The Buyer Who Doesn't Exist: Why B2B Teams Are Surveying AI Copies of Their Customers — and the Evidence Trap Hiding in the Answers

Imagine you could survey five hundred CISOs by lunch. Not a panel vendor's promise of "fielding in three weeks," not a $40,000 invoice for a study that trickles back with a 4% response rate. Five hundred security leaders, matched to your ideal customer profile, reacting to your new pricing, your new category story, your half-finished product concept, with the results sitting in a dashboard before your coffee gets cold.

For anyone who has ever tried to run real research on B2B buyers, that pitch lands somewhere between a miracle and a con. The hard part of B2B insight work was never the analysis. It was access. The people whose opinions actually move a deal, the CFO who signs, the CISO who can veto, the VP of ops who has to live with your software for three years, are precisely the people who will not take your fifteen-minute survey for a $50 gift card. So teams settle. They talk to five customers instead of fifty, extrapolate from a lopsided panel, and quietly caveat every deck with "directional."

Synthetic respondents promise to erase that constraint overnight. Feed a model your ICP, and it generates hundreds of AI "copies" of your buyers that will answer questions, sit for interviews, and complete surveys on command. It is one of the fastest-moving ideas in B2B research right now, and also one of the most dangerous, because the failure mode is not that it doesn't work. The failure mode is that it works just well enough to fool the people who most want to believe it.

For CMOs, product marketers, customer-insight leaders, RevOps teams, and product managers at any B2B company that has ever struggled to get enough of the right buyers in a room, this is about a tool that is already inside your insight pipeline, whether you approved it or not, and the discipline that separates a genuine shortcut from expensive self-deception.

What a synthetic respondent actually is

Strip away the marketing and a synthetic respondent is a large language model, prompted or fine-tuned with demographic, firmographic, and behavioral data, asked to role-play a research participant so its output can be treated as if a human said it. Ask it what a 500-person manufacturer's IT director thinks about your onboarding, and it will produce a fluent, plausible, confident answer. Ask a thousand of them, and you get a distribution that looks, at a glance, exactly like survey data.

The category has moved from novelty to line item with startling speed. In its 2026 Market Research Trends report, based on a Q3 2025 survey of more than 3,000 research professionals across 14 countries, Qualtrics found that 69% of market research professionals have already used synthetic responses, and 87% reported being satisfied with the results. The company now sells a synthetic panel built on a model fine-tuned specifically for research, trained on a foundation of more than 200 million third-party global respondents, which it claims delivers 12 times the accuracy of a general-purpose LLM. It is expanding that capability from the US into the UK, Ireland, Canada, Australia, and New Zealand through the first half of 2026. The underlying synthetic data generation market, worth roughly $587 million in 2026, is projected to grow at about 35% a year toward $8.8 billion by 2035.

This is no longer a lab experiment. It is a purchased capability sitting in the research stacks of companies you compete with.

Why B2B fell for this harder than anyone

Consumer researchers have panels. They can buy a representative sample of 2,000 US adults without much trouble. B2B researchers have always lived with a structural scarcity that synthetic respondents seem tailor-made to solve: the total population of, say, healthcare CIOs at hospital systems above 1,000 beds might be a few hundred people worldwide, and every vendor in the category is trying to reach the same list.

That scarcity is exactly where the synthetic pitch is sharpest. A wave of B2B-specific vendors has emerged to exploit it. NewtonX launched what it billed as the first B2B synthetic personas solution, built on identity-verified professional data rather than generic web scraping. Evidenza generates hundreds or thousands of synthetic customers for a given product category and lets teams "survey AI copies of their customers," explicitly targeting the hardest-to-reach B2B buyers: the CEOs, CIOs, CFOs, and CMOs that traditional survey methods find prohibitively expensive or nearly impossible to recruit. In one profiled engagement, a team replaced a 12-month research grind with a roughly one-month sprint, generating synthetic buyers and distilling them into buying-center personas that connected brand and demand work. Bain has published its own analysis of "synthetic customers earning their stripes."

The appeal is obvious and, in narrow doses, real. When your alternative is guessing, a fast, cheap, always-available proxy for an audience you can't otherwise reach feels like a genuine unlock. The problem starts the moment "proxy for an audience I can't reach" quietly becomes "the audience." And the people who do this work for a living can see that line coming.

