What the independent research says
The single source we point people to most is a research brief from NORC at the University of Chicago, a long-standing social-science research organization. NORC puts survey-fraud rates at roughly 15 to 30 percent across the industry, and up to about 45 percent on some platforms. On a typical online panel, that means somewhere between one in seven and one in three responses may not be from a genuine, unique respondent, and on the worst platforms it can approach half. That is not an outlier reading. The Insights Association estimated that around 40 percent of all nonprobability survey interviews in 2025 were fraudulent, on the order of two billion responses. Kantar, one of the largest research firms in the world, went as far as to call panel fraud “the new ad fraud,” and has found researchers discarding up to 38 percent of collected data on average over quality problems. Greenbook has written about the pervasive threat of tech-enabled fraud in survey research. Read together, these sources describe roughly the same range, from about a third to as much as half of responses on some platforms.These are industry figures from NORC, the Insights Association, Kantar, a Dartmouth study in PNAS, and Greenbook, not VerifyYou measurements and not a claim about your specific account. They describe how much junk is typically sitting in online data, so you can judge whether the problem is worth solving for you.
Why it is getting worse, not better
A few years ago, faking a survey response at scale took real effort. Generative AI changed the economics. A large language model can now produce fluent, human-sounding answers in bulk, and a bot farm can spin up convincing synthetic respondents cheaply. The cost to fake a person has fallen, and when the cost of an attack falls, the volume rises. A peer-reviewed Dartmouth study in PNAS found that AI agents can now pass standard quality checks at near-perfect rates, 99.8 percent in one proof-of-concept study, which is why the trap questions many teams rely on are quietly losing their grip. This is why catching fakes after the fact is such a hard game. Passive detection, the kind that watches behavior and tries to spot a bot from the outside, is a cat-and-mouse loop. Defenders learn the latest tell, attackers learn the new defense, and the synthetic responses look a little more real each round. You can win rounds, but you do not get to stop playing.What it costs you
The figures stay abstract until you turn them into line items. When a meaningful share of your data is not from real, unique people, the cost shows up in three places:- Wasted incentives and spend. Every payout paid to a bot or a duplicate is money spent on traffic that was never going to become a real respondent. The fake portion is pure waste.
- Contaminated datasets. Fraudulent and duplicated responses do not just miss value, they poison what is left. A decision drawn from a panel that is a third synthetic is drawn partly from noise.
- Decisions made on data that is not from real people. This is the quiet one. If you set strategy or product direction off survey results, and a large slice of that data came from bots and survey farms, you are steering by a compass that has been tampered with.