From idea to impact
See how real people will react to your headline, statement, or script in minutes, before you launch.
simulate
Synthetic people read your message.
analyze
Reactions scored on emotion and intent.
predict
See how each group will respond.
optimise
Find the version that lands hardest.
trust
Every answer traces back to its source.
See it in action
Each synthetic person is built like a real one. They have a job, a city, local news they read, and online groups they follow. They react the way real people would, not the way a generic audience would.
The message we tested
“New plant based protein bar. Only 3g sugar.”
Maya R.
35 · marketing manager · Brooklyn, NY
City parent · busy schedule · health conscious
backgroundReads NYC parenting forums and local news
‘Only 3g sugar' jumps out. That's the number I check at the bodega. But show me the protein first. That's what makes me buy a bar like this.
Cal D.
54 · farmer · rural Iowa
Skeptical of marketing · wants real ingredients
backgroundReads farming forums and local Iowa news
‘Plant based' reads like a coastal pitch to me. Tell me what's actually in it. Soy, pea, what. Otherwise it sounds like a candy bar with extra steps.
Jordan T.
28 · software engineer · Seattle
Reads ingredient labels · skeptical of buzzwords
backgroundFollows Seattle tech and food news
Sugar's fine. Ingredient quality matters more to me. If it's whole food protein, I'll grab one for the office. If it's just sweetener and isolate, hard pass.
Devon K.
41 · HVAC technician · Atlanta suburbs
On the job · needs real food, not snacks
backgroundReads trade forums and Atlanta weather news
I burn 4,000 calories on a service day. 3g of sugar tells me it's not enough food. Show me the calorie count. Sugar isn't what I'm watching for.
Same message. The Brooklyn mom looks at the sugar. The Iowa farmer wants to know what's in it. The Seattle engineer wants real ingredients. The Atlanta tech needs more calories. Your written summary catches all of this and tells you who to rewrite for.
How it works
Pick from a template, or describe your audience in plain English. Age range, where they live, what they do for work. Or just paste your creative brief and we'll figure it out.
We create a balanced mix of synthetic people. Each one is shaped by where they live, their job, the local news in their area, and the online groups they follow. They all read your message and react.
Color coded scorecards show how each group reacted. You get the words that worked, real quotes, and a written summary of what to do next. Everything is sourced and ready to share.
Real examples
Three real examples. Two from political campaigns, one from a brand crisis. For each one, you'll see the actual message, what three different people said about it, and what the report recommended. Every test took under 15 minutes.
A campaign tested its closing line on voters in four swing state cities before spending on TV ads. The audience had a balanced mix of Democrats, Republicans, and independents.
Closing line · version 1
“We're not asking for a handout. We're asking for a fair shot.”
Voters in 4 swing state cities · mixed political views
What 3 people said (out of 14 groups tested)
Wayne K.
58 · pipefitter · rural Pennsylvania
Republican · working class · small town
That's the language. Nobody's giving us anything. We're earning it. Sounds like someone who's actually been on a job site.
Priya S.
42 · school administrator · Atlanta suburbs
Independent · college educated · suburban
‘Fair shot' is fine, but ‘not asking for a handout' sounds defensive. Like I should already disagree. I want to know what they'll actually do, not what they're not asking for.
Jordan A.
31 · designer · Phoenix
Democrat · young professional · city
It reads like a 1990s campaign clip. I don't disagree with it. I just don't feel anything. Where's the vision?
The takeaway
Strong with the base. Weak in the suburbs. The summary flagged ‘not asking for a handout' as the phrase that turned suburban voters off, and suggested a more hopeful rewrite that keeps the working class voters who loved the original.
A clean energy plan was tested with two different opening lines across all 50 states. The fundraising team loved the bolder version. The actual data told a different story.
Version A: about jobs
“A clean energy plan that creates two million American jobs.”
Version B: about Big Oil
“A clean energy plan that finally takes on Big Oil.”
1,000 voters across all 50 states · split between the two versions
What 3 people said (out of 17 groups tested)
Cal D.
54 · farmer · rural Iowa
Republican · rural · skeptical of DC fights
‘Two million jobs' is something I can talk about with my brother-in-law. ‘Big Oil' just sounds like the same DC fight everyone's tired of. Show me the jobs number.
