Fake Users Just Hit a $2B Valuation. Should You Test on Them?

On July 30, Simile announced a $200 million Series B at a $2 billion valuation. The round, led by Greenoaks, landed just five months after the startup left stealth with a $100 million Series A from Index Ventures. CVS Health Ventures joined too, and CVS is also a marquee customer.

Simile sells synthetic users: AI-simulated people that companies interview, survey, and test products against instead of recruiting humans. Founder Joon Sung Park built his Stanford PhD around this idea. His famous "Smallville" project put 25 generative agents in a simulated town where they formed relationships, spread invitations, and threw a party nobody scripted. Now investors are paying unicorn prices for the commercial version, and every founder doing customer discovery has a new question to answer.

A synthetic user is a language model prompted to behave like a specific person: a 34-year-old nurse in Ohio who is skeptical of subscriptions, a procurement manager who has been burned by bad software. You ask it questions. It answers in character, at any hour, in any volume.

Vendors build panels of thousands of these personas, tuned with survey data and behavioral research, then sell access for marketing and product research. TechCrunch described the category as something akin to vibe coding for product mock-ups, and that comparison cuts both ways. Fast and cheap first drafts, with all the reliability questions a first draft implies.

The money is chasing a real cost problem. Recruiting 20 humans for interviews takes weeks and thousands of dollars. Panels get slower and pricier every year, and response rates keep sliding. A synthetic panel answers in minutes for a fraction of the cost.

Simile is not alone. Aaru raised a Series A in December at a headline $1 billion valuation for AI-simulated research. When two startups in the same young category command ten-figure valuations within months of each other, the venture market is declaring that a chunk of the market research industry, an industry worth tens of billions annually, is up for grabs. Whether the tools deserve that confidence is a separate question, and it is the one that matters to you.

There is also a tell in who invested. CVS Health Ventures put money into a vendor it already buys from, which usually signals the product solves a real internal pain, in this case the grind of testing marketing and product ideas at enterprise scale. Enterprise pain and founder pain are cousins, not twins. What a Fortune 50 research team does with synthetic panels is not automatically what you should do with them at the idea stage.

Used well, simulated respondents are a sharpening tool. They shine at the steps before real customers get involved.

They stress-test your interview script, so you find the confusing question before wasting a real conversation on it. They generate objections you had not considered, which is worth a lot when you have never sold anything before. They let you rehearse a pitch cheaply and compare five different ways of framing the same feature. And they are useful for exploring segments you cannot easily reach yet, the way a map is useful before you visit a city. None of that replaces meeting the city. All of it makes the visit more productive.

A concrete example. Say you are building scheduling software for dog groomers. Before cold-calling thirty groomers, you run your interview script past a synthetic groomer persona and discover your third question assumes they all use booking software already. Half do not. You fix the question, and your real interviews start smarter. Twenty minutes of simulation just improved thirty conversations. That is the honest ceiling of the tool, and it is a useful ceiling.

Simile's stated mission is to simulate all eight billion people on earth with full accuracy. TechCrunch called that goal preposterous, and the reasoning is worth internalizing: the entire reason market research exists is that humans are unpredictable, moved by emotion as much as logic.

Models are trained on what people say, mostly online, and people misreport their own behavior constantly. Synthetic users inherit that gap without the correcting force of a real wallet. They also tend toward agreeableness. Ask a simulated customer if they would pay $29 a month and you will hear yes far more often than the market will ever say it. For a first-time founder, that failure mode is dangerous precisely because it feels like validation. You can accumulate a folder of enthusiastic fake evidence and mistake it for demand.

Here is a sequence that takes the speed of synthetic research without inheriting its blind spots.

First, draft your assumptions: who has the problem, how painful it is, what they do today. Second, run synthetic interviews to pressure-test the script and surface objections. Treat every output as a hypothesis, never a fact. Third, talk to at least ten real humans in your target market before building anything. Their contradictions are the data. Fourth, write down what the synthetic panel got wrong. That list tells you exactly how much to trust it next round.

Keep the whole trail in one place: questions asked, assumptions made, what real customers contradicted. A spreadsheet works. So does a structured planning tool like Foundra, which keeps your customer research sitting next to your business plan so the evidence and the strategy stay connected.

Enterprise synthetic-research contracts are priced for CVS, not for you. But the underlying technique is nearly free. A frontier model subscription costs about $20 a month, and careful prompting gets you a serviceable synthetic panel for script testing.

Compare that with what real conversations cost: mostly time and mild social discomfort. Ten customer conversations might take two weeks of persistent outreach. That is the single best two weeks you can spend before writing code, and no budget line item competes with it. The honest budget for early validation is $20 in software and 20 hours of talking to strangers. Founders who skip the second half because the first half felt like progress are the ones who ship products nobody wants.

Run any synthetic finding through this filter before it touches a decision.

• The simulated users agreed with almost everything. Real segments push back.

• The insight cannot be traced to any real-world source, survey, or behavior you can verify.

• You are using the output to justify skipping human interviews rather than to prepare for them.

• The finding involves willingness to pay. Simulated wallets open easily. Real ones do not.

• You feel relief instead of surprise. Good research changes your mind at least once. If nothing you learned was uncomfortable, you probably learned nothing.

• Simile hit a $2 billion valuation on July 30, five months after emerging from stealth, and Aaru raised at a $1 billion headline figure in December. Synthetic research is now a funded category.

• Simulated users are excellent for rehearsal: sharpening scripts, surfacing objections, testing framings.

• They are unreliable for validation, especially pricing, because they inherit self-report bias and agreeableness.

• The winning process is synthetic first drafts followed by at least ten real conversations, with a written record of where the simulation missed.

• Your cost of entry is about $20 and some nerve.

Can I use synthetic users instead of customer interviews? No. Use them before interviews, to prepare, and after, to explore variations. The decision-grade evidence still comes from humans with real constraints and real money.

Why did Simile raise so fast? Growth-stage investors are paying for category leadership in AI markets early. A $200 million round five months after a $100 million round says more about venture competition than about settled product truth.

Are synthetic panels at least good for surveys? They approximate broad directional patterns better than individual decisions. Treat results as a cheap pilot that tells you which questions deserve a real sample.

What tools do I need to try this? Any frontier chat model works for informal synthetic interviews. Write detailed personas, instruct the model to be disagreeable, and vary the framing across runs.

How many real interviews are enough? Patterns usually stabilize somewhere past ten conversations within one segment. If every answer still surprises you, keep going. Consistency is the signal to stop, not fatigue.

1. Synthetic-user startup Simile raises $200M at $2B valuation 5 months after $100M Series A (TechCrunch, Jul 30, 2026)
2. Simile: Series B announcement
3. Stanford HAI: Computational agents exhibit believable humanlike behavior (Smallville)
4. AI synthetic research startup Aaru raised a Series A at a $1B headline valuation (TechCrunch, Dec 2025)