Evaluate our predictive intelligence on your own A/B tests
Send us past A/B tests with the results removed. Our model scores every variant. You compare its ranking with the outcomes you observed.
- 40+A/B tests
- 2+variants per test
- 0results shared with us
- 1workbook, sent by email
- 1Your spring picks are 20% off until Sunday Predicted first
- 2Spring savings end Sunday: shop now
- 3Don't miss our biggest spring sale
Predicted first = observed winner
You hold the outcomes. The model scores without them.
The comparison is only valid if the outcomes are withheld until the scores have been delivered. Everything happens in one workbook that you keep.
Step 1: Fill in the workbook
On the “Enter messages” sheet, add one row per variant and name its campaign. Leave out every result.
Step 2: Send it back
Reply to the person at Persado who reached out to you, with the workbook attached. Leave the “Score vs Response” sheet empty.
Step 3: We return it scored
Every message scored five times, within its own campaign, in about two business days.
Step 4: Enter your outcomes
Add unique clicks and messages delivered for each message in “Score vs Response”. The workbook computes click-through rate and the three metrics.
- Scoring
- Each message scored five times, compared only within its own campaign
- Winner test
- Two-proportion test on each campaign’s top two messages; close calls are set aside
- Metrics
- Top-1 accuracy, Spearman rank correlation, run-to-run stability
- Workbook limits
- 40 to 300 campaigns, up to 1,200 messages, one channel and one industry
Your workbook back, scored, with metrics that calculate themselves
- A Persado score for every messageFive runs per message, shown side by side so you can check the ranking is stable.
- A results sheet you completeEnter unique clicks and messages delivered. It computes click-through rate, tests whether each campaign has a clear winner, then calculates top-1 accuracy, Spearman rank correlation and every campaign where the model was wrong.
- A worked example to followA filled-