Rita C.
47 · ops manager · suburban Detroit
Independent · suburban · works in auto industry
Version A I could email to my dad. Version B I'd skip. It feels like it's already decided I'm angry at someone.
Sam L.
29 · climate organizer · Brooklyn
Democrat · young · activist
Honestly Version B reads better to me. It names the problem. But Version A is the one I'd want my parents in Ohio to see. They'd call B ‘too much.'
The takeaway
Version A won with every single group. About 1.8× more people said it would persuade them. Version B was loved by the political base but turned off the swing voters who decide elections. The summary recommended using A for TV ads, and B only for fundraising emails to existing supporters.
A consumer electronics brand wrote three different apology statements after a battery recall. Before publishing, they tested all three with their customers.
Apology · version 3 (the winner)
“We're recalling every unit sold since March. If you own one, we're refunding it in full and shipping the safer version free. Here's what we got wrong, and here's the fix.”
600 people: loyal customers, recently churned customers, and people considering buying
What 3 people said (out of 9 groups tested)
Nadia H.
38 · loyal customer · suburban Boston
Bought 3 times before · trusts the brand
Okay. They said what went wrong, they're paying to fix it, and they gave a real timeline. That's the bar. I'd buy from them again.
Marcus B.
44 · former customer · rural Tennessee
Stopped buying last year · vocal on social media
‘Here's what we got wrong.' Finally. Most companies skip that part. I won't say I'm coming back, but I'd stop telling people not to buy them.
Elena R.
26 · considering a purchase · Los Angeles
New shopper · researching alternatives
Refund, free upgrade, and they admitted what went wrong. That's actually handled well. I'd put them back on my shortlist. The other two versions sounded like a lawyer wrote them.
The takeaway
Three apologies tested in 12 minutes total. The version that admitted the mistake and offered a refund earned 2.4× more trust than the first draft. Even from former customers who were warning others not to buy. They published Version 3 the next morning.
Also tested with
What you get back
See how your message scored on the things that matter. Does it grab attention. Does it feel positive. Would people actually act on it. Does it feel meant for them.
Joy, fear, surprise, anger. See exactly which emotions your message triggers, and where different groups feel differently about it.
One click and you can see who loves it, who's lukewarm, and who tunes out. Find the audience your message is quietly losing, before you spend the budget.
The words that worked, real quotes from people, what to test next, and where every finding came from. Ready to share with your team or your client.
Three studies. Same answer.
Tested three ways. Once against a personality inventory the model could have seen during training. Once against a political survey published after our training cutoff. Once head to head against asking the AI directly. The headline numbers are below. The full receipts follow if you want them.
Synthetic personas reproduce real personality distributions on the most-replicated scale in psychology.
See the receiptSame accuracy on a survey released after the model's training cutoff. Not memorized. Predicted.
See the receiptAsked the same AI without the persona pipeline. Off by 0.7 to 1.3 points; we were off by 0.1 to 0.2.
See the receiptThree different validations. Two land at 91% match. One shows where plain AI without personas falls apart. Full methodology, per-item breakdowns, and figures below.
First, a quick primer · how the tool works
We generate synthetic personas from public population distributions: U.S. Census demographics for age, region, education and income, plus published personality and value profiles. Each persona has a complete, internally consistent identity.
When you put a message, a poll, or a script in front of them, each persona reacts in their own voice, shaped by their personality, their life context, and the local culture they live in. The same persona stays consistent across every question.
To make sure the personas actually behave like real humans, we run published surveys through them and compare the answer distributions to the real respondents. Below are two of those tests.
Headline · Population Match Score
Population Match Score = 100 × (1 − |synthetic mean − real mean| / scale range), averaged across items. A score of 100 means our synthetic personas' average answer is identical to real respondents'; 50 means our error is half the scale. We chose this metric over R² because it stays interpretable across rating scales (1-5 personality, 1-7 ideology, 0-100 thermometers) and answers a plain question: how close are we to real people, on the scale the researcher used?