The adoption chasm nobody in the sales deck mentions

Here is the number that should stop every revenue leader mid-sentence. When User Interviews fielded its 2026 State of Synthetic Users report, surveying 150 research professionals in May 2026 (93% of them UX and user researchers, 62% inside enterprises of 500-plus employees), it found a population that is about as AI-saturated as any in the enterprise: 97% use AI somewhere in their research workflow, and 81% use it regularly. These are not Luddites clutching their clipboards.

And yet only 8% of them regularly use tools that generate synthetic participants. The single largest group, 28%, has never used them and actively chooses not to. Total lifetime trial sits at 29%, a stunning rejection rate for a technology that has been marketed aggressively into this exact buyer base for three years. Asked their overall sentiment, 47% described themselves as skeptical and another 17% as outright opposed, putting 64% of the profession on the negative side of the ledger. Genuine enthusiasm registered at 3.3%: five people out of 150. Not a single respondent said they had no significant concerns. And exactly one person endorsed synthetic users as a standalone tool, valid without any human participants. That is the entire addressable market for the "replace your panel" pitch: one researcher in 150.

This gap matters because it is not a technology-adoption lag waiting to close with the next model release. It is an informed verdict from the people closest to the work. When 96% of a market has heard of a product, most have researched it, and they still walk away, more webinars will not change their minds. They have already done the homework.

Where the evidence actually breaks

The most useful thing about the skepticism is that the peer-reviewed literature independently backs it up, and the picture it paints is not "synthetic research is worthless." It is far more precise, and far more actionable, than that.

On the encouraging side: in one evaluation spanning 57 real consumer surveys with 9,300 human participants, synthetic respondents reproduced roughly 90% of human test-retest reliability and better than 85% distributional similarity. For structured, upstream tasks, ranking features, sanity-checking a screener, exploring which of three messages lands, that is genuinely useful signal at a fraction of the cost and time.

Push past that boundary, though, and the same models fall apart in ways that are invisible until they cost you. A study in the journal Political Analysis, pointedly titled "Synthetic Replacements for Human Survey Data? The Perils of Large Language Models," found that AI-generated responses could match the headline numbers of a gold-standard survey while the underlying statistics were broken underneath: variance too tight, roughly half of the between-variable correlations pointing the wrong way, and results that shifted with nothing more than a change in prompt wording. Separate work replicating classic experiments documented a "hyper-accuracy distortion," where models return suspiciously clean, low-variance answers precisely where real humans are noisy and contradictory. Benchmarks against the World Values Survey showed accuracy holding for wealthy, English-speaking, Western populations and degrading sharply everywhere else. And when researchers asked LLMs to forecast the 2024 European election results, the output was, in the plain description of the analysis, generally disastrous.

Translate that for an operator. Synthetic respondents can help you generate hypotheses and pressure-test instruments. They cannot support a pricing decision, a demand forecast, a TAM claim, or a product-market-fit verdict, no matter how large the synthetic sample size, because the statistical machinery those decisions rely on is exactly what breaks first. The confidence stays high while the reliability drains out, which is the most expensive combination in all of research.

The real risk is human, not technical

If the problem were only that the models are imperfect, you could manage it with disclaimers. The deeper risk lives in the org chart. When researchers in the User Interviews study named their biggest concerns, quality and accuracy led at 89%, but the next two were not about the model at all. 80% worried about stakeholders over-trusting AI-generated findings, and 78% worried about synthetic data amplifying bias against underrepresented buyers.

Read that first one slowly, because it describes a principal-agent problem, not an engineering one. Researchers are not mainly afraid the AI will be wrong. They are afraid it will be wrong convincingly, that an executive who never wanted to fund proper research in the first place will treat fluent synthetic output as equivalent to validated customer evidence, skip the human study, and ship. Fluency reads as truth to anyone who can't audit the method, and synthetic output is nothing if not fluent.

The governance data turns that fear into a live exposure. 63% of organizations have no stance whatsoever on synthetic users; the decision is left to individual researchers. Another 15% don't know whether a policy exists. Only 11% have any formal position. Combine that with the 29% who have already tried the tools, and the conclusion is uncomfortable: synthetic data is very likely already flowing into insight pipelines, undocumented and ungoverned, inside companies whose leadership believes their roadmap and pricing rest on real buyers.

