Methodology

How FirstReaction simulations work

By Pavan, Founder, FirstReaction · Updated

Short answer

FirstReaction turns your audience description into saved buyer archetypes, then has a multimodal AI model react to your ad as each one, in a fixed format with realistic benchmark ranges. You get scroll-stop, engagement and likelihood-to-act scores per archetype, a projected click-through rate, a creative ranking and prioritized fixes.

What a FirstReaction simulation is

A simulation is a structured prediction of how a defined audience would react to one Facebook or Instagram ad. It's built from three things you give it (an audience, an ad, and your brand's industry) and returns scores, reasons and fixes for each type of buyer in that audience. It doesn't use your ad account, and it doesn't spend anything on Meta.

What goes in

  • An audience. You describe who you sell to in plain words, plus targeting (age, location, interests, platforms) and how warm they are: cold, lookalike, warm, hot or retargeting. FirstReaction turns that into up to 5 archetypes: specific buyer types, each with goals, doubts and buying habits. You can edit them, and they're saved, so every version of an ad is tested against the same people.
  • The ad. Headline, primary text, description, call to action, and up to 10 images and 5 videos (up to 2 minutes each).
  • The brand's industry, which sets the benchmark ranges described below.

How a simulation runs

  1. Videos are watched once. Each video is turned into structured notes (what happens in the opening seconds, pacing, on-screen text and sound) and kept, so the same video reads the same way in every later run.
  2. One model reads the whole ad. A multimodal AI model (one that reads text, images and video) gets the archetypes, the ad copy, the images and the video notes in a single call, and answers in a fixed format: every score, reason and fix has a set place and range. Answers are checked against that format and repaired where they break it, before you see them.
  3. Numbers are anchored to real benchmarks. AI models tend to be optimistic about ads, so each run is given realistic ranges for the audience's temperature and the brand's industry, and asked to place average ads in the middle of them, strong ads at the top and weak ads at the bottom.
  4. Runs are kept steady. The scoring call runs with randomness turned off (temperature 0) and a fixed seed, so the same ad scores close to the same each time. Not always identical, which is why we average several runs for anything we publish.

Benchmark ranges

The starting ranges, before the industry adjustment (an e-commerce brand is a little above them, a finance brand well below):

Audience temperatureShare who stop scrolling (hook rate)End-to-end link click-through rate
Cold10–22%0.3–1%
Lookalike14–26%0.5–1.4%
Warm18–30%0.8–2%
Hot25–38%1.5–3.5%
Retargeting32–48%2–5%

What the scores mean

Scores run from 0 to 100. 70 and above is strong, 40–69 is moderate, and below 40 is weak.

ResultWhat it measures
Scroll-stop probabilityHow likely each archetype is to stop on the ad in a busy feed
Engagement scoreHow strongly the ad holds attention and interest once they've stopped
Likelihood to actHow likely they are to click and take the next step
SentimentWhether the reaction is positive, neutral or negative, and why
Copy scoresHeadline, primary text and call to action, each scored on its own
Engagement funnelHook rate (stop), hold rate (watch or swipe, for video and carousels) and click rate, which multiply into a projected click-through rate
Creative rankingEvery image and video ranked, with strengths, weaknesses and fixes
Suggested improvementsUp to 10 changes, each with a priority and the reason behind it
Policy checkAn advisory check against Meta's advertising policies, not Meta's own review

How to read the results

  • Compare, don't forecast. The scores are most useful side by side: version A against version B, against the same audience.
  • Look per archetype. An average can hide the one buyer type you're losing. The fix for that buyer is often the biggest win.
  • Read the reasons. Each score comes with an explanation; that's what tells you what to change.
  • Confirm live. Take the strongest one or two versions into a real test before scaling.

Your data

Your audiences, ads and results stay private to your account and are never used to train AI models. Nobody else sees a simulation unless you create a share link for it, and you can ask us to delete your account and data at any time. The privacy policy has the details.

Frequently asked questions

Is FirstReaction's predicted click-through rate accurate?

It's calibrated, not measured. Each run is anchored to realistic hook-rate and click-through ranges for your audience temperature and industry, so the numbers land in a believable range. Use it to compare versions of an ad; confirm the winner with a live test.

Why can two runs of the same ad give slightly different scores?

AI models aren't perfectly repeatable. FirstReaction turns randomness off and uses a fixed seed, so scores stay close, often identical, but can move a few points. For important decisions, run each version two or three times and compare the averages.

Does FirstReaction need access to my Meta ad account?

No. A simulation only needs your audience description and the ad itself. Nothing is published or spent on Meta.

Is my ad data used to train AI models?

No. Your audiences, ads and results stay private to your account and are never used to train AI models.