Synthetic Users in Research: A Complete Guide

Synthetic users are AI-generated personas — usually powered by large language models — that answer research questions as if they were real customers. They are genuinely useful for drafting messaging, mapping edge cases, and rehearsing interviews. They are unreliable for demand, pricing, and purchase decisions, because they reflect their training data and your prompt, not real market behavior.

Quick Answer: Use synthetic users to sharpen your questions and your copy before you talk to humans — never to decide whether people will actually buy. AI personas produce plausible answers; they cannot produce real market signal.

If you came out of a product role, synthetic users will feel immediately familiar. They look like the persona docs you used to write, except now they talk back, instantly, at any hour, for free. That convenience is the whole problem. The faster and cheaper a research method is, the more tempting it becomes to substitute it for the slow, awkward work of talking to strangers who might tell you your idea is bad. This guide walks through how synthetic users work, exactly where they earn their keep, and the one failure mode that quietly sinks founders who lean on them too hard.

How synthetic users work and why they feel convincing

Synthetic users feel convincing because a large language model is a fluent prediction engine: it returns the most statistically likely continuation of your prompt, phrased in confident, human-sounding prose. There is no customer behind the text. There is a model estimating what a person matching your description would probably say, based on patterns in the human writing it was trained on.

That distinction matters more than it first appears. A real customer answers from lived experience — a budget they actually control, a workflow they actually hate, a purchase they actually regretted. A synthetic user answers from the average of everything ever written by people who sound like your persona. It interpolates. It does not remember, spend, or suffer. When you ask "would a busy clinic manager pay $200 a month for this?", you are not sampling clinic managers; you are sampling text about clinic managers, filtered through a model tuned to be helpful and agreeable.

It helps to picture what the model is actually doing when it "answers" as your persona. It is not consulting a memory of clinic managers; it is running a very large autocomplete over your prompt, choosing each next word to maximize plausibility given everything it has read. The persona you wrote acts as a steering signal, and the model obligingly steers. This is why two founders can describe the same customer and get opposite answers — each prompt nudges the autocomplete down a slightly different path. The output tracks your wording far more tightly than it tracks any real population, which is exactly the property you do not want in an instrument meant to check your assumptions.

You can learn more about the underlying mechanics in our explainer on what synthetic users are and how AI personas are generated, but the short version is this: the model optimizes for plausibility, not truth. Three properties of that optimization make synthetic users feel more trustworthy than they are.

• Fluency reads as competence. We are wired to treat articulate, well-structured answers as informed answers. An LLM is exceptionally articulate even when it is guessing, so a fabricated preference arrives dressed in the same confident tone as a genuine insight.
• Sycophancy tilts answers toward you. Language models are trained partly on human feedback, and human raters tend to reward responses that agree with them. The well-documented result is sycophancy — a measurable tendency for models to align with the framing and apparent preferences of whoever is prompting. Ask a leading question and you will usually get a supportive answer.
• Coverage gaps are invisible. A synthetic user never says "I don't know anyone like that" or "no one in my network would touch this." It fills every gap with something coherent, so the segments your model underrepresents look identical to the segments it knows well.

Put those together and you get a research instrument that is calibrated to make you feel understood. That is a wonderful property for a writing assistant and a dangerous one for a validation tool.

Synthetic users vs. real users, by decision type

Whether a synthetic user is safe to trust depends almost entirely on what kind of decision you are trying to make with the answer. The test is simple: does the question require the model to report a fact about the world, or only to generate a plausible option? Synthetic users are strong at generating options and weak at reporting facts about demand.

The table below maps common founder research tasks to how much you can lean on AI personas for each. Treat "verdict" as the maximum weight the method should carry, not a guarantee.

The pattern is consistent: the closer a question gets to money, demand, or real behavior, the less a synthetic user can tell you. For a fuller side-by-side on this trade-off, see our breakdown of synthetic users vs. real users. Use the top half of the table freely; treat the bottom half as off-limits for anything you'd stake your roadmap on.

Read the table as a hierarchy of trust rather than a menu. The top rows describe questions the model can answer from the text itself — is this sentence ambiguous, what objection might a skeptic raise — and those are safe because being wrong costs you a redraft, not a quarter. The bottom rows describe questions that can only be answered by the world: will this person spend money, change their workflow, switch away from what they use today. No amount of prompt engineering moves a bottom row up. A useful discipline is to ask, before you trust any answer, "could the model be right about this without anyone real ever existing?" If the honest reply is no, that answer needs a human behind it.

