3.11 Synthetic Persona Screening
At a Glance
Other names AI Persona Screening · LLM Persona Simulation
In Brief
Prompt a large language model (LLM) with a detailed customer description, then run a simulated interview by asking the same open-ended questions you would ask a real customer, varying the persona across segments. The output is a set of candidate hypotheses — possible pain points, objections, language patterns, and segment differences — that you then validate with real humans. It is preparation for customer interviews, not a substitute for them.
Common Use Case
You are preparing for your first round of customer discovery interviews but you are not sure what questions to ask or what objections to expect. You prompt an AI with a detailed description of your target customer and run a simulated interview conversation. The AI raises objections you had not considered, which you add to your real interview guide.
Helps Answer
- What pain points might this type of customer have?
- What words does the target customer use to describe their problem?
- What objections might someone raise about this solution?
- Are there related customer groups we have not thought about?
- What questions should we ask in our first real interviews?
Description
Synthetic persona screening is a structured exercise where you write a detailed customer description, prompt an LLM to role-play that persona, and interview it the way you would interview a real person: open-ended questions about their day, their problems, and their reaction to your value proposition. You run the conversation again across a few persona variations, then pull out the recurring pain points, objections, and phrasing. What you get is a hypothesis list — candidate pain points, plausible objections, language patterns, and segment-level differences — not validated findings. It expands your hypothesis space before you talk to anyone real; it does not measure or evaluate anything on its own.
A synthetic persona is not a substitute for talking to real customers. Conditioning a model on detailed demographic backstories can produce useful population-level patterns, but individual responses should not be expected to match a real respondent, and known LLM shortcomings — training-data bias toward English-speaking and Western populations, sycophancy, coherence bias, and a tendency to invent opinions — carry through into the persona. The output tends to be too shallow for deeper research and is most useful for broad attitudinal questions. Even vendors who sell synthetic-respondent platforms position their tools as a discovery aid rather than a replacement for real research.
The reason to run it anyway is cost and speed: a session runs for a few dollars in under an hour, and it often surfaces objections and language a founder had not considered. Used as a substitute for real research, it produces confident answers that may be wrong. Used to prepare for interviews with real people, it sharpens the questions and the screener.
How to
Prep
-
Define your target persona in detail. Write a rich description including demographics, job title, industry, daily workflow, goals, frustrations, tools they currently use, and any relevant psychographic details. The more specific, the better. “A 35-year-old ops manager at a 50-person logistics company who spends 3 hours a day in spreadsheets reconciling shipment data” is far more useful than “a logistics professional.”
-
Choose persona variation axes. Decide in advance which dimensions you will vary across runs — company size, seniority, industry vertical, geography, level of technical sophistication. Three to five variations is usually enough to surface segment differences without drowning in transcripts.
-
Pick your platform. Any general-purpose LLM chat tool works well for persona role-play with custom instructions; dedicated synthetic-respondent platforms package the workflow with built-in persona libraries and synthesis. The platform changes convenience, not the validity caveats.
-
Craft the system prompt. Instruct the LLM to role-play as this persona. Be explicit: “You are [persona]. Respond as this person would — with their vocabulary, their priorities, their level of technical sophistication. Do not be helpful or agreeable beyond what this person would naturally be. If this person would be skeptical, be skeptical.”
Execution
-
Run a discovery conversation. Ask the same open-ended questions you’d ask in a real customer discovery interview:
- “What’s the hardest part of your day?”
- “Walk me through the last time [problem context] happened.”
- “What have you tried to solve this?”
- “What didn’t work about those solutions?”
-
Probe for objections. Present your value proposition and ask the synthetic persona to react honestly. Push for specifics: “What would make you hesitate to try this?” and “What would your boss say if you proposed buying this?”
-
Vary the persona. Run the same conversation with 3 to 5 variations along the axes you chose in prep. Note where responses diverge — those divergence points are often the most valuable hypotheses.
-
Extract and categorize hypotheses. Pull out distinct pain points, objections, language patterns, and segment differences. Tag each one as a hypothesis that needs real-world validation.
-
Design your real interview guide. Use the hypotheses to craft sharper screener questions and interview prompts for actual customer discovery interviews.
Analysis
The output is a hypothesis list, not a finding. Treat every insight with the same skepticism you would apply to your own brainstorming.
Look for:
- Recurring themes across persona variations. If the same pain point surfaces across multiple persona descriptions, it’s a higher-priority hypothesis to test.
- Surprising objections. Objections you didn’t anticipate are often the most valuable outputs, because they expand your hypothesis space rather than confirming what you already believe.
- Language patterns. The specific words and phrases the LLM uses can be useful starting points for ad copy, landing page headlines, and search keyword research — but only after you confirm real customers actually use those terms.
- Segment boundaries. Where persona variations produce dramatically different responses, you may have found a meaningful segmentation boundary worth exploring.
Do NOT treat any of the following as validated:
- Willingness to pay
- Feature priorities
- Emotional intensity of pain points
- Market size or demand signals
-
Sycophancy bias LLMs are trained on human feedback that rewards agreeable, helpful responses. A synthetic persona will almost always be more receptive to your idea than a real human would be. If the AI persona says “that sounds interesting, I’d definitely try it,” that is virtually meaningless. Real customers hedge, deflect, and politely lie — and they’re still more honest than an LLM playing a character.
-
Training data bias The LLM’s responses reflect its training corpus, which overrepresents English-speaking, tech-savvy, Western perspectives. If your target customer is a rural small business owner in Southeast Asia, the synthetic persona will be drawing on very thin data. To partially counteract this, prompt for outside perspectives — “What would an ethnographer observe that wouldn’t appear in a standard business analysis?”
-
Confirmation bias Because you write the persona prompt, you inevitably encode your own assumptions about the customer. The LLM then reflects those assumptions back to you, creating a closed loop. Actively prompt for disagreement and skepticism to partially counteract this.
-
Coherence bias LLMs generate internally consistent narratives. Real customers are contradictory — they say one thing and do another, hold incompatible priorities, and change their minds mid-sentence. Synthetic personas are unrealistically coherent.
-
Substitution trap The method is so fast and cheap that founders run dozens of synthetic sessions and skip the harder work of recruiting real humans. No quantity of synthetic interviews compensates for zero real ones.
-
Anchoring bias Once you’ve seen the AI’s list of pain points or persona behaviors, it becomes the boundary of your imagination. You design your interview guide around the AI’s framing and never surface insights that didn’t appear on the list. Generate your own hypotheses before prompting AI. If you can’t produce ideas that aren’t on the AI’s list, you’ve been anchored.
Learn more
Case Studies
Radius Insights / MilkPEP
ran a concept test with AI synthetic respondents alongside ~200 real respondents per concept; synthetic and real top-two-box scores aligned on four of six measures, and the divergences themselves became the finding.
Got something to add? Share with the community.
