Skip to content
Synthetic Audiences

Synthetic research for marketers

Know what works before you spend the budget

Digital twins of your audience answer your question in minutes, at a fraction of the cost of a research study.

See real cases
Twins answer

Discount or transparent cancellation?

Transparent cancellation90%
50% off first month70%
Plain offer (control)25%

Sell control over payments, not a discount.

Where do we launch first?

Adults in the brief100%
Can use the app32%
Category habit4%
Your real market1 in 200

The real market is 200× smaller than the headline.

Who are we missing?

Personas first19%
Segments first98%

Share of the market covered: 19% became 98%.

Real runs, shown as shares. Product anonymised.
  • Minutesfor a full cycle, once your data is in place
  • A fractionof the cost of a classic study: tens to hundreds of thousands of dollars
  • Beforeyou commit the media or production budget

Who it is for

Four decisions you no longer make blind

  • Marketing leads
  • Product and growth teams
  • Agencies and consultants

  • Where do we launch?

    Your real market, country by country, not the headline population.

  • What do we offer?

    Which feature, offer or price wins, measured against a control.

  • Who do we talk to?

    Segments that cover the market, each with a portrait of real life.

  • What do we say?

    Messages and videos checked on twins, before you pay for production.

Why teams choose it

Faster, cheaper, and it shows what you didn't ask

Classic study cost: typical range for a commissioned market study. The 19% → 98% figures come from a real run.

Real runs

Three cases, real numbers

Shown as shares. Products anonymised.

A mobile entertainment app picks where to launch in an emerging region.

18 of 20 choices agree · 6 runs · 2 models

The headline population is 200× the real market.

Results on a synthetic audience. Product anonymised.

Why trust it

Why you can believe the numbers

Data and real people, not a model's imagination.

Official statistics

Population, income, connectivity, language, each with source and year.

Real personalities

30 Big Five traits of real people in every twin.

  • O
  • C
  • E
  • A
  • N

They disagree like people

Not a chorus. As many twins per segment as your question needs, each with their own answer.

chorus twins · ● persona

Numbers from code

The model reads and writes. Code computes every figure: the same on every rerun.

Straight talk. Twins compare options far better than they predict exact rates, so every result is a comparison against a control.

Your data

Tell us about your product. We find the data.

No data room needed to start. Add your own data whenever you are ready.

  • Open datastatistics and household surveys we collect
  • Voice of the audiencereviews and forums about you and competitors we collect
  • Your datacampaigns, analytics, CRM you connect
Twins of your audience built in minutes, sharper with every source you add

Security

Your data stays yours

Raw personal data never goes to model providers, in any mode.

Every value has a kind

Open

Public statistics, surveys, published reviews.

Commercial secret

Your campaigns, pricing, plans and analytics.

Personal

Names, contacts and records of your customers.

Three ways to process it

1 · Model providersno training, no storage 2 · Privacy gatewayour model strips personal details 3 · Your own modelnothing leaves you
🌐 Open ✓ ✓ ✓
💼 Commercial ✓ ✓ ✓
🔒 Personal only totals ✓ cleaned ✓
  • Isolated at database level
  • Keys encrypted
  • Region of your choice
  • Export or delete everything

How to start

Three steps from question to decision

  1. Tell us about your product

    Company, product, the decision you face. We gather the data.

  2. Ask your question

    Twins of your audience answer, against a control.

  3. Decide with evidence

    Numbers, ranges and a clear mark of how well each is proven.

Run a campaign and bring the results back: the platform tunes the twins to your market, and what it learns stays in your product's library.

Early access

Test before you pay for it

Coming soon

Access is by invitation while we finish the first version. Requests open here soon.

The full process, six stages

1. Brief

Three short steps: company (brand, tone, competitors), product (what it is, price, how it makes money, where it sells) and goal from a catalogue: enter a new market, launch a feature, find new segments and more. The next product needs only the product and goal steps.

2. Market

One funnel per country, each step a share of the one before: population, able to use, total addressable market, serviceable market, obtainable market. The model only chooses which factors apply, from a fixed set; code does the arithmetic. Where no data exists, the step is an assumption with a range, and an agent can search the web for data to replace it. You decide whether to use what it found.

3. Segments & personas

Two models each propose segments three times, using measurable bounds only; code computes sizes, overlaps and coverage (target 80–95% of the market). Each segment gets a portrait from national household surveys with sources, and one persona you can interview, alone or in a group.

4. Twins test

Interviews produce hypotheses. Twins of each segment answer them with proven research tools: rating scales, forced choice, best-worst ranking, price sensitivity. Every variant is rated on its own, there is always a control, shares come with a 95% range, and twins never see which variant is "ours". Synthetic audiences rank options better than they predict exact rates, so every result is a comparison.

5. Messages & materials

Messages are built on hypotheses the twins confirmed, with a channel chosen from the segment's portrait. Videos, images, texts and audio scripts are drafted or uploaded; each twin watches part by part. You get a map of where interest drops and why, a suggested fix, and a before-and-after on the same twins.