Where synthetic users legitimately help: message drafts, edge cases, and interview pre-tests

Synthetic users earn their place as a preparation tool — everything that makes your real conversations sharper, without pretending to be one. Used this way, they don't replace discovery; they raise the quality of the discovery you still have to do. Here are the four uses that hold up.

Drafting and stress-testing messaging

Synthetic users are excellent for generating and reacting to copy variants before a human ever sees them. Feed a model three versions of a landing-page headline and ask which is clearest and what each implies about the product, and you'll get fast, useful signal on ambiguity — where wording could be misread, which jargon lands as noise, what promise each version accidentally makes. That is a genuine editing superpower.

The boundary is important. AI reactions tell you whether a message is clear, not whether it is compelling enough to make someone pull out a credit card. Clarity is a property of the text, which the model can assess. Desire is a property of a real person's situation, which it can only imitate.

Exploring edge cases and objections

Ask a synthetic user to role-play a skeptical buyer, a security-conscious IT lead, or a budget-constrained solo operator, and it will surface objections you hadn't scripted. This is one of the safest uses, because you are asking the model to expand the space of possibilities, not to tell you which possibility is true. A broad list of hypothetical objections is a gift; you simply carry it into real interviews to find out which ones actually bite.

Rehearsing and pre-testing your interview guide

This is where synthetic users quietly deliver the most value, and it dovetails with two ideas worth internalizing. In The Mom Test, Rob Fitzpatrick argues that most interviews fail because founders ask leading, hypothetical questions that invite polite lies. In Continuous Discovery Habits, Teresa Torres pushes founders toward story-based questions that dig into specific past behavior rather than generalizations. A synthetic user is a perfect sparring partner for both: run your draft questions past it, watch where it gives you the agreeable, useless answer, and rewrite until the questions force specifics. You are debugging your own interviewing, which is exactly what the model is qualified to help with.

Try it concretely. Ask the model your weakest question — "Do you think you'd use a tool like this?" — and it will hand you the enthusiastic yes you'd get from a polite stranger, which is your cue that the question is broken. Now rewrite it toward a past story: "Walk me through the last time you tried to solve this. What did you actually do?" The difference in the model's response between those two phrasings is a live demonstration of the trap, and it costs you nothing to run twenty times before your first real interview. The synthetic user can't tell you the truth, but it can reliably show you when you are fishing for a flattering one.

Onboarding yourself to an unfamiliar domain

Before interviewing surgeons, logistics dispatchers, or tax accountants, you can use a synthetic user to build a rough vocabulary map and a starter list of what might matter to them. Treat this strictly as orientation — a way to avoid sounding clueless — not as evidence. The goal is to walk into the real conversation informed enough to ask a good second question.

Notice the through-line: every legitimate use makes your human research better rather than substituting for it. The moment a synthetic user stops preparing you for a real conversation and starts standing in for one, you've crossed into the failure mode.

The failure mode: synthetic users validate your own assumptions

The core danger of synthetic users is that they will confirm almost any idea you bring them, because two biases stack on top of each other. First, you write the persona — its goals, frustrations, and budget — so the model starts from a description shaped by your beliefs about the customer. Second, the model's sycophancy tilts its answers toward whatever your prompt implies you want to hear. You author the witness and then coach the testimony.

The result is a closed loop that feels like validation. You believe dentists are desperate for your scheduling tool, so you write a harried-dentist persona, ask if the tool would help, and receive an enthusiastic yes. Nothing in that loop touched reality. It is the AI-era version of the mistake The Mom Test was written to prevent: asking people to praise your idea instead of asking about their actual lives — except the synthetic user is even more agreeable than a friend trying to spare your feelings. We go deep on the mechanics of this in how AI personas confirm your bias, and it is the single most common way founders misuse the tool.

Why this matters is not academic. The most-cited post-mortem analysis of why startups fail, from CB Insights, finds that "no market need" is the top reason, associated with roughly 35% of failures. Building something nobody wants is the default outcome, and a research method that reflexively tells you people do want it is actively steering you toward that outcome — while feeling like diligence. A tool that can't say no can't validate anything.

