In April 2025, eight people across two customer research companies in the United States were indicted for conspiring to bill clients up to $10 million for fraudulent survey data. They didn't hack systems or forge documents. They filled surveys with fake respondents, then asked their clients to pay for it.
A few months later, a study published in the Proceedings of the National Academy of Sciences showed how large language models can evade every major detection method built to catch them, completing online surveys at roughly $0.05 each with a 99.8% success rate across 43,800 evaluations. With the average survey incentive sitting between $1 and $2, that's a profit margin above 95% for anyone willing to deploy fake respondents at scale.
Two serious signals. One structural problem, more than two decades in the making. Before looking at what panel quality actually requires in 2026, it's worth understanding how the market research industry got here.
Market research, past and present
Two decades ago, sampling in paid market research was tightly controlled within double opt-in online panels. It wasn't perfect, affordable, or anywhere close to today's smartphone-powered scale, but respondents were recruited carefully, profiled in depth, and treated as a scarce resource. Customer research companies invested in their respondents because they understood that respondent quality was the product.
Ten years later, the programmatic era disrupted that model. Sample exchanges let buyers pull respondents from dozens of sources simultaneously, optimizing for speed and cost. Efficiency went through the roof. Turnaround times and prices fell off a cliff.
What also dropped was accountability. When a respondent passes through three routers before reaching a survey, which party in the chain is actually responsible for quality? The panel company points at the exchange. The exchange points at the supplier. The supplier points at the router. All that fingerpointing had an economic impact too: cost-per-interview rates compressed to the point where serious investment in respondent quality became irrational. For many operators, it turned into a race to the bottom.
The research industry has been measuring volume and calling it quality for years. Those are not the same thing.
A generation of consolidation played its part too. Many of today's largest panel companies are assemblages of ten or fifteen acquisitions, each bringing new respondent pools, new technology stacks, new processes, and new opportunities for quality to fall through the gaps. Companies built through merger after merger naturally valued scale over the careful management of individual respondents.
By the 2020s, inattentive, unrepresentative, and fraudulent respondents were just part of the package. The industry had split two things that were never meant to be apart, like an atom: speed and volume on one side, quality on the other. We're dealing with the fallout now.
The threat of AI
The industry's traditional defenses were built for a different kind of threat: unsophisticated operations like click farms and duplicate accounts. VPN checks, IP blacklists, and basic digital fingerprinting were designed to catch them. Those threats still exist, but they're being dwarfed by something the old tools were never built to handle: AI-powered fraud.
Generative AI models can maintain coherent demographic personas, generate plausible open-ended responses, pass attention checks, avoid trap questions, and complete surveys at human-like speeds. The PNAS research tested this across nine different language models and found that AI agents evaded detection across every method the research industry has at its disposal. The tools built to catch bots simply don't catch these bots.
An independent benchmark published by the ACFE in early 2026 tested five leading fraud-detection platforms against identical respondent pools across consumer, B2B, and healthcare audiences. Block rates differed by up to 35 percentage points between vendors analyzing the exact same respondents. The platforms frequently agreed on roughly how many respondents to exclude, but rarely on which specific ones deserved the boot. One platform would flag a respondent for device spoofing; another would pass the same respondent as valid. Fraud, it turns out, is not a universally defined category. No single tool catches everything.
That has a clear implication for our industry. A research company that deploys one fraud-detection tool and considers the problem solved isn't protected at all, it has a single layer of defense against a multi-vector attack. The greatest accuracy occurs where multiple systems converge on the same failure decisions, at the pressure points where layered coverage produces overlap that no individual tool can replicate alone.
AI is being deployed on the detection side too, and that's genuinely promising. Machine learning models can identify behavioral signatures that rule-based systems miss: micro-timing patterns, response trajectories across a full survey, coherence between answers given at different points. AI-assisted review of open-ended responses can flag text that's statistically unlikely to have been written by a human under normal survey conditions. These are valuable measures, but they're caught in an arms race with the fraud methods they're built to catch, one that simply doesn't end.
