Using AI to create synthetic users & digital twins

What is the difference between synthetic personas and digital twins?

When I talk about synthetic users, I am usually talking about two different things: personas and digital twins. People often mix them together, but they behave very differently in research.

Personas, which are composite, not a person. When I build a persona, I am blending insights from many people who share similar traits and contexts. In marketing language, it sits between an ICP and a segment. If I am working with a baby products company, for example, I might have younger urban moms, rural moms with more space and different budgets, dads who do most of the shopping and grandparents who spoil the kids. Each persona is defined by demographics, attitudes and archetypes and their role and KPIs on the client side.

A key point for me: for personas, I strip out PII. I am not trying to mirror any one individual. I am trying to represent a recognizable type in the market.

A digital twin is the opposite. A twin is based on one real individual with a clear digital footprint and often a direct relationship with me or my client. To build a twin, I pull together:

• Interview transcripts that I or my team conducted
• Public talks, podcasts, conference sessions
• Blog posts, articles, books and trade pieces
• Profiles or write ups that show how they think and what they care about

What I am trying to capture is not just biography. I want their belief system, decision logic and operating style. I use digital twins for people who are:

• Hard to schedule with, like physicians, CEOs, or skilled trades
• Critical stakeholders, where one conversation is never enough
• Participants I have already interviewed once, but cannot keep calling back

I always get consent when I do this for client work. In some cases I will offer to remove any identifying markers so the twin is de identified.

TLDR; Personas are composite and anonymized. Digital twins are grounded in a real person, used with permission and built from their own words.

Why is “trusted sources only” non negotiable for synthetic users?

Most of the negative press around synthetic users is aimed at a pattern I try to avoid: people let the model invent everything. If you say, “Give me a synthetic Head of Product for a SaaS company,” and stop there, the model will happily blend internet stereotypes into something that looks polished but has no lineage. It feels like insight, but it is an average of content, not a person or segment.

My approach is the opposite. I start from trusted, cited sources and treat synthetic users as structured research artifacts.

For every persona or twin, I want:

1. Traceable sources
   • URLs for long form content: talks, interviews, blog posts
   • Full text for my own interview transcripts
   • Screenshots, PDFs and other artifacts when they carry important context
2. Citations by default

   Even opinions should be anchored in specific places. I want to be able to say, “This comes from what they said in this talk,” or “This came from this paper.” If I cannot point to a source, I do not consider it part of the brain.

3. Tools that respect sources

   I prefer tools that automatically surface citations, show if a link is broken or suspect and let me limit answers to my own corpora when needed. In other words, a synthetic user is only as trustworthy as the body of data behind it. The safest way to work is to make that body small, clear, and fully cited.

How can someone try this in a low stakes way?

I like to call the knowledge base for a persona or twin a brain. Technically, it is just a curated corpus, a bit like a small RAG system. But the “brain” metaphor is useful when I explain the idea to clients. Here is how I build one.

Step 1: Map the digital footprint

For a twin, I start by mapping where their public presence actually lives. That usually includes:

• Personal or company website
• LinkedIn, if it is accessible in my tooling
• YouTube talks and podcast appearances
• Blog posts, essays or public articles
• Academic profiles like Google Scholar, where relevant

If it is someone I have interviewed, I add:

• Full interview transcripts
• My observation notes from the session

I also disambiguate people with the same name so I am not mixing two lives together.

Step 2: Consolidate into a single document

Once I have a list of sources, I consolidate. Instead of uploading 20 different files, I:

• Drop all URLs into a single document
• Paste in excerpts that matter
• Embed transcripts and key screenshots when needed
• Export the whole thing as one PDF

That PDF is the brain for that twin or persona. It is much easier to manage and reason about one file per brain.

Step 3: Name it clearly and date it

Every brain file follows a simple naming convention, for example:

SaaS_Product_Leaders_2025-12-05_MB.pdf

Where I include:

• Segment or person name
• Date
• My initials or the researcher’s initials

This tells me at a glance how fresh the brain is, and who curated it.

Step 4: Upload into my analysis environment

I then upload this PDF into a workspace that lets me:

• Restrict responses to this corpus alone when needed
• Get citations in every answer
• Generate extra views like mind maps, infographics, flashcards, and explainers

Once that is in place, the twin is ready for structured interviews. If this sounds abstract, the easiest way to start is to build a twin of yourself.

That is often what I do in workshops. We:

1. Take a participant’s public footprint
2. Build a small brain from their talks, posts and website
3. Ask the twin to summarize their philosophy or critique a piece of work

How do I interview synthetic users without getting fooled?

The value is not in having a twin. The value is in how you talk to it and how you validate its answers. I try to ask grounded questions that the sources can reasonably answer.

For example, “How has this person talked about risk and uncertainty in past talks?”, “When they evaluate a new vendor, what tradeoffs do they emphasize?”, “Based on their public writing, how might they respond to this concept?”. These are grounded in the data. If I know a rough true range for an answer, I can tell when the twin is drifting.

What I avoid are fantasy prompts like: “Invent a new product for them.” That is not research any more, that is creative writing.

I routinely ask twins obviously irrelevant questions as a stress test. If I have a twin of a Delta Airlines pilot and ask them how they would design a candy line for toddlers, and I get a confident, detailed answer, that tells me the system is stepping outside the brain. A good synthetic user should say something closer to: “This is outside my area and the sources do not cover this.”

For each twin, I carry a mental map of what is in scope including the core domain, where they are experts, adjacent topics they mention sometimes. This also helps me rule out areas that are clearly irrelevant.

When a synthetic user speaks with a lot of confidence far outside the domain of the sources, I treat that as a model problem, not a human insight.

Lastly, in the setups I use, I make it explicit that I want no sycophancy and just direct, specific feedback. I want the behavior of a careful, slightly conservative participant, not a hype machine.

Where do synthetic users actually help in research?

I do not see synthetic users as a replacement for research. I see them as a way to extend what good qualitative work already does. Here is where they have been most useful in my practice:

• Concept testing: I use twins and personas to pressure test early ideas. (“How would this segment react to this pricing concept?”; “What objections or risks would this stakeholder likely raise?”). It is directional, but it helps me refine what I take into real interviews.
• Internal education: Synthetic users can also be teaching tools. I have taken the same brain used for research and turned it into flashcards, short explainers and simple quizzes to reinforce who we are really building for.
• Post interview follow up questions: For busy people and CEOs who are hard to get time with, if I have follow ups after an initial conversation, I use their twin to answer those questions for me.

When there is real money or policy on the line, I still go back to live humans and proper methods.

What would you say to skeptics of synthetic personas and digital twins?

There is a lot of skepticism about synthetic users right now, and honestly, much of it is justified. If your process is describe a generic buyer persona and let the model hallucinate one without showing any sources — then yes, that should be challenged. You are mixing marketing fantasy with a probabilistic model and calling it insight.

However, if you start from grounded, cited data, keep the corpus narrow and intentional validate the behavior with pressure tests and use the outputs as directional, not definitive, then synthetic users can become a practical extension of research, not a shortcut around it.

It is hard to change minds with arguments alone. It is easier to show a twin built from real interviews and public talks, walk through the sources, and compare that to a generic, invented persona. The difference is obvious once you see it side by side.