How results are made
Four stages between your upload and the answer.
Nothing on the result page is a guess typed by a model. Each card is computed the same way every time, from what people could see, who they are, what they said, and how the versions compare.
The eye
First, what could be seen.
Before anyone reads, we work out what a glance takes in. For ads and pages, an attention model trained on real eye-tracking gives every part a chance of being looked at in about two seconds; each simulated person gets their own handful of landings and sees only what they landed on. For emails and posts, a reading model reads the first line and skims the rest, the way people do. Text that is too small or too faint to read is never passed on.
That is where “noticed by” comes from: the share of the audience whose glance reached each part.
Noticed by
Headline
91%
Button
52%
Logo
34%
Small print
3%
5 landings in a 2-second glance
The audience
Then, who is looking.
One sentence becomes a few hundred people. Each has a handful of fixed facts drawn to match the real mix of your market (role, company size, industry, country, seniority) and five traits: how sceptical they are, whether they think things through, their appetite for risk, how much the topic matters to them, and whether they know your brand.
The audience is frozen. Every version and every rerun meets exactly the same people, so a change in the result is a change in the thing, not in the crowd.
How audiences are builtHeads of security at mid-size European banks
The people
Each one meets it where it lives, and says what they think.
A person is not asked to “evaluate this ad”. They are scrolling LinkedIn on their phone, or opening their inbox, and this comes past. They answer a few short questions in their own words: what they noticed, what they think it is, whether it matters to them, what they doubt, and what they would do next. Nobody is asked for a number, and nobody sees two versions side by side.
Their words are what you read on the result page: the quotes, the themes, the unusual reactions.
Noticed
What I think this is
Does it matter to me
“Nothing in my environment deploys in 10 minutes. That line makes me trust the rest less.”
What I do next
DP · CISO · insurance · 3,800 staff · sceptical
The judge and the statistics
Finally, the answer, and how sure to be.
The verdict does not come from counting who liked what. The two versions are compared directly, ten times over, in five wordings and both orders, and the verdict is “stronger” only when the comparisons agree. The people never decide the verdict; they explain it.
Their words are placed on scales by meaning and grouped into themes. If they clearly point against the judge, the result becomes “too close to call”. Seven checks (do the comparisons agree, is the judge sure, do the people agree, does it hold on a rerun and on a second model) become one label: strong signal, some signal, or too close to call.
Compared with your current ad
Clearly stronger
9 of 10 comparisons agree · asked in five wordings, both orders
Comparisons agree
Judge is sure
The people do not point the other way
Holds on a rerun
Holds on a second model
Three rules every result follows.
01
Everything is a comparison.
Your new version against your current one, or against typical examples of its kind. Never a click rate, a conversion rate or a sales forecast: those need real data, and a simulation cannot give them honestly.
02
People explain. They do not decide.
The verdict comes from direct comparison. The simulated people give you the reasons, the objections and the words, which is what they are good at.
03
“Too close to call” is a real answer.
Most real launches compare versions that are close. When they are, the result says so, instead of inventing a winner.
Test your next launch on your customers first.
Your first audience can be ready this week.
Get early access