A 60-person AI company just got valued at $2 billion — nearly six-tenths of NIQ, the market-research giant with 38,800 employees and a $3.4 billion market cap. Simile has been around for barely two years. The gap in headcount could not be starker; the gap in valuation is closing fast.
This is not a vanity round. It is the market re-pricing a $160 billion industry that has spent decades failing to produce a single large winner. Of the four listed global research giants, three have long traded below $4 billion. Only Gartner clears $10 billion — and its revenue mostly comes from IT research subscriptions, not classic consumer research. The sector was considered a structurally bad business. Then AI walked in, and a handful of two-to-three-year-old startups (Simile at $2B, Aaru at $1B, Listen Labs at $500M) are now valued in the same zip code as incumbents that took decades to build.
Where it started: 25 AI villagers
Simile's origin story is a Stanford-Google paper from April 2023. Researchers built Smallville, a pixel-art town, and dropped in 25 AI agents, each given a simple persona — "John Lin runs a pharmacy and has a son in college." No script, no tasks. The agents woke up, ate breakfast, opened shops, chatted with neighbors, built relationships, and spread rumors. Before Valentine's Day, one agent proposed a party; the news rippled through town, and a crowd of AI residents actually showed up. No pre-written plot.
The paper's thesis: give a large model memory, reflection, and planning, and agents exhibit continuous, interacting "social behavior." It now has 7,000+ citations — and its first author, Joon Sung Park, is Simile's founder.
The pivot to business came from a November 2024 follow-up: the team interviewed 1,052 real people for ~2 hours each, built AI agents of them, and asked the agents the General Social Survey. The agents reproduced their human counterparts' answers and behavior patterns with high fidelity. When Fortune 500 CEOs and board members visiting Stanford started asking "can we use this for market research?", Park had his answer.
How Simile works: build people, then simulate
Traditional research has two unsolvable problems. First, people say one thing and do another — a consumer ticks "would buy" on a survey, then flips at the shelf over a $5 price gap or a friend's comment. Second, you cannot simulate knock-on effects: drop prices 10% and competitors respond, consumers start waiting for the next sale — you only see this after launch.
Simile's answer has two steps:
- Build people. Clients feed in authorized interviews, historical surveys, and behavioral data. CVS Health's collaboration is built on ~2.9 million authorized responses from 400,000+ participants across 200+ behavioral scenarios, generating 100,000+ "digital twin consumers" with distinct demographics, incomes, and preferences.
- Run simulations. Drop a new ad, price, or product into the environment; the AI consumers choose, talk, and influence each other, projecting what happens next. CVS internal tests show simulated results matching historical human feedback up to 95% after calibration. And the timeline collapses: what took 4–6 weeks now runs in 15–30 minutes.
The business model is classic B2B subscription — $150K to several million dollars a year — with disclosed customers including CVS Health, Deloitte, and Telstra.
Two routes to the same endgame
The AI research pack splits into two strategies, both attacking the same thing: headcount.
- Replace the respondents (Simile, Aaru). Simile does deep modeling of real individuals; Aaru simulates populations at scale — thousands of agents with varied ages, jobs, incomes, and politics. During the 2024 US election, Aaru ran ~5,000 agents to predict New York's Democratic primary; the result missed the actual vote count by fewer than 371 votes.
- Replace the researchers (Listen Labs). Real respondents stay, but the human work — designing questions, moderating, probing, synthesizing — goes to AI. One researcher used to cap out at a handful of deep interviews a day; AI chats with hundreds simultaneously. Listen Labs passed 1 million real interviews in its first nine months.
Traditional research is labor-stacked: every project means new questionnaires, recruited respondents, interviews, data cleaning, and analyst hours. Revenue growth requires more people; costs never decouple from revenue. NIQ's numbers tell the story: 2023–2025 revenue of $3.34B, $3.97B, and $4.20B, with operating costs eating 88.6%, 84.9%, and 82.4% of it — the improvement driven largely by layoffs and efficiency cuts, not a structural fix.
What actually changed
The moat of the old giants — a global interviewer network, tens of thousands of researchers, decades of consumer panels — is being reclassified from an asset to a cost. When consumers can be simulated directly, the data-collection and analysis pipeline starts as a fixed investment (data licensing, model training, customization) and then runs on electricity and tokens. Marginal cost per simulation keeps falling. Market research is moving from a labor-intensive professional service to a software business that can scale and copy.
That cost-structure shift — not any single customer win — is why capital is assigning Simile and its peers software multiples. If the logic holds, the entire $160 billion industry gets re-underwritten.
What to do about it
- If you commission research: run a parallel pilot. Test one pricing or ad decision through AI simulation against your historical data before trusting it. Calibration with proprietary data is the difference between a toy and a 95% match — bring your own panels.
- If you're an incumbent: your sample libraries and interviewer networks are no longer the moat. The moat is proprietary data access and validation rigor. Both are defensible; neither is automatic.
- If you're an investor: the tell is unit economics, not revenue. Watch whether marginal cost per project actually approaches zero, and whether validation holds as simulations scale to new verticals.
- The caveat: simulation is not reality. Aaru's 371-vote miss in a primary is impressive; a product launch that misfires is not. AI research is a decision-support layer, not a decision replacement — at least until the 100-year track record exists.