Start with the question
Upload a spot. Paste a URL. Send a script. Share a campaign. Or just tell us what you are debating.
What does Comprenda see in your marketing that everyone else misses? We will show you what the existing senior intelligence already supports, where the evidence stops, and what is worth proving next.
Most people arrive with a hunch rather than a decision. That is the normal starting point, and it is enough to work with.
You do not need a data strategy, an experimentation roadmap, or to know which Comprenda product you want. Start with the business decision.
Pick whichever of these sounds closest to your situation. It will fill in a first draft you can rewrite. The wording is not the point, the question behind it is. Each one links to what came back the last time we were asked something of that shape.
Or start from a decision that keeps recurring
How an engagement runs
Most of what people expect to be a study turns out to be a lookup, and it is worth knowing which one you have before committing to either.
Read the record, days
We take the question and screen it against messages that already ran where behaviour was observed and joined to a settled outcome. You get what appears knowable, what does not, and whether the question is already answered. No new data collection, and often no new data from you at all.
Settle it with a test, months
Where the record cannot decide, we randomise the next move at the moment it matters. On the screener leg of a book running about 76,000 conversations a month where the customer speaks, the first state reads out in about four weeks; on the transfer leg alone, about 8,000 a month, four to five months. Lower volume moves proportionally.
Required
The conversations or the creative, and the decision you want settled. If your calls already run through infrastructure we hold, the first engagement needs no new data agreement at all. That is usually the fastest possible start.
Helpful, not required
Your own outcome column. If you have a clean one we grade against it. If you do not, we read the outcome out of the interaction and tell you how far that reader is certified.
Opens up more
Adherence or persistency data, if you hold it. We do not have it today. With it the question changes from which message converts to which message produces a customer who stays, a different and larger question.
How your data is handled
Access is read-only and aggregate. Person identifiers are hashed before they reach analysis and are never used to single anyone out. Raw text stays inside the analysis environment; anything that leaves it is scrubbed, and every finding published on this site is de-identified and carries the population it was measured on.
We can work within customer-specific security and data-handling requirements, and we will tell you plainly which parts of a question we can answer without your data leaving your side.
The read
The first clock above, as a named piece of work. You send the thing you are about to commit to. What comes back is a read of each element in it against messages that already ran where behaviour was observed and joined to a settled outcome, with the population, channel and outcome printed on every line.
What you send
A script, a spot, a landing page, an agent prompt, or the question your team keeps debating. One is enough. You do not need an outcome column or a data agreement to start.
What comes back
For each element: what is known, on which people and which channel; where the evidence stops; what would resolve it; and what to test first. Every line carries its scope. Every element closes with a verdict: deploy, do not deploy, or not enough evidence.
The clock
Working days. The read is a lookup against a record that already exists, not a study, so it runs on the first clock above. Where the record has nothing on an element, the read says so and names the step that would fill it, rather than waiting on one.
What it is not
It is not a forecast of your ad or your script. It is evidence about the elements they are built from. It is not a randomised test; it is the way to choose which test is worth running, and the one instrument that can tell you a question has already been answered. Where the record has nothing on your population, the read says unknown and stops there.
Before the money goes out
A television flight, a new creative platform, a script rolled out across a floor. The commitment is made weeks before any outcome exists. The usual way to de-risk that is to recruit a panel of older adults and ask them what they think of it.
The ceiling on asking
We measured how consistent real older adults are with themselves. The same person, contacted on two different days a month apart, both calls connected, each with an independently settled outcome: 191 pairs agree 59.7% of the time, against 58.8% expected by chance, kappa +0.022.
On the outcome that pays, an older adult is close to chance with their own earlier self. That is a fact about the population, not a criticism of panel research, and it bears on any instrument pointed at them. Every survey, every copy test, every simulator, ours included. Nobody gets to be more reliable than the thing they are measuring.
Repeat-contacted seniors are a biased sample: the dialer chose who to call back. It is also the population anyone actually acts on. The 45-day row recovers to kappa +0.11 on 100 pairs and sits inside the noise band. It is not evidence the consistency returns.
So the useful question is not how to ask better. It is what you can learn without asking. Many of the elements inside a new flight, a claim, an order, a way of naming a cost, a reassurance, have appeared before somewhere that behaviour was observed, and some of those are joined to a settled outcome. Reading that is not a forecast of your ad. It is evidence about the elements your ad is built from, with the population, channel and outcome printed on it.
And the synthetic substitute is worse than nothing here
Replacing the panel with a simulated one does not remove the problem, it inverts it. Across scenario replications, up to 83% of effects that were null in humans came back significant in the model, with effect sizes two to three times too large.
Cui, Li & Zhou, Nature Computational Science 5:627–634 (2025). The 83% is the top of a 68–83% model-dependent range, on scenario replication. A tool that manufactures findings where there were none is not a cheaper panel; it is a faster way to be confidently wrong.
What a copy test gives you
Stated response from a recruited panel, on a population at chance with itself on the outcome that matters. Weeks, and a cost that scales with sample.
What the record gives you
What comparable message elements already did, on people who were not being observed for a study, graded on money that actually settled. Days, and it gets cheaper as the record grows.
Neither replaces a randomised test. Both are ways of choosing which test is worth running. And the record is the one of the two that can tell you a question has already been answered.
Received
We’ll take a look and come back with what appears knowable, what isn’t yet, and the most useful next step.