It gets worse where the seniority gap widens. Qualtrics found leaders consistently more bullish than the people who actually run the studies: 79% of research leaders expressed confidence in synthetic data quality versus 61% of individual contributors, and 68% of leaders considered themselves synthetic-data experts versus 41% of their frontline teams. The enthusiasm is concentrated exactly where the hands-on knowledge is thinnest. That is the precise recipe for a confident bad call, and the reason "our team uses AI research" is not, on its own, reassuring.

The disciplined way to use it

None of this argues for a ban. It argues for a boundary, and the good news is that the practitioners, the academics, and the vendors' own honest case studies all converge on the same one. In the User Interviews data, 41% of researchers accept synthetic users as a preliminary tool and 29% as a supplement, roughly 70% support for a bounded, upstream, human-anchored role. Even the teams selling the value describe a "blended" approach: synthetic breadth on the outside, human depth in the middle, and never the reverse. Gabb, a Qualtrics customer, described using synthetic data as a "cultural radar" to cut early testing from a week to hours, then validating high-stakes decisions with human panels.

Here is how to stay inside the line that works.

1. Draw a bright line between exploration and decisions. Write it down, literally. Synthetic respondents are cleared for instrument QA, hypothesis generation, edge-case exploration, and interviewer training. They are prohibited from pricing studies, demand forecasting, PMF validation, and any analysis with real capital riding on it. The disqualified list is not a matter of taste; it maps directly onto the statistical failures the research has already measured.

2. Use synthetic breadth to sharpen human depth, not replace it. The highest-return pattern is a sandwich. Run synthetic first to pilot your interview guide, catch a confusing survey question, and generate the hypotheses worth testing. Then spend your scarce, expensive human sessions on the questions that actually matter, and use synthetic once more only to help interpret, never to decide. You get more out of the ten real CISOs you can reach precisely because the AI absorbed the wasted motion around them.

3. Label every synthetic artifact, end to end. The top practitioner fear is stakeholder over-trust, and the cheapest fix in this entire playbook is a labeling rule: any chart, quote, or persona that contains synthetic-participant data is marked as such wherever it travels, so no executive ever mistakes a simulated buyer for a real one. Unlabeled synthetic data is how a "directional gut-check" quietly becomes "the customers told us."

4. Write the one-page policy before procurement writes it for you. With 63% of organizations governing this by accident, a single page, approved uses, prohibited uses, mandatory labeling, and a human-validation threshold for anything decision-grade, eliminates most of the risk at close to zero cost. It also produces something increasingly valuable: a clean, documented trail of how your customer evidence was generated. Acquirers and investors have started asking, and "we're not sure" is becoming an expensive answer.

5. Protect the research function, on purpose. 45% of researchers expect synthetic tools to be used as a justification for cutting research headcount. A team that reads the technology as an existential threat will never evaluate it honestly, and a vendor whose implicit ROI case is "fire your researchers" is selling against 92% of its own users. Position synthetic capacity explicitly as exploration and QA leverage that makes your researchers more valuable, not as a way to have fewer of them.

The buyer who doesn't exist

The seduction of synthetic respondents is that they solve B2B's oldest, most genuine pain. You really can't get 500 CISOs on a call, the access problem really is real, and a tool that manufactures an audience on demand really does feel like a cheat code for a game that was rigged against you. That pull is not stupid. It is the reason smart teams are about to make an avoidable mistake.

The mistake is confusing an audience you can generate with an audience that exists. A synthetic CISO has never fought a procurement committee, never burned political capital defending a purchase, never sat through a Q4 budget freeze and killed a deal you thought was closed. It will answer your question anyway, cleanly and confidently, and that clean confidence is exactly the tell. Real buyers are messy, contradictory, and occasionally wrong in ways that teach you something. The model gives you the average and calls it insight.

So use the tool for what the evidence says it's good at: the fast, cheap, upstream work of figuring out what to ask before you spend real access asking it. Draw the line at anything that moves money, and defend that line in writing, because the people above you who never wanted to fund research will be the first to argue it doesn't matter. The companies that get this right in 2026 will make sharper decisions than competitors drowning in cheap synthetic confidence. The ones that get it wrong will build a year of roadmap, pricing, and positioning on the opinions of buyers who were never there, and they won't find out until the real ones vote with their budgets.