Marketers spent a decade learning to verify every ad impression, but Jaxon Thomas, managing partner at Ekas, argues the research behind their biggest decisions now needs exactly the same scrutiny, before bad data quietly spends the budget for them.
A decade ago marketers faced a reckoning. Digital media overtook offline in spend, which attracted organised criminals and fraudsters because security and verification were still fairly nascent or going unused. There was easy money to be made with an almost victimless form of fraud.
The industry faced something of a dilemma – the bots got so bad there were billions of dollars disappearing into the ether every year. But, as ever, we got wise and found ways to ensure the inventory we were buying was real, with viewability standards, traffic verification and supply-chain scrutiny becoming mandatory.
The buyers who took it seriously pulled ahead as their spend worked because it was doing its job and being seen by real humans.
The research that sits underneath your biggest decisions is going through a similar reckoning right now, but most marketers may not have clocked it yet and it may be compromising the data you base your decisions on.
Segmentation, brand tracking, message testing, the innovation pipeline. Research has underwritten marketing’s largest bets for decades and in higher-stakes categories like healthcare, it shapes decisions where getting the audience wrong carries a cost well beyond a misfired campaign. You trusted what landed in the deck to be parsed, informed and human-centred because it really could only have been created by people.
That trust now needs the same scepticism that was aimed at your media spend, because the fraudsters are getting into the research and they are more sophisticated than ever before, with AI bots able to pose as real people and infiltrate research panels with alarming ease.
Analysis suggests the share of usable responses in online surveys has fallen from roughly 75 per cent to 10 per cent in recent years, driven by fraud. [2024 review, Frontiers in Research Metrics and Analytics]
An autonomous AI agent evaded state-of-the-art detection almost flawlessly, passing 99.8 per cent of standard attention checks while mimicking a real respondent down to the typos. [2025 Dartmouth study, PNAS]
Yes, the checks most research has leant on were beaten by this agent 998 times out of 1,000.
Before you panic, let’s take a moment to understand what this all means and how the industry is in fact responding to this threat.
The main responsibility in this area sits with the panel provider companies which have thousands of respondents across the region answering questions on all kinds of topics.
I’ve spent a career around research panels, and Ekas has been operating for more than 50 years across both consumer and healthcare research. In that time, we’ve seen ‘enterprising’ people try every trick going to game the system – mostly for money, occasionally just for the sport of it.
For most of that time the scams were clumsy and small because the systems were, by today’s standards, primitive. These were the days when the telephone was used to actually speak to people rather than message them. Therefore fraud was easier to detect and mitigate.
On the whole, it still is, as long as the right processes are in place. I think of it like the security operation at a good venue, because it works in much the same way.
First, it’s about controlling who comes through the door, and that’s where the smartest technology now sits. A photo-ID check to confirm a real face, digital fingerprinting that recognises devices caught misbehaving before, and fraud scoring that weighs dozens of signals in a heartbeat, all run before a respondent reaches the first question.
In more sensitive areas like healthcare research the bar is higher still, because you’re not only confirming a real person but the right kind of person, a verified specialist rather than someone claiming to be one, which is its own discipline.
It’s the fraudster’s own AI turned back on them at the entrance. Get that right and most of the problem never touches your data.
What slips past the door gets caught by the security watching the floor, the checks for suspiciously fast or nonsense answers, and the rescreening that confirms people are who they claimed to be.
AI can do a lot of this work too, flagging the patterns across thousands of responses that a person would miss and pointing a researcher at the few that warrant a second look, which gives the experts more time and headspace for the calls that genuinely need a human.
And the controls are measurable.
We are blocking around 40 per cent more survey attempts than before ChatGPT launched, with removal rates a fraction of those seen in open consumer research. [Ekas, M3MR group data, 2025]
None of this is an argument against AI in research because it is going to be an integral tool for us all going forward. It’s an argument for knowing the difference between the synthetic data you invited in and the kind that slipped in because security was on a break (that’s one big positive of AI, it never rests or loses concentration).
So the next time a tracker or a segmentation lands on your desk which demands your attention, make sure you treat it the way you’d treat a media buy you couldn’t see. Ask your research partner how they control who gets into the panel, what they catch behind that, and who reviews what gets flagged.
If the answer is a shrug or not well explained, you have a right to be sceptical. If you hear a version of the above, you can be sure they’re on the cutting edge of the fight against this kind of fraud.
Bad data doesn’t announce itself, it sits inside a perfectly clean-looking chart, an innocent spreadsheet and sends you confidently in the wrong direction.
But, you won’t know until the campaign it shaped has already spent its budget finding out. With the right due diligence, it doesn’t have to be that way.
