---
title: Message testing: which version actually lands, before you spend on media
description: Test two to five versions of a headline, a value proposition or an ad copy against synthetic personas. What it settles, what it cannot, and how many responses an answer needs.
---

# Message testing

Five versions of a headline, and everyone in the room has a favorite. Message testing replaces the room with the people the message is written for: you put the versions in front of a synthetic population calibrated on public US data, and you read which one is understood, believed, and acted on. Before the media budget, not after.

## What a message test settles

A message test is not a beauty contest. It answers three questions that come up in the same order in almost every launch.

### What it says

Does the reader restate your proposition in their own words, and is that restatement the one you intended? A message that needs a second sentence to be understood has already lost half its audience.

### Whether it is believed

A credible claim and a claim that sounds like marketing are read differently, even when both are true. The gap between "I would consider this" and "I do not buy it" is usually in the wording, not in the offer.

### Which one to run

Between two versions, the useful output is not a winner in the abstract. It is the segment where the order flips, and the sentence that carries the difference.

## How a test runs here

The work is in writing the versions. Once they exist, a run takes minutes and the report arrives with the distributions and the verbatims.

| Step | What you do | What you get |
| --- | --- | --- |
| 1. Write the versions | Paste two to five variants of the same asset: headlines, a value proposition, an email subject line, two ad copies. | A survey where each persona reads the versions in the same order. |
| 2. Set the population | Choose the audience the message is for: age, income, region, or the whole panel. | An audience calibrated on the American Community Survey, not on whoever answered first. |
| 3. Ask the follow-ups | Add the questions that make the preference usable: what it means, what it promises, what it would take to act. | A report that separates comprehension, credibility and intent, instead of one average. |

## What to ask, beyond "which one do you prefer"

A preference question alone gives you a ranking with no reason attached. Four questions turn it into a decision.

- **Restate it.** "In one sentence, what is this offering?" The answers show whether the version survives being read quickly.

- **Believability.** "How believable is this claim?" on a five-point scale, plus the reason in the persona's own words.

- **Intent.** "How likely are you to act on this in the next month?" A preference with no intent behind it is a compliment, not a signal.

- **The objection.** "What would stop you?" This is where the next version is written, and it is usually the most useful block of the report.

## How many versions, how many responses

Five versions is a comfortable ceiling: beyond that, the survey gets long and the reader stops reading carefully. For the sample, the rule of thumb is the same as for any proportion: a rough read needs around 25 responses per segment, a comparison between two versions needs around 100, and a claim about a sub-segment needs several hundred.

The [sample size calculator](/us/sample-size-calculator) gives the number for a given margin. On the free plan you can test one question for nothing; the paid plans exist for the rhythm, not for the number of personas per run. The [cost page](/us/cost-of-market-research) lays out what each plan covers.

## What it does not do

This is a synthetic population, calibrated on public statistics. It tells you how a message is read by an audience that resembles the one you target, which is what a pre-launch decision needs. It does not measure the effect of a media plan, it does not replace an A/B test on live traffic, and it does not give you a margin of error as if it were a probability sample. The [reliability page](/us/fiabilite) states what the engine does and does not guarantee, and it is worth reading before you quote a number from a report.

## Test your first pair of versions

The shortest path is to take the two versions already on the table, paste them into one survey, and see whether the room's favorite survives contact with the audience. It takes minutes, and the free plan is enough to start.

[Run a message test](/us/login)

Related: [how the synthetic population works](/us/personas-synthetiques), [concept testing](/us/concept-testing), [pricing research](/us/pricing-research), and the [plan comparison](/us/comparatif).