Home / Ad testing / Synthetic ad testing
Synthetic ad testing: what a modelled audience can and cannot tell you
Synthetic ad testing shows an ad to modelled respondents instead of real people. EchoDepth's synthetic audience can model up to 3,000 people for one ad. Each is built from census age and gender profiles for the audience you choose, then given a job, a region, a media diet, a level of category use and a relationship with your brand. Each person watches the ad the way it would reach them: in a TV ad break, as a skippable YouTube pre-roll, in a muted social feed, or among five other ads as a recall test does. They answer straight after and again the next day. The report shows who was won over, lukewarm, turned off or never noticed it, whether they linked it to the brand, and which channels and segments it fits best. It is a model of an audience, not a measurement of one: version 0.1 has not yet been calibrated against real panels, so we report it as a pre-test, never as measured viewers.
Updated 19 October 2026 by Jonathan Prescott, Founder, Cavefish.
How each modelled person is built
Every respondent is assembled from the same layers, and the report states which layers were modelled and how, so nothing is passed off as asked.
| Layer | What you choose | How it is derived |
|---|---|---|
| Who they are | The general public, or a profession such as teachers, nurses and midwives, HR, marketing or IT professionals, small business owners or senior leaders | For a profession, sex and social grade follow an approximate profile of that work |
| Age and gender | Age bands from 16–24 to 65+; women, men or both | Ages follow the working-age census, not a survey of the profession |
| Where they live | The UK by default, or any list of countries | Each respondent has a country, region and area, and either has children or does not |
| How they watch | TV-first, mixed or social-first | Each person gets a media diet; which channels they use is modelled from approximate UK reach by age and that diet, not asked |
| Category use | Heavy, light or non-users | Sets how familiar the category is to each respondent |
| Brand relationship | Customers, lapsed, considering, rejecters, never heard of them | Sets what each person already thinks of the brand before they watch |
| Memory | Nothing to set | Recall straight after and next day comes from each person's modelled memory; forgetting uses reasoned parameters, not yet fitted to real recall data |
How a synthetic audience test runs
From a brief to a report in minutes.
Describe the audience and the ad
Choose who the ad is for using the layers above, and how many people to model, up to 3,000. Add the brand, the category, competitors and what the ad is meant to say, so each respondent's takeaway can be checked against the intended message.
Each person watches in their own channel
A respondent sees the ad the way one of their own channels plays it: an ad break on live TV, a skippable pre-roll on YouTube, a muted feed on social, or a cinema screen. For a classic recall test, everyone sees it among five other ads instead.
They answer straight after and the next day
Did they notice it, did they like it, which brand was it for, are they more likely to consider it, and what did they take from it, in their own words. The next day they answer again from memory.
Read the report
Everyone is placed in one of four groups: won over, lukewarm, turned off or didn't notice it. You can arrange the audience by age, gender, children, area, where they live, media diet, brand relationship and channel; see which channels and segments index highest; read each person's profile and reaction; and see the moments where the modelled audience drifts away, with suggested fixes. Results come with ranges, and any segment with fewer than 30 people is not reported on its own.
Synthetic audience vs real viewers
EchoDepth offers both. They answer different questions, and the report never mixes one with the other.
| Synthetic audience | Measured test with real viewers | |
|---|---|---|
| Who watches | Up to 3,000 modelled people built from census profiles | Real, consenting people from your audience, on their own webcam |
| What it reports | Won over, lukewarm, turned off or unnoticed; brand linkage, liking, consideration and next-day recall; channel and segment fit | Second-by-second facial response from 44 FACS Action Units, plus what viewers say afterwards |
| Speed | Minutes | Results in less than 24 hours from the start of fieldwork |
| Best for | Screening many ideas, animatics and rough cuts; choosing channels and segments; deciding what to test | The edit that will carry the media budget |
| Status | Version 0.1, not yet calibrated against real panels | Measured, not predicted |
What it cannot do yet
- It is not calibrated: results have not yet been checked against real panels.
- Memory and forgetting use reasoned parameters, not ones fitted to real recall data.
- Channel use is modelled from approximate UK reach figures, not asked of anyone.
- Music and sound effects are not read yet, so respondents react to the pictures and words only.
- A long film is read in beats of a few seconds, so very short moments blur together.
- A respondent can only link the ad to a brand that appears on screen or in the words.
Where it fits
Use a synthetic audience early, when you have more ideas than budget: to compare animatics, to find where a long film loses people, to see which channels and segments an ad suits, and to check that the brand is clear before you pay for a measured test. For a sense of scale, the report also places the film on a memorability model trained on real next-day recall for 2,181 ads. Then put the two or three strongest edits in front of real viewers, 15 or more per audience segment, and decide on what actually happened. The full sequence is in how to pre-test a video ad, and the methods are compared in what is ad pre-testing?
Synthetic testing questions
How the measured test works is set out in the methodology.
What is synthetic ad testing?
Synthetic ad testing shows an ad to modelled respondents, built from population data, instead of to real people. Each modelled respondent watches the ad and answers questions such as whether they noticed it, whether they liked it and which brand it was for. It is fast and cheap, but it is a model of an audience, not a measurement of one.
How many people can a synthetic audience model?
EchoDepth can model up to 3,000 people for one ad. Results come with ranges rather than single numbers, and any segment with fewer than 30 people is not reported on its own.
Where does the synthetic audience data come from?
Each modelled person starts from census age and gender profiles for the audience you choose. For a profession, sex and social grade follow an approximate profile of that work. Each person is then given a location, a media diet, a level of category use and a relationship with your brand. Which channels they use is modelled from approximate UK reach by age and their media diet, not asked.
Are synthetic respondents accurate?
Only as accurate as their calibration against real people. EchoDepth's synthetic audience is version 0.1 and has not yet been checked against real panels, so we label every result as a modelled pre-test. Use it to screen ideas and decide what to test, not to sign off the edit that carries the budget.
What is the difference between a synthetic audience and a predictive ad score?
A predictive score gives one number for the ad from a model trained on past ads. A synthetic audience models individual people from your audience, each with a profile and a channel they watch in, who answer in their own words straight after and the next day, so you can see who was won over, who was not, and why.
Can a synthetic audience replace testing with real viewers?
Not for the final decision. A synthetic audience cannot show what real faces did in each second, and it reads only the pictures and words of the ad, not the music. EchoDepth uses it before a measured test, to narrow many ideas to the few worth putting in front of real viewers.