6. Campaign calibration

Winning variants go to a real campaign. Results come back, the twins' forecast is compared with what happened, segment weights are updated, and each hypothesis is confirmed or rejected with "forecast vs fact" numbers. What is learned stays in your product's library for the next research.

Enter a new market

A consumer mobile entertainment app — free with ads, with a paid tier — choosing where to launch in an emerging region of dozens of countries.

The funnel

Market funnel as shares
StepShareBased on
Adults the brief is about100%population statistics
Able to use the app (own a smartphone)32% of themdata, by country
Have the category habit13% of those (8–20%)found data, replaced the model's 8–25% assumption
Want the product's format12% of thosefound data
Market the product can serve≈ 0.5% of the start (0.2–1.3%)computed

Reproducible

Parameters were chosen six times by two models; 18 of 20 choices agreed, and the two that differed are shown with the numbers of each option. Recomputing gives the same numbers, twice in a row.

Segments

Three segments by age, 36%, 29% and 35% of the market — together all of it. Code computed their sizes in each country.

Market potential

For every segment × country cell the platform answers six separate questions: how many people, will they try it (asked of the twins), how much money, how easy to reach, how crowded, what one user costs. There is no weighted score: you sort by any column and decide.

Results on a synthetic audience. Product anonymised.

Test a new feature

In interviews, two personas named the same barrier: not the price, but a hidden way to cancel and fear of an unexpected charge. That became a hypothesis: transparent cancellation will win more subscribers than a 50% discount.

Four offers, rated one by one

"Would subscribe" (4 or 5 of 5), older segment

Share of twins who would subscribe, by run
OfferFirst runRepeat, hints removed
Control10%25%
Discount15%70%
Transparent cancellation95%90%
Manual renewal70%35%

20 twins per segment in two segments. Transparent cancellation led the control by 85 points, and still by 65 after the repeat. Manual renewal fell from +60 to +10: part of its first result had been led by the description — the repeat exists to catch exactly this.

In a third test on two segments, transparent cancellation again beat the discount: 40% vs 20% and 70% vs 10%. In the younger segment the discount moved no one, and no offer reached more than a quarter in the repeat — a separate question for the next research.

Verdict: confirmed on twins. The next step is a campaign that can turn it green.

Results on a synthetic audience. Product anonymised.

Find new segments

The first run wrote personas first and fitted segments to them; together they covered 19% of the market. In the next runs the market was cut into segments first: 98% covered with five segments, then 100% with three or four.

Four segments, age × income

Shares of the market: 40%, 26%, 20% and 14%. Each has a portrait from national household surveys:

Portraits of two segments from survey data
Who they are in life25–44, better-off25–44, lower income
Working now57%41%
Married or living with a partner77%75%
Watch TV at least once a week79%51%
Listen to the radio at least once a week43%32%
No TV, radio or newspaper in a week20%39%
Use the internet daily30%26%
Have a bank or phone-money account64%51%

Each figure has its source, years and how much of the segment it covers. The model writes two separate parts: what the data shows, using only these figures (code checks), and hypotheses about daily rhythm, home and worries — marked as assumptions, to be checked on twins or with data.

Personas are then written with their segment's portrait, so their jobs, families and media habits are plausible for the segment.

Results on a synthetic audience. Product anonymised.

The Big Five (OCEAN)

The most studied model of personality: five broad traits — Openness, Conscientiousness, Extraversion, Agreeableness and Neuroticism (emotional sensitivity) — each split into six narrower traits, 30 in all.

Our twins take these 30 scores from questionnaires answered by real people — a pool of over 100,000. Each segment's group is picked to lean the way that segment does, yet within one segment you still get the cautious and the curious, the anxious and the calm.

Why it helps. A language model asked to "be a woman of 30" tends to give the same polite answer every time. Real profiles bring real spread, show who reacts to what, and keep twins from agreeing in chorus.

The voice of your audience

On request we collect what people say about your product and your competitors — app store reviews, review sites, forums, comments — in your countries and languages.

Out come motives, barriers and complaints with their frequency, what people praise and criticise in competitors, and real phrases in their own words. They feed personas (real motives and manner of speech), hypotheses and ad messages.

Personal names and nicknames are removed. Reviews come mostly from the very happy and the very unhappy, so we use them for motives and barriers, not for shares — and label them so.

How your data moves

Every value has a kind: 🌐 open, 💼 commercial secret or 🔒 personal. Before each model call the platform checks the most private kind in the prompt and refuses to send what the chosen mode does not allow.

1. Model providers

For open and commercial data. Providers neither train on your data nor keep it. Personal data reaches them only as totals computed by code.

2. Privacy gateway

For personal data. Our own model removes names, contacts and other personal details first; only the cleaned result goes on.

3. Your own model

Everything runs on a model you host. Nothing leaves your side.

Your account: isolated at database level · keys encrypted · region of your choice · export or delete everything

Raw personal data never reaches model providers in any mode.