Free synthetic audience ad test
The synthetic audience ad test shows how an ad might land with a cohort of 1,000 virtual people before you spend on it. Paste the ad, say what you sell and who it is for, and pick how warm the audience is. Each virtual person scrolls past, stops, taps or intends to act, and says what holds them back. You get a funnel per 1,000 virtual impressions, results by profile and the main objections. Free, no signup, a signal and not a forecast.
Each virtual person answers alone, so this is not a group discussion. Version 1 reads text only. Last updated 15 November 2026.
Test your ad
A signal, not a forecast. Version 1 tests the text of one or two ads. Images and video files are not uploaded.
Want real people instead? Human research from $1,200.
How the test works
The ROIMOBI synthetic audience ad test is a free tool that runs one ad, or two, past 1,000 virtual people. Each person scrolls past, stops, taps or intends to act, and says what holds them back. The tool returns a funnel per 1,000 virtual impressions, seven readings of the creative, the main objections and a view by profile. It is a signal and not a forecast of CTR or sales.
1Every person sees the ad once.
21,000 virtual people answer, each one alone.
3Code counts the people, applies the weights of the audience mix and does every sum.
4You get the funnel per 1,000, the objections and, for two ads, a verdict.
Fig 1 · Illustration with example numbers
- Scrolls past 359
- Stops 321
- Taps 282
- Intends to act 38
- 1.The cohort1,000 virtual people, shown as 1,000 dots, one virtual impression each. Each has a profile from the factor grid.
- 2.The ad is shownEvery person sees the ad once. They answer alone, and no one sees another person's answer.
- 3.The sortEach dot moves to the furthest step its person reaches: scrolls past, stops, taps or intends to act. The ones who hold back say why.
- 4.The readYou get the funnel per 1,000, the readings of the creative, the objections and the table by profile.
Jev, a decision model by TypeSafe AI, answers one narrow question per person: how far does this person go with the ad, from scrolling past to tapping with the intent to take the offer. The answer is a probability for each step. Code counts the people, applies the weights of the audience mix and does every sum. The same 1,000 people see both ads, so a difference between two ads is not a difference between two audiences. See the run on 54 live ads for how Jev is used on this site.
Who the virtual people are
Every virtual person is generated from a visible factor grid, so you can see who answered. The grid has 1,044 distinct profiles: 1,008 for people who are in the target and 36 for people outside it. The panel is 1,000 of them, picked with a fixed seed, and it is the same for every ad. The grid uses situations and attitudes only. It has no age, gender, income or ethnicity, because models flatten identity groups.
| Factor | Options | Why it is in the grid |
|---|---|---|
| Fit with the audience | Core audience, Close to the audience, Outside the audience | Separates the people the ad is for from the people it only reaches, and gives the control check its group. |
| Stage | Unaware of the problem, Feels the problem, Comparing solutions, Uses a competitor, Tried and gave up, Knows the product, Waiting for an offer | The same ad reads differently to someone who has not met the problem and someone waiting for an offer. |
| Urgency | Needs it within weeks, No deadline | A deadline changes how much reason a person has to act now. |
| Price attitude | Tight with money, Pays for clear value, Pays easily | Drives objections to the price and to the commitment. |
| Decision style | Decides on the spot, Researches first | Quick deciders act on the ad itself, careful ones look for more before they pay. |
| Trust in ads | Distrusts ads, Neutral toward ads, Open to ads | Sets how much a claim in the ad has to carry on its own. |
| Moment | Scrolling fast, Browsing with time | Sets the chance to stop at all. |
The audience mix sets how much each person counts. A cold mix has many people who have not looked for a solution, a warm mix has more people who are comparing, and retargeting has mostly people who know the product. The weights below are starting values chosen for this tool. They are assumptions, not measurements of any real audience, and the tool lets you switch the mix after a run without a new request.
Cold audience
Assumption, not measured. Stage and urgency count target people only.
- Fit with the audience
- Core audience 30%, Close to the audience 40%, Outside the audience 30%
- Stagetarget people only
- Unaware of the problem 35%, Feels the problem 28%, Comparing solutions 14%, Uses a competitor 9%, Tried and gave up 6%, Knows the product 5%, Waiting for an offer 3%
- Urgencytarget people only
- Needs it within weeks 25%, No deadline 75%
- Price attitude
- Tight with money 25%, Pays for clear value 50%, Pays easily 25%
- Decision style
- Decides on the spot 35%, Researches first 65%
- Trust in ads
- Distrusts ads 30%, Neutral toward ads 50%, Open to ads 20%
- Moment
- Scrolling fast 65%, Browsing with time 35%
Warm audience
Assumption, not measured. Stage and urgency count target people only.