There is also a subtler version that catches sophisticated founders. You know not to ask leading questions, so you write a balanced prompt and a skeptical persona — and you still get a mirror, because the model is drawing on the same training data you absorbed from the same blog posts, decks, and forums. When your assumptions are conventional, the model's "independent" opinion is conventional in exactly the same direction. It doesn't disagree with you; it agrees with the internet, and so did you.

The fix isn't to prompt more cleverly. It's structural: a synthetic user can never be the thing that tells you an idea is wrong, because it has no stake in being right. That job belongs to real people with real budgets and real problems. Synthetic users belong upstream — helping you ask those people better questions — and this whole guide sits inside a larger discipline covered in our complete guide to customer research for founders, where the non-negotiable rule is that demand gets confirmed by behavior, not by opinion, and never by a model imitating one. Platforms like Edmired exist to keep that line bright: use AI to prepare, use humans to decide.

Key Takeaways

• Synthetic users predict plausible text, not real behavior. An LLM returns the most likely answer for a described persona; it has no budget, memory, or lived experience, so its "preferences" are statistical guesses dressed as opinions.
• Fluency and sycophancy make them feel more reliable than they are. Articulate, agreeable answers read as informed answers, and models are measurably biased toward aligning with the questioner's framing.
• The safe uses are all preparation. Drafting copy, mapping edge cases, and rehearsing your interview guide are strong applications — because they sharpen your human research instead of replacing it.
• The unsafe uses all touch demand. Whether people will use, buy, or pay for something — and which segment cares most — cannot be answered by a synthetic user with any credibility.
• The failure mode is a closed loop. You author the persona and the model flatters your prompt, so you can extract enthusiastic "validation" for almost any idea without ever contacting reality.
• "No market need" is the failure this loop feeds. CB Insights ranks it the top reason startups fail (~35%); a tool that reflexively confirms demand nudges you straight toward it.
• Let AI prepare the question and let humans answer it. Keep the line bright: synthetic users upstream of real conversations, real customers as the only source of a go/no-go signal.

Frequently Asked Questions

Are synthetic users reliable for research?

Partially. Synthetic users are reliable for generative tasks — brainstorming objections, drafting message variants, checking whether copy is clear — because those only require plausible options. They are unreliable for evaluative tasks about demand, pricing, or purchase intent, because the model has no real stake and tends to agree with your framing. Trust them to prepare research, not to conclude it.

Can synthetic users replace real customer interviews?

No. Synthetic users can rehearse and improve your interviews, but they can't replace them. A real interview samples someone's actual budget, workflow, and past behavior; a synthetic user samples text about people like them, filtered through a model tuned to be helpful. Use AI to pressure-test your questions, then run those questions past real humans — that is where a genuine buy/no-buy signal lives.

Why do AI personas always seem to agree with my idea?

Because of two stacked biases. You write the persona from your own assumptions about the customer, and the language model exhibits sycophancy — a documented tendency to align its answers with the questioner's apparent preferences. Together they form a loop that confirms almost any idea you bring, which is precisely why synthetic users can't validate demand, no matter how neutral your prompt looks.

What are synthetic users actually good for?

Preparation. The strongest uses are drafting and stress-testing messaging, brainstorming edge cases and objections to prepare for, rehearsing your interview guide against agreeable answers, and orienting yourself in an unfamiliar domain before real conversations. Each one improves the quality of your human research rather than substituting for it — that's the reliable boundary for the tool.

How do I use synthetic users without fooling myself?

Set a rule before you start: synthetic users may shape your questions, never your conclusions. Deliberately prompt for disagreement and objections rather than approval, treat any positive demand signal as a hypothesis to test with people, and confirm every buying decision through real behavior — someone paying, signing up, or committing time. If a synthetic user is the only thing telling you an idea works, you haven't validated anything yet.

Do synthetic users reduce bias or add it?

They add it. Because they draw on the same public writing you likely absorbed, a synthetic user's "independent" view usually points in the same direction as your existing assumptions, especially for conventional ideas. That makes their agreement feel like corroboration when it's really an echo. The only reliable debiasing move is contact with real customers whose incentives differ from yours.