Per-item breakdown
Every dot is one item. The shaded zone marks items above 90% match. The headline is not an average over wild swings. Both studies have their bulk of items clustered between 87% and 99% match, on the same scale, against completely different real-world populations.
The Big Five (OCEAN) is the most replicated personality instrument in psychology. We ran 500 synthetic personas through the standard 10-item version on a 1-to-5 agreement scale, then compared their distributions to 345,443 real Americans from the Open Psychometrics dataset. 91.60% Population Match Score. The synthetic mean lands within roughly a third of a scale point of the real mean on a typical item.
Honest caveat: this dataset has been on the public web since 2018 and was likely in our model's training data. It's the strict benchmark, not the leakage-resistant one. That's why we ran a second study (below).
Per Big Five trait
Synthetic personas reproduce the real population's average score on every Big Five trait. Side-by-side bars show the synthetic mean (lighter) next to the real mean (darker) for each trait.
To rule out memorization, we ran the same kind of comparison against the Cooperative Election Study 2024 Common Content: 60,000 American adults, fielded by YouGov for Harvard, first released on Harvard Dataverse on April 3, 2025, after our model's training cutoff. Twelve numeric items (job approval, ideology ratings, self-rated ideology) were pre-registered before the run.
Result: 91.58% Population Match Score, within a tenth of a percentage point of the retrospective IPIP run, on data our model could not have memorized. Ten of twelve items within half a scale point; all twelve within one. When we split the personas by political party and check whether our Democrats look like real Democrats and our Republicans look like real Republicans, we get 88.35% within-cohort match. A model that had only memorized population averages would fail the within-cohort test by construction.
The receipt
Each dot is one CES item. Left: Democrat personas predicting the Democrat real mean. Right: Republicans predicting Republicans. Both panels hug the diagonal. Our personas reproduce the partisan structure of the response, not just the population average. CES Common Content DOI: 10.7910/DVN/X11EP6 · first released 2025-04-03 · model cutoff: Q1 2025.
What this does and doesn't prove: the personas reproduce the central tendency of real Republican and real Democrat ratings on items the model couldn't have memorized. Distribution shape (variance on bimodal partisan items) is still narrower than the real distribution. We surface this as the next open problem rather than hide it.
We re-ran the CES 2024 study with the persona pipeline turned off. Same AI, same 12 questions, same 500 respondents.
On the headline overall-mean number, the two methods are essentially tied. A vanilla AI is fluent at producing the average American answer when the average is all you're asked to predict. The structural difference shows up where real life varies, across partisan subgroups. Below: a single CES question (approval of Joe Biden, on a 1 to 4 scale) split between real Democrats and real Republicans. Each method's prediction sits next to the truth.
Real Democrats: what they actually said
“Do you approve of the way Joe Biden is handling his job as President?”
Reactiphi was within one tenth of a point of how real Democrats answered. The chatbot missed by seven times more. It can't see the partisan structure because it has no demographic input.
Real Republicans: what they actually said
“Do you approve of the way Joe Biden is handling his job as President?”
Same question, mirrored answer: real Republicans land near “strongly disapprove.” Reactiphi tracks them within two tenths of a point. The chatbot lands at neutral, missing by over a full point on a four-point scale, because it's anchored to one “average American” voice that doesn't exist.
Honest caveat: a vanilla chatbot still wins on a few items where the real population mostly agrees (Trump approval, Republican- party rating). The persona advantage shows up on items where real Americans split along visible demographic lines.
What this means
Two independent validations on completely different real-world populations both land at roughly 91% Population Match Score. One study the model could have memorized. One it couldn't. Both measurements landed in the same band. That's evidence our personas model people, not memorized distributions. And on the questions where Americans actually disagree with each other, our personas hold structural accuracy that asking a chatbot directly cannot.
For the technically minded: the literal scores are 91.60% and 91.58%, within the sampling noise of either study at this size. The magnitude is the signal, not the third digit.
Synthetic people are a strong first signal. Pair with real-world research for high-stakes decisions. Methodology and per-item gaps are surfaced openly above. A copy of the validation report is available on request.
Sign in and run your first test. Free trial. No credit card. No surprise bills.
Helpful as a first signal. Pair with real research for big decisions.