- Fit with the audience
- Core audience 60%, Close to the audience 30%, Outside the audience 10%
- Stagetarget people only
- Unaware of the problem 5%, Feels the problem 20%, Comparing solutions 25%, Uses a competitor 12%, Tried and gave up 8%, Knows the product 20%, Waiting for an offer 10%
- Urgencytarget people only
- Needs it within weeks 25%, No deadline 75%
- Price attitude
- Tight with money 25%, Pays for clear value 50%, Pays easily 25%
- Decision style
- Decides on the spot 35%, Researches first 65%
- Trust in ads
- Distrusts ads 30%, Neutral toward ads 50%, Open to ads 20%
- Moment
- Scrolling fast 65%, Browsing with time 35%
Retargeting
Assumption, not measured. Stage and urgency count target people only.
- Fit with the audience
- Core audience 75%, Close to the audience 20%, Outside the audience 5%
- Stagetarget people only
- Unaware of the problem 0%, Feels the problem 5%, Comparing solutions 15%, Uses a competitor 8%, Tried and gave up 7%, Knows the product 40%, Waiting for an offer 25%
- Urgencytarget people only
- Needs it within weeks 25%, No deadline 75%
- Price attitude
- Tight with money 25%, Pays for clear value 50%, Pays easily 25%
- Decision style
- Decides on the spot 35%, Researches first 65%
- Trust in ads
- Distrusts ads 30%, Neutral toward ads 50%, Open to ads 20%
- Moment
- Scrolling fast 65%, Browsing with time 35%
What you get back
- The dot grid: 1,000 dots sorted into scrolls past, stops, taps and intends to act, with the count per 1,000. The label under it reads: panel answers, not delivery metrics.
- A mix switch that re-weights the same answers in your browser, with the weights in use.
- For two ads, a verdict (B ahead, A ahead or too close to call), the differences per 1,000 and the share of people each ad wins.
- Seven readings of the creative: hook strength, follow-through, control check, selectivity, sceptic gap, price gap, cold pull. Each comes with a sentence on how to read it.
- The objections as a bar list with the top objection named.
- A table of taps by profile, labelled as model readings. With an email you also get the stop and intent tables and a text and CSV download.
- The cost of the run: questions asked, input tokens, US dollars and seconds.
The result never shows a predicted CTR, conversion rate, cost per acquisition or return on ad spend, and it gives no match percentage with human research.
A worked example
Sample product and ads written for this page. Not a real advertiser or client. The numbers below are the real output of the test for these two ads: Jev (jev-1.14.0) answered on 15 November 2026, and code did the counting under the cold audience mix. Try an example in the tool above to see the same result with the mix switch.
- What is sold
- A Spanish learning app with ten minute lessons and a voice coach. First week free, then $9.99 a month.
- Who it is for
- Adults who plan a trip to a Spanish-speaking country in the next few months.
- Ad A: primary text
- Speak Spanish on your next trip, guaranteed! Ten minute lessons built around the phrases travellers use most, with a voice coach that checks your accent. Start today.
- Ad A: headline
- Learn Spanish fast
- Ad A: button label
- Learn more
- Ad A: offer line
- First week free, then $9.99 a month
- Ad B: primary text
- Ten minute Spanish lessons for travellers. Learn the phrases for the airport, the hotel and the table. Start your free week.
- Ad B: headline
- Travel Spanish in 10 minutes a day
- Ad B: button label
- Start free week
- Ad B: offer line
- First week free, then $9.99 a month
| Step | Ad A | Ad B | B minus A |
|---|---|---|---|
| Scrolls past | 359 | 302 | -57 |
| Stops | 641 | 698 | +57 |
| Taps | 320 | 366 | +46 |
| Intends to act | 38 | 81 | +43 |
Stops, taps and intent count everyone who got that far, so each is a part of the one before. Panel answers, not delivery metrics.
Ad A
- Scrolls past 359
- Stops 321
- Taps 282
- Intends to act 38
Ad B
- Scrolls past 302
- Stops 332
- Taps 285
- Intends to act 81
Verdict: B ahead. Ad B is ahead on both taps and intent, with +46 taps and +43 intents per 1,000 for B minus A. Ad B goes further than ad A for 80.6% of the virtual people, and ad A goes further than ad B for 4.8%. The top objection for ad A is no reason to act now (67.5% of the people asked), and for ad B no reason to act now (81.1%).
| Reading | Ad A | Ad B |
|---|---|---|
| Hook strength | 67.8% | 75.9% |
| Follow-through | 14.9% | 26.5% |
| Control check | 14.9% | 17.1% |
| Selectivity | 3.21× | 3.18× |
| Sceptic gap | 0.37× | 0.54× |
| Price gap | 0.32× | 0.48× |
| Cold pull | 0.69× | 0.69× |
The control check is the tap share among people who have no need for the product: 14.9% for ad A and 17.1% for ad B. Both are under 30%, so the panel is not flagged as generous for these ads. Even so, the level of any one number is a panel answer, and the gap between two ads is the safer reading. The ad rewrite changed the words and kept the product, so the gap comes from the ad. A live test still decides, because a virtual tap is not a click.
