Simile, a market research startup with fewer than 60 employees, just raised at a $2 billion valuation. NIQ, one of the five largest research firms on the planet, employs roughly 38,800 people — and trades at about $3.4 billion. Put differently: the entire labor force of a legacy research giant is now priced about the same as a startup that simulates consumers instead of interviewing them.
That's the headline. The underlying signal is bigger: capital is quietly repricing a $160 billion industry (ESOMAR's 2025 estimate) that never produced a proper winner.
The pivot: agents that act like people
The origin story matters. Simile's founder Joon Sung Park is the first author of the Smallville generative-agents paper, the Stanford study where 25 AI agents with one-line personas woke up, ate breakfast, opened shops, gossiped, and spontaneously organized a party — no script, just memory, reflection, and planning. The paper has been cited more than 7,000 times.
The company-forming moment came in late 2024. Park's group ran two-hour interviews with 1,052 real respondents, built AI agents from the transcripts, and had those agents answer General Social Survey questions. The results were convincing enough that visiting Fortune 500 CEOs and board members started asking one question: can you do this for market research?
Simile's answer: feed a system authorized interviews, historical surveys, and behavioral logs; generate 100,000-plus digital twin consumers spanning different ages, incomes, preferences, and habits; then drop new ads, prices, or product concepts into the simulation and let the agents choose, talk to each other, influence each other, and project consequences forward.
CVS Health is the flagship case — built on roughly 2.9 million authorized responses from more than 400,000 participants covering 200+ behavioral scenarios. After calibration with the client's own data, internal tests showed simulated results matching historical real-human feedback up to 95% of the time. And the timeline collapsed from four to six weeks to fifteen to thirty minutes.
The structural shift: people go from asset to cost
Traditional research is a labor business. Every project means redesigning the questionnaire, recruiting respondents, running interviews, cleaning data, and paying analysts to write it up. NIQ's cost base tells the story: operating costs plus period expenses ate 88.6% of revenue in 2023, 84.9% in 2024, and 82.4% in 2025 — the improvement mostly came from layoffs and efficiency cuts, not a better model.
That's why investors never loved this category. Three of the four listed global leaders have long traded below $4 billion: Nielsen went private in 2022 and its NIQ split re-IPOed at roughly $3.3 billion; Ipsos sits around €1.6 billion; Forrester around $200 million. Only Gartner clears $10 billion, and its core is IT research subscriptions, not consumer polling.
Once consumers can be simulated, the thing that defined this industry — a global network of interviewers, thousands of researchers, decades of sample panels — stops being an asset and starts being a cost. Simile's real moat is now the messy authorized data in front of the model plus its calibration loop. After the system is built, each additional simulation run costs roughly electricity and tokens.
It's not just Simile
The same repricing is running across the sector. Aaru, valued at $1 billion, simulated roughly 5,000 agent-voters from demographic data and claims its prediction of New York's 2024 Democratic primary missed the actual result by fewer than 371 votes. Listen Labs, at $500 million, ran more than a million real interviews in its first nine months — then lets AI take over the design, live follow-ups, synthesis, and analysis that used to be an analyst's job.
What to watch
The caveat is real: simulation is calibrated on historical truth, and history doesn't contain the next surprise. A 95% match to past human feedback doesn't guarantee the next launch will behave like its digital twin. The winners here will be the ones holding proprietary, authorized, high-quality data — not the ones with the biggest model.
For buyers of research, the practical move is hybrid: run the simulation for speed and breadth, keep real pilots for the highest-stakes calls, and treat simulated results as a prior, not a verdict. For anyone building in knowledge services, the playbook is now visible: find the labor-intensive layer, ask whether agents can absorb it, and remember that in this transition the people-network moat is what gets repriced first.