Key Takeaways
- ISO published the fourth edition of 20252, the international market research standard, in September 2026, adding clause 4.1.5 on the use of AI. Compliance is mandatory from September 2029
- Every requirement in that clause governs the researcher's own tools: which AI, which version, which stage, told to the client, disclosed to participants
- Conveo raised a $50 million Series A on September 2, 2026 to run consumer interviews with an AI moderator, bringing its total to $55.8 million
- A NORC at the University of Chicago review published in April 2026 puts industry fraud rates at 15 to 30%, reaching 45% on some platforms
- The same review describes a synthetic respondent that passed 99.8% of standard data quality checks, and a study where 34% of active survey participants said they used an LLM to help write their open-ended answers
Edu here. Two things landed in market research this month, about three weeks apart (different outlets, no overlap in the coverage), and both were written up as the industry finally growing up about AI.
On September 2, a company called Conveo raised $50 million to run consumer interviews with an AI moderator instead of a human one. Later in the month ISO published the fourth edition of 20252, the international standard the serious agencies certify against, and for the first time it contains a clause about using AI in research.
I read the coverage of both waiting for the other half of the sentence. Both are about the researcher's side of the microphone. The AI that has actually broken survey research is sitting on the other side, typing the answers.
ISO 20252:2026, published in September 2026, introduced clause 4.1.5 covering the development and use of AI in market research. Every requirement in it concerns the provider's own tools: the type of AI, the version, the research stage, disclosure to the client and to participants. None of it addresses whether the respondent supplying the data is a human being.
What the standard actually asks of you
ISO 20252 is the quality standard for market, opinion and social research, run by technical committee TC 225 with ESOMAR in liaison. It first appeared in 2006, was revised in 2012, restructured in 2019, and the fourth edition landed this September. The Market Research Society puts the transition deadline at September 2029, though most certified agencies will move across at their next recertification rather than waiting.
I should say upfront that I have not read the document. You have to buy a copy, I am not certified to anything, and a few dozen pages of clause numbering was not going to happen this week. So what follows is assembled from the ISO catalogue entry, the MRS summary and a certification consultancy's clause-by-clause writeup (which means I am reading other people's reading, and if a clause reference below is off by a digit, that is why).
The AI clause is 4.1.5, "development and use of AI and related technologies". Here is what it asks a provider to do, and who each requirement is about.
| Requirement in clause 4.1.5 | Whose behaviour it governs |
|---|---|
| Describe the type of AI used and the research stage it was applied in | The researcher |
| Record and tell the client when AI generated a research insight | The researcher |
| Get explicit client consent before passing project data through AI | The researcher |
| Document the version of the AI used | The researcher |
| Inform participants when they are interacting with AI | The researcher |
| Run monitoring procedures for security, privacy and quality | The researcher |
I want to be fair about this, because it's a genuinely useful clause and it's overdue. If you commission a study in 2030 from a certified agency (assuming they are certified at all, and plenty of good shops are not), you will know whether a model wrote the analysis, which version, and at which step. That is more than you can find out today from most vendors. The old edition had nothing to say on any of it.
The 2026 edition also splits the old Annex A, moving panel development and lifecycle management into a new Annex G, and strengthens the data quality requirements around sampling. So the panel side was not ignored. It just sits in the annexes, where the AI clause does not reach.
Fifty million dollars for a faster interviewer
Conveo's round closed on September 2, led into by DST Global Partners, Balderton Capital, Visionaries, 6 Degrees Capital and Y Combinator, taking the company to $55.8 million raised. I think that is the biggest round research tooling has seen this year, though I did not check every deal. The product runs video conversations with consumers where an AI interviewer adapts its follow-up questions to what the person just said, then pools the findings across studies into something a marketing team can search in plain language. The company says more than 400 enterprises use it, over 50 of them in the Fortune 500, and names Google, Unilever, AB InBev and Canva.
The pitch is not silly. Surveys are shallow, focus groups are slow and expensive past a certain scale, and one reported customer is running something like 2,000 interviews a month across seven markets, which no human moderation team is doing (not at that price, anyway).
The counterweight is thinner than the funding announcement suggests. Implicator.ai could not independently verify the 400-enterprise and 50-Fortune-500 figures, noting an industry directory listing more than 45 enterprise brands instead. It also points at a Mission Field comparison across 44 interviews where the AI could not read body language, facial expression, boredom or hesitation, and weighted outlying answers too heavily. Human moderators drew out more reflective and emotionally expressive responses. That comparison was small and was not peer reviewed (44 interviews is not a lot to hang a verdict on), so treat it as the only evidence available rather than as settled. Maybe a bigger comparison lands the other way.
The problem is on the other side of the microphone
In April 2026, NORC at the University of Chicago published a literature review on fraudulent respondents and bots in nonprobability surveys. It runs to six pages (the reference list alone is about eighty entries), and it is the most sobering thing I have read about research this year.
The headline estimate is that fraud runs at 15 to 30% across the market research industry, reaching as high as 45% on some survey platforms. Those are the averages. The individual published cases are worse.
Goodrich and colleagues ran a survey of US beekeepers in 2023 and found 4% of 2,622 responses were legitimate. Krawczyk and Siek collected 981 responses in four hours in 2024, of which three were authentic. Imes and colleagues reported that over 95% of responses recruited through Twitter were bot-generated. Pozzar and colleagues put a health study at 94.5% fraudulent back in 2020, before any of this got easy.
Those are the extreme end (surveys that got audited hard enough to find out), and NORC presents them as documented failures rather than as an average. Still, four of them in one literature review is a pattern.
Two findings in that review are the ones I keep coming back to. The first is Westwood, in 2025, building an autonomous LLM-based synthetic respondent that passed 99.8% of standard data quality checks while producing coherent, persona-consistent answers that could not be told apart from human ones. The second is that the open-ended question stopped working. Everyone relied on that filter, because a bot could not write a plausible sentence about why it switched shampoo. Yee reported in 2025 that traditional strategies like domain-knowledge checks and open-ended questions have become ineffective against LLM-powered bots.
And then there is the finding that is not about fraud at all.
In one study of active online survey participants, 34% said they used an LLM to help them answer open-ended questions, mostly to express themselves better. Nobody is cheating there, or not obviously. They are people, they are eligible, they are giving you their real opinion, and the words you are about to code as customer language were written by a model (partly, in most cases, but you cannot see the seam). The review notes that LLM-generated responses tend to be more homogeneous and to gravitate toward common cultural touchstones, which suppresses exactly the natural variation you were paying for.
So what do you ask now
If you commission research, the disclosure question is about to answer itself. From 2029 a certified provider has to tell you what AI it used and where. Good. Ask the other one instead: how did you establish that these respondents were people, what share of collected responses did you throw out, and were those rules written down before fieldwork started?
NORC's own answer (not mine, and probably the better one if you do quantitative work) is to move toward probability-based panels: verified contact details, real sampling frames, PIN-based entry, established recruitment. That does not eliminate fraud, but it removes most of the doors an automated attack walks through. It also costs considerably more than buying completes from an open panel, and open panels exist for exactly that reason.
My answer is different because my work is different. I read what people said when nobody asked them. There is no incentive to collect, no screener to beat, no $2 payout waiting at the end. Nobody writes a 400-word complaint about their magnesium at one in the morning to farm a gift card. That does not make a Reddit corpus clean (planted marketing content is real, I wrote about it two weeks ago, and I still cannot tell you how much of it I have coded across twenty studies). It makes it broken in a different direction, and the two failure modes probably do not overlap much, so they are useful to hold against each other.
The thing I genuinely cannot resolve is that 34% number. Some of that group is a real person with a real opinion using a model to say it more clearly (not a data quality problem at all, maybe even the opposite). Some of it is a model answering on their behalf. Nobody has separated the two, and I have no idea how you would design a study that could.
Nobody fakes a complaint they were never paid to write.
Insightios reads the threads where your market argues with itself, codes what repeats across unrelated communities, and hands back the objections, phrasings and buying triggers ranked by how common they are, with the real quotes attached. Flat fee, fixed turnaround.
Frequently asked questions
What does ISO 20252:2026 say about AI?
The fourth edition, published in September 2026, adds clause 4.1.5 on the development and use of AI and related technologies. It requires a research provider to describe the type of AI used and the stage it was applied in, record and tell the client when AI generated an insight, get client consent before passing project data through AI, document versions, inform participants who interact with AI, and monitor for security, privacy and quality.