Cost of this run: 1,000 virtual people per ad, 2,484 questions asked, 406,980 input tokens, $0.0171, 0.9 seconds.
Method and limits
- It is a signal and not a forecast. Virtual taps are not clicks, and the funnel per 1,000 virtual impressions does not predict CTR, conversion or sales.
- Synthetic respondents lean positive. In our own checks, people outside the target still tapped, at 14 to 21 percent in an early test, which is far above real click rates. The control check shows this on every run and adds a note when it passes 30%. Read the gap between two ads, not the level of one.
- The profile table is a model reading of described profiles. It is not a finding about real groups, and the published work on synthetic respondents warns that models misportray and flatten them.
- Repeat runs of one ad moved the funnel by up to 5 per 1,000 when measured on 2026-11-15. The verdict says too close to call unless the gaps are larger than that.
- The audience mixes are assumptions. The weights are printed so you can judge them.
- Each virtual person answers alone. The test does not show how opinions spread in a real group, and it does not read images or video.
- The model version is pinned to jev-1.14.0 and printed with the example, because model behaviour can drift over time.
For the evidence behind these limits, see LLMs that replace human participants can misportray and flatten identity groups (Wang, Morgenstern and Dickerson, 2025) and Synthetic Replacements for Human Survey Data? (Bisbee and others, 2024). The method is also published as an open-source agent skill: the open-source method and synthetic-audience-panel on GitHub.
Synthetic panel, real panel, live test, other tools
| Option | Good for | Limit |
|---|---|---|
| This test | A fast filter on wording and offer fit, funnel, objections and two ads on one panel, free | Virtual people and a model's reading of an ad, with no real behaviour |
| Real human panel or interviews | Real reactions and follow-up questions | Slower and paid: the Human research service starts from $1,200 |
| Live A/B test with spend, for example a platform's own split test | Real clicks and real conversions | Needs budget and time, and costs money on weak ads |
| Other synthetic panels, for example Admetrics Panels | A second opinion from a different method | Methods differ, so compare outputs rather than trust one |
| Chat assistant prompt such as act as 100 customers | Free-form feedback | Answers vary with the wording and the run, and there is no fixed factor grid |
| Enterprise ad pre-testing, for example Kantar or System1 | Benchmarks and human data | Quote-based and built for large brands |
The options do different jobs. Use this test to choose which ad to try first and which objection to answer, then confirm with real people or a live test with a small budget. To read the ad text itself, use the free ad creative checker.
Want real people to read the ad?
A senior growth practitioner can read your market and your ads by hand and hand you decisions and briefs. The offer, its scope and its limits are described on the ad checker page, from $1,200, excluding taxes.
Questions about the synthetic audience ad test
- What is a synthetic audience?
- A set of virtual people described by stated traits, used as a stand-in for a target group. Here code builds 1,000 people from a visible grid of factors, and a decision model answers for each one on their own.
- Is a synthetic focus group the same as a real focus group?
- No. A real group has people who talk to each other and a moderator who reads the room. Here each virtual person answers alone, nobody influences anyone, and the result is a signal on how far each person would go with the ad.
- How accurate is an AI focus group for ads?
- Vendors publish match percentages against human research, and those figures come from their own tests. This page publishes none. Treat the result as a filter that compares wording and surfaces objections, then check it with real people or a live test.
- Can it predict my CTR or sales?
- No. Virtual taps are not clicks, and the funnel per 1,000 virtual impressions is not a forecast. Real results depend on the platform auction, the audience you target, the landing page and the budget.
- Is it really free?
- Yes. There is no account. The example run on this page, two ads on 1,000 people each, cost $0.0171 in model fees, and the page shows the cost of your run. The full report is optional and needs an email.
- How many virtual people are in a test?
- 1,000. They are drawn from a grid of 1,044 distinct profiles with a fixed seed, and the same 1,000 see every ad. The audience mix changes how much each person counts, not who is asked.
- Can I test two ads against each other?
- Yes. Both ads go to the same 1,000 people, so the difference between them is not a difference between two audiences. The page gives a verdict only when the gaps are large enough to read.
- Why do the virtual people object to my ad?
- Each person answers from a profile, so objections show which kind of person stops and why: price, trust, no reason to act now, an unclear offer, or this is not for me. Treat repeated objections as the useful part.
- How is it different from asking a chat assistant to act as 100 customers?
- A chat assistant answers in prose and the answer can change with the wording and the run. Here the profiles come from a fixed grid, the decision model answers typed questions with probabilities, and every ad meets the same people.
- Does it read images or videos?
- Not in version 1. Paste the text, and for video the first three seconds as words.
- Is my ad text stored?
- No. Your text, product and audience go to Jev, a model run by TypeSafe AI, to answer the questions. ROIMOBI does not store them, log them or send them to analytics. For the vendor's own handling of data, read TypeSafe AI's legal page.
Let’s make your growth add up
Tell me about your app and where growth is stuck. You can start with an audit, then move to a build and a run. Pick the step that fits.