The buying metric is changing
For years the industry has bought on cost per complete. That is becoming a weak number, because a complete tells you a response arrived, not that a real person left it. The figure that matters now is cost per quality response, what you actually pay once the bad data is removed, and on that measure the picture is stark.~40%
of nonprobability survey interviews in 2025 were fraudulent, on the order of two billion responses.
Insights Association
99.8%
of standard survey quality checks passed by AI agents in one proof-of-concept study.
Westwood, PNAS
51%
of all internet traffic is now automated, surpassing human traffic for the first time.
Imperva Bad Bot Report, 2025
These are industry figures from third parties, the Insights Association, a peer-reviewed PNAS study out of Dartmouth, Imperva, Kantar, and Research Defender among them, not VerifyYou measurements and not a claim about any specific account. The full citations are in the PDF above.
A large industry, taking on water at the source
Market research is a $142 billion global industry, and its core promise is simple: gather authentic human perspectives so organizations can make better decisions. The threat to that promise used to be methodological. It has been overtaken by something more literal. A growing share of online respondents are not who they claim to be, and an increasing number of them are not human at all. The U.S. Department of Justice has indicted operators who ran a survey fraud ring for a decade, billing clients ten million dollars in fabricated data, and Research Defender has flagged roughly a third of raw responses across hundreds of sample sources as fraudulent.As capital flows into AI, human data becomes scarce
There is a funnel underneath all of this, and it runs one way. Investment in AI drives model capability, capability drives scale and reach, and scale fills the internet and the panels with non-human content. That is what makes verified human data the scarcest, most valuable input in the insights economy, and the market has started to price it. The AI training data market, worth about 11 billion by 2030, and human-written content already costs several times its AI-generated equivalent. When the honest share of the world’s data shrinks, the premium on the honest share rises to match.Three layers of cost, each more expensive than the last
The damage from survey fraud reaches well past the wasted incentive. It operates on three levels, and each one reprices what a study is worth. The first is operational: when a third of responses are fraudulent and traditional cleaning catches only a fraction, teams over-recruit, re-field, and burn labor chasing bad records. The second is decisional: contaminated data that survives cleaning shapes strategy, and a decision drawn from a poisoned set is drawn partly from noise. The third is compounding: every study that ships on bad data erodes confidence in the method itself. Kantar’s own tracking shows the share of collected data discarded for quality problems climbing toward the mid-forties in percent, and once you price fraud back in, cost per quality response can run many times the headline cost per complete.Detection has a ceiling, verification does not
The industry answered the crisis with an expanding toolkit: device fingerprinting, behavioral analysis, trap questions, CAPTCHAs, and post-collection statistical cleaning. Each catches some fraud. None catches enough. Research Defender’s data suggests the large majority of survey fraud now evades traditional cleaning outright, and detection runs on a losing asymmetry: defenders have to catch every form of fraud, while an attacker only has to find one gap. Every technique you add is a target the other side learns to beat. The deeper point is about the question being asked. Detection answers one thing, is this session automated. Verification answers a different one, is this person real, unique, and present. The first question has a ceiling. The second does not.A different model, with different economics
Detection accepts a contaminated input and tries to filter it afterward, which is reactive, expensive, and incomplete by design. Verification moves the question to the door and confirms a real, unique, live human before the first question loads. The two models have genuinely different economics.
The cost of detection scales with the size of the problem, and the problem is growing. The cost of verification is fixed per respondent. It also solves something detection ignores entirely, which is respondent friction: one verified credential means one fewer login, one fewer onboarding, one fewer reason to abandon a study.
”With tools like VerifyYou, AI becomes a force for good. It helps prevent AI-driven fraud and can be done with much lower friction than what we did over the years.”
President and Founder of Rare Patient Voice, now part of Konovo