B2B decision simulation with synthetic data asks AI decision questions through personas built from decision-makers’ profiles, instead of the people themselves, and measures how the brand is represented in those answers. A focus group explores the perception in the buyer’s mind; a simulation measures the information the buyer gets from AI. As the first step of buying increasingly happens in AI, that information becomes a measurable, repeatable data source. Its value depends on building brand-specific personas and validating findings against reality.
An industrial software company wants to understand who will evaluate its new product and how. The classic method is clear: focus groups with plant managers, maintenance leads and IT managers in the sector. In practice, getting them into one room for two hours takes months, incentives are expensive, and those who come often aren’t the real decision-makers.
Decision simulation with synthetic data approaches this from a different angle. We explained what synthetic data is in what is synthetic data. This article focuses on what simulation shows, and doesn’t show, compared with a focus group, in light of academic research.
Why are focus groups hard in B2B?
In consumer research, recruiting participants is relatively easy. In B2B, targets are few, their time is limited and they’re reluctant to sit at a table with competitors. According to 6sense’s 2024 research, the average buying group has 11 people; building a focus group that represents the whole group (users, finance, information security, procurement, management) is nearly impossible.
There’s also the observer effect. In focus groups, participants tend to describe how they think they should behave, not how they actually behave in a real decision. They don’t always say they started supplier research with AI or got their first list there. Yet in G2’s 2026 research, 51% of B2B software buyers say they start research with an AI chatbot more often than with Google.
What does academic research say about synthetic data?
How far language models can represent human behavior has become a serious research field. Two studies show both the potential and the limit clearly.
“Out of One, Many” (Argyle et al., Political Analysis, 2023). Researchers conditioned GPT-3 on the demographic backstories of real respondents to major US election surveys and asked the same questions. The response distributions of this “silicon sample” closely matched real survey distributions on many questions. The researchers called this property “algorithmic fidelity”: a properly conditioned model can emulate the response patterns of different subgroups.
“Using LLMs for Market Research” (Brand, Israeli and Ngwe, Harvard Business School, 2023). The study showed GPT could produce estimates of product preferences and willingness to pay of realistic magnitude, comparable to human studies. The same study notes the limit: estimates are sometimes inaccurate and occasionally even wrong-signed, and while the model gives reasonable aggregate results, it struggles to capture heterogeneity between individuals.
Both findings teach the same lesson to anyone working with synthetic data: a well-conditioned model can capture patterns, but used to predict individual decisions, it can be wrong.
Two different uses of synthetic data
There’s a critical distinction here that most synthetic data debates skip.
| Synthetic respondent | Decision simulation with synthetic personas | |
|---|---|---|
| Question | “What would this person think or prefer?” | “What does this person see when they ask AI?” |
| Model’s role | Answers in place of a human | The information source the human actually uses |
| What’s measured | An estimated opinion | AI’s actual answer |
| Core risk | The model misrepresents the human | The persona misrepresents the real buyer |
In the first use, the model imitates a person, and the limits academic studies point to apply directly. In the second, the model isn’t what’s being imitated; it’s the system being measured. Buyers really do ask ChatGPT, Gemini and Perplexity; the simulation measures the answer to that question. The core risk is then the accuracy of the persona, not the model. Recro’s decision simulation is built on the second use.
Focus group and simulation side by side
| Focus group | Decision simulation | |
|---|---|---|
| Measures | Participants’ perceptions, reasons, feelings | The brands, reasons and sources AI presents to the decision-maker |
| Time | Weeks to months | Days |
| Repeatability | Low; every group differs | High; the same question set can be rerun periodically |
| Scope | A limited number of participants | Many personas, languages and tools |
| Weakness | Observer effect, recruiting difficulty | Doesn’t measure human emotion or relationships |
The two methods don’t replace each other. Simulation answers “what is AI telling my customer?” A focus group or in-depth interview answers “how does my customer feel when they hear it?”
What does the quality of a synthetic persona depend on?
The value of the simulation depends on how closely personas are built to real decision-makers. A generic “procurement manager” written at a desk produces generic questions. A persona derived from real sales calls, RFPs and lost bids produces questions that carry real conditions. Argyle et al.’s finding supports this: algorithmic fidelity depends on how rich and realistic the conditioning profile is.
A persona is a set of decision criteria, not a job title: what they ask, what they filter on, what they trust, what they consider a risk. The decision question map is those personas turned into a question set.
The limits of simulation
- It doesn’t predict the real buyer’s decision; it measures the information the buyer gets from AI.
- Answers vary. In SparkToro and Gumshoe.ai’s 2026 research, the chance of two answers to the same question giving the same brand list was below 1 in 100. That’s why the pattern across repeated runs is read, not a single result.
