From Respondent Pools to Behavioral Ecosystems: What Sampling Looks Like in Five Years

Before I get to the future, I need to say something about the present that I don’t think is controversial, even though some people treat it like it is: the current model is broken. Not “needs tuning.” Broken.

Online sample struggles to reach low-income respondents and many ethnic and racial subgroups. It’s chronically bad at hitting specific geographies at usable size and cost. Any low-incidence audience becomes slow and expensive regardless of panel size. And that’s just the U.S. — every one of these gaps gets wider globally, not narrower. Layer on removal rates that, depending on which benchmark you look at, run somewhere between one in four and six in ten respondents getting cut post-survey (for US B2B studies), and you get a system that is, structurally, not working the way anyone pretends it is at a sales dinner.

I don’t think that’s an alarmist take. I think the alarmist take is pretending it isn’t true.

So — with that as the floor we’re standing on — where does this actually go?

First, a vocabulary problem

“Synthetic” is doing an enormous amount of work in this industry for one word covering at least five genuinely different products: synthetic personas (AI stand-ins for a segment, directional, not individual-level), synthetically-derived insights (aggregated findings pulled from existing data, no respondent-level output), simulated individual-level data (a fully generated dataset structured like a traditional panel), digital twins (an AI replica of one specific, known person), and simulated conversations (AI or human-moderated interviews with synthetic participants).

*Ask yourself, next time someone tells you they have a synthetic offering: which one of these five are they actually describing? Most conversations skip that question entirely, and it’s the only question that matters.*

The money is moving fast regardless of which flavor wins. Aaru, a synthetic research startup, raised a Series A exceeding $50 million in December 2025 at a billion-dollar valuation. That’s not a niche experiment anymore.

The uncomfortable middle: it’s not the vendors driving this

Here’s the part that surprised me most in preparing for this talk: it’s increasingly brands, not research suppliers, pushing synthetic adoption. Companies like Microsoft, Procter & Gamble, and Kroger already sit on enormous first-party and syndicated data — transactional, behavioral, CRM — and AI now makes it feasible to mine that internally for business answers without ever going through a traditional supplier.

Brands have historically never built their own SaaS. They didn’t build competing CRMs or programming platforms. AI plus their own proprietary data might be the first exception to that pattern in decades.

*If that’s right — if the center of gravity for “synthetic” moves inside the brand rather than staying with suppliers — where does that leave the rest of us? Do suppliers still matter for the market and competitive intelligence a brand can’t self-generate, or does that shrink too?*

I don’t have a confident answer. I don’t think anyone does yet, and I’d be suspicious of anyone who claims otherwise.

The thesis I actually believe

Here’s where I land, and I’ll say it plainly because I think it’s the load-bearing idea of this whole talk: **the panel becomes the ground truth. Synthetic becomes the scale layer.**

Not a replacement. A multiplier. Synthetic extends coverage to audiences that are hard to reach with human-only samples, closes attrition-driven gaps, and enables designs that simply aren’t possible otherwise. But a multiplier cuts both ways. Feed a model a panel that’s chronically weak on the populations I mentioned earlier, and synthetic doesn’t fix that gap — it encodes it and reproduces it at scale, with total confidence and no visible seams. Garbage in, garbage out was always the risk. The version we’re building toward is the same problem, just harder to spot once it’s running at scale.

The question isn’t whether a model is fluent. It’s whether it was ever shown the audience you actually need.

That reframes what a panel is for. Not “trust us, we validated these people, here’s their opinion.” More like: we’re building something closer to a behavioral ecosystem — identity and quality verified on an ongoing basis, passive behavioral signal layered on top of stated answers, increasingly activated in real time rather than on a fixed fielding schedule. What are people doing, not just what do they remember doing.

There’s a strange side effect buried in that shift that I didn’t expect going in: the amount of verified data required to make a respondent useful for a serious synthetic product is itself a fraud deterrent. Faking a survey response is trivial. Simulating a coherent purchase history, browsing behavior, and device pattern convincingly enough to hydrate a model is a wildly different, much higher bar — closer to fabricating an entire life history than checking a box. At some point, the effort required to fake it convincingly exceeds the effort it would take to simply be a real, honest respondent.

The question I’m not going to resolve for you

Which brings me to the one I’m deliberately leaving open, because I think sitting with it uncomfortably is more useful than a confident answer: **does the traditional survey even survive?**

Instead of boosting a human sample with synthetic respondents, why not skip straight to asking the model the business question directly? Will a survey dataset look the same in five years, or does it become a continuous input that just feeds a larger model? And if it does — where do sample suppliers actually sit in that world?

I’ll tell you what I don’t believe: I don’t believe humans stop answering questions entirely. What makes us human is precisely our unpredictability — the thousand small, irrational, context-dependent reasons someone chooses one backpack brand over another, and no amount of transactional data fully explains that in advance. But I also don’t think the traditional survey — the fielding schedule, the fixed instrument, the fifteen-minute grid — survives unchanged. Somewhere between “surveys are dead” and “nothing changes” is where I actually think we land, and I genuinely don’t know exactly where on that line.

*So here’s what I’d ask you to sit with, the same way I asked a room of researchers and students to sit with it: if verified human identity becomes the scarce asset everything else gets trained against, is your organization treating it as an asset worth protecting — or still as a checkbox on a vendor scorecard?*

I don’t think that’s a rhetorical question. I think it’s the one worth answering honestly before the next five years answer it for you.