The root cause nobody wants to fix
The industry conversation about data quality focuses overwhelmingly on fraud detection, on building better walls to keep fraud out. Necessary, but not sufficient.
The deeper problem is the supply of genuine, engaged human respondents. Fraud doesn't exist in a vacuum, it fills a gap, and that gap exists because honest people have largely stopped wanting to take surveys.
The economics of panel participation, from the respondent's side, are poor. Screening rates are high (a respondent may attempt ten surveys to complete one). Surveys are long, repetitive, often poorly designed. Compensation is low relative to the time required. Being routed from platform to platform before qualifying, or not qualifying, is frustrating enough that many ordinary people just opt out.
What's left is a self-selected pool: people highly motivated by the incentives, people who've learned to game the system, or automated agents. None of these groups produce high-quality data.
It's a vicious cycle the industry has declined to address directly. A bad experience drives away honest respondents. Their departure creates a gap. Fraudsters fill the gap. Quality declines. Researchers add more detection layers and more screener questions. The experience gets worse. More honest respondents leave. The cycle accelerates.
The answer can't be only better fraud detection. It has to involve treating respondents as people whose time has value, whose experience matters, and whose participation is worth investing in. In practice, that looks like:
- Fair compensation, including for screenouts and quota fulls, not only for completed interviews.
- Responsive support when something goes wrong.
- Positive survey experiences, carefully designed not to exhaust respondents' patience.
- Feedback mechanisms that actually close the loop.
It's not pure empathy. Engaged respondents who feel respected don't rush. They don't drop out mid-survey. They give thoughtful answers. They come back. The respondent experience isn't a nice-to-have sitting alongside the data quality program, it's inherent, and it dictates outcomes far more than the industry admits. Panel health and respondent experience aren't adjacent to data quality. They're the foundation it's built on.
What panel quality actually requires in 2026
Effective panel quality is a daily practice built across three stages. Here's how we approach it at Make Opinion.
Before the survey
Exclusion and verification. Automated checks on VPNs, proxies, and suspicious IP patterns block known bad actors before they enter. Digital fingerprinting measures behavior across a respondent's lifetime in the panel, catching duplication and flagging deviations. Screener rotation disrupts path cloning, and survey links are actively rotated to stop fraudster communities from sharing entry points.
During the survey
Behavioral monitoring and consistency checking. Timing analysis distinguishes genuine engagement from speed-running, examining the core survey separately from demographics since fraudsters idle on the final page to make total time look normal. Answer consistency checks and AI-assisted text review catch the rest.
After the survey
Reconciliation and feedback. Post-survey experience ratings provide a signal no fraud-detection tool can replicate, surfacing both experience issues and manipulative patterns. As researchers ourselves, we also manually review flagged cases for template-like responses automated systems weren't trained to catch.
The ACFE's benchmarking findings underline why this layered approach isn't optional. The greatest accuracy occurs where multiple systems converge; a single tool leaves exploitable gaps. Multiple overlapping layers, each approaching the problem from a different angle, produce a defense that's genuinely difficult for fraudsters to penetrate.
So, back to the core question: why can't quality simply be a feature for online market research platforms? Because none of this is a solved problem. Fraud evolves, and our tools have to evolve with it. That's why quality is a commitment rather than a feature, one that requires ongoing operational and multi-layered investment in staying ahead, or at least keeping pace.
The combination of proactive support and a clean panel is the only way I've seen to guarantee quality while saving budget and hours of post-field work.
Back to basics
There is no panel company in the world that has solved data quality. Anyone who claims otherwise is either mistaken or trying to sell you something. The threat landscape is evolving faster than any single organization can track, and AI-enabled fraud will keep getting better. New vectors will keep emerging.
It's a moment to get back to basics. What distinguishes serious operators from non-serious ones isn't the absence of fraud in their data. It's the daily choice to invest in the people and processes required to minimize it, to close the feedback loop when things go wrong, and to treat the respondents whose answers underpin every finding with the respect their participation deserves.
Quality is not a feature you switch on, a certification you hang on the wall, or a tool you deploy and forget about. It's a commitment and a practice, built from the respondent up. And the best part? It's visible in the data your company generates.