- Models prefer different sources. Measurement in a single tool carries that tool’s bias; several tools are used.
- It doesn’t measure human factors like relationships and trust, which determine much of a B2B decision.
- Results are time-bound. As models and sources change, answers change; measurement should be repeated periodically.
How are results validated?
Findings are validated from two directions. The first is external reality: claims in answers are checked against their counterparts on the web, so information AI made up isn’t reported as the brand’s gap. The second is sales reality: is the persona the simulation shows the brand to be weak with also the one where objections and losses happen most in sales? Is the competitor that leads in the simulation the one sales meets most often? If overlap is high, findings are reliable; if it’s low, the question set or personas are reviewed.
How does Recro use synthetic data?
The Recro Insight Model and its application, the Recro B2B Focus Group Simulation, design synthetic data for the second use above: measuring AI’s actual answer through the eyes of brand-specific personas.
One simulation · 4 key pillars · 90+ steps
What sets Recro apart is four pillars built from scratch for each brand. Together they form a single simulation of 90+ steps, with each pillar producing the input for the next.
- Persona builder: Decision-maker profiles that represent the brand’s real buyers, specific to its sector and sales structure.
- Question builder: The brand-specific decision questions these people ask AI along their buying journey.
- Report builder: A report that reads the answers against the brand’s goals and competitors, together with the sources behind them.
- Action recommendation builder: A prioritized list of actions that closes the sector-specific signal gaps.
Because the core risk of synthetic data is persona accuracy, Recro’s weight sits on the first pillar:
- The persona builder derives personas from the brand’s real buyer structure and can optionally connect to anonymized CRM data. The synthetic persona then represents the real opportunities the brand wins and loses, not a desk-written type. No personal data is used; only roles and decision criteria are extracted.
- The question builder derives questions from persona attributes and sequences them in a natural buying journey flow, covering far more decision moments than a two-hour focus group can.
- The report builder reads answers through answer analysis, the links that source the conversation and KPIs set against the brand’s goals. Daily outputs are verified in a monthly reality check and finalized through human review.
- The action recommendation builder turns findings into a sector-specific signal list and prioritized actions. Reasons for not appearing are classified as Communication, Content and Authority GAPs.
Sample finding format · illustrative
- Persona: Maintenance lead · Automotive plant
- Question: Verification · “Which predictive maintenance software works with existing sensor infrastructure?”
- Finding: The brand is listed, but framed as “for large-scale facilities”; missing from mid-size facility questions.
- Sales check: Sales often hears “too big for us” objections at mid-size facilities.
- Action: Publish mid-size facility cases and tiered pricing as text.
Executive summary
- Synthetic data serves two different purposes: estimating a person’s opinion, or measuring the answer a person gets from AI. Recro’s simulation is built on the second.
- Academic findings show well-conditioned models can capture patterns but can be wrong when predicting individual decisions. Value depends on how realistic the persona is.
- Simulation doesn’t replace focus groups; it complements them. Findings should be validated against external sources and sales reality.
Frequently asked questions
Where do synthetic personas come from?
From the brand’s real buyer structure, the sector’s decision structure and customers’ own words. A persona is defined as a set of decision criteria, not a job title: what they ask, what they filter on, what they trust.
Aren’t AI answers variable? How can results be reliable?
Answers vary, so the pattern across repeated runs is read, not a single answer, and several tools are used. Reliability is tested by how well findings match external sources and sales data.
Is personal data used?
No. Personas represent roles, not real people. If a CRM connection is used, data is anonymized and used only to extract roles and decision criteria.
Is simulation cheaper than a focus group?
It’s usually faster and broader, because it doesn’t require recruiting and covers many personas, languages and tools. But since it answers a different question, cost comparison alone shouldn’t decide.
How often should simulation be repeated?
For most B2B brands, continuous monitoring with monthly reviews is ideal. Interim measurement can follow major model updates, content changes or competitor moves.
What do your decision-makers see when they ask AI? Recro simulates your buying committee’s questions without a focus group, using personas built for your brand.
Request a demo insight report →
Sources
- Argyle et al. · Out of One, Many: Using Language Models to Simulate Human Samples (Political Analysis, 2023)
- Brand, Israeli, Ngwe · Using LLMs for Market Research (Harvard Business School, 2023)
- 6sense · The B2B Buyer Experience Report for 2024
- Demand Gen Report · Half of B2B software buyers now start their research with AI chatbots: G2 (2026)
- Search Engine Land · AI recommendation lists repeat less than 1% of the time
This article was prepared with AI assistance and published under the review of Mehmet Semih İpek.