When do research agencies have to comply with ISO 20252:2026?
There is a three-year transition from the 2019 edition, so compliance becomes mandatory in September 2029 according to the Market Research Society. Most certified organisations will move across earlier, at their next recertification audit, rather than waiting for the deadline.
How much online survey data is fraudulent?
A NORC at the University of Chicago literature review published in April 2026 cites industry estimates of 15 to 30%, reaching as high as 45% on some survey platforms. Individual published studies report far worse: one US beekeeping survey found only 4% of 2,622 responses were legitimate, and a social-media-recruited health study found 94.5% fraudulent.
Can AI bots pass survey quality checks?
Yes. The NORC review describes work by Westwood (2025) building an autonomous LLM-based synthetic respondent that passed 99.8% of standard data quality checks and produced coherent, persona-consistent answers indistinguishable from human ones. Open-ended questions, long used as the practical bot filter, no longer reliably work against LLM-powered responses.
Are AI-moderated interviews as good as human ones?
The published comparisons are thin. One small, non-peer-reviewed comparison across 44 interviews found the AI moderator could not read body language, facial expression, boredom or hesitation, weighted outlying responses too heavily, and that human moderators drew out more reflective and emotionally expressive answers. Speed and cost are the clearer wins.
What should you ask a research agency about data quality?
Ask how they established the respondents were real people, what share of collected responses they discarded, and what the screening rules were before fieldwork began. From September 2029 an ISO 20252 certified provider must also disclose its own AI use, so the disclosure question answers itself. The respondent-identity question does not.
What to do with this
If you are a DTC brand and you have bought panel research in the last two years, go and find the fieldwork notes (the appendix nobody opens). Not the deck. The part that says how many responses came in and how many survived screening. If that number is not in there, ask for it, and notice how long it takes to arrive.
If you are about to buy AI-moderated interviews, the speed is real and the scale is real. Just be clear with yourself that you are buying a cheaper way to ask, and that asking was already the weak instrument. I am not sure speed is the thing that was missing. A faster version of a method with a known bias gives you the same bias sooner.
And keep at least one read of your category that nobody was recruited for. Support tickets, reviews, the subreddit where your customers complain about you by name. If you want the method rather than the service, the guide to using Reddit for DTC research covers how I gather and code a corpus, and why surveys fail DTC brands covers the part of this argument that predates AI entirely.
All in all, I think September was a good month for market research governance and a bad month for anyone who assumed the governance was about the data. A standard can make you disclose your own tools. It cannot make the person on the other end exist.
Sources
- International Organization for Standardization. (2026). ISO 20252:2026, Market, opinion and social research, including insights and data analytics, Vocabulary and service requirements. Fourth edition, published September 2026, ISO/TC 225. Link Retrieved September 22, 2026.
- Market Research Society. Quality standards, ISO 20252. Confirms the three-year transition from the 2019 edition, mandatory compliance by September 2029, the split of Annex A and the new Annex G for panel management. Link Retrieved September 22, 2026.
- Assent Risk Management. (2026). ISO 20252:2026 has been published, here's what you need to know. Clause-by-clause summary, including the six requirements of the new clause 4.1.5 on AI. Link Retrieved September 22, 2026.
- NORC at the University of Chicago. (2026). Fraudulent respondents and bots in nonprobability surveys: A literature review. Research Brief, April 2026. Source for the 15 to 30% industry fraud estimate, the Westwood 99.8% figure, the 34% LLM-assisted responding figure, and the Goodrich, Krawczyk and Siek, Imes and Pozzar case studies. Link Retrieved September 22, 2026.
- Tech.eu. (2026). Conveo raises $50M to scale its AI-powered consumer intelligence platform, September 2, 2026. Source for the round size, investors, $55.8 million total, and the customer figures as reported by the company. Link Retrieved September 22, 2026.
- Implicator.ai. (2026). Conveo lands $50 million for always-on consumer intelligence. Source for the unverified customer counts, the 2,000 interviews a month figure, and the Mission Field comparison across 44 interviews. Link Retrieved September 22, 2026.
- Insights Association. Work underway to update ISO 20252 market research standard. Confirms ISO/TC 225 targeted AI and data fraud prevention in the revision, with ten countries and ESOMAR participating. Link Retrieved September 22, 2026.