only on the ideas that earned it — and you haven't quietly outsourced your roadmap to a model's best guess about who your customers are.
That's the whole discipline a $2B valuation is quietly betting most teams will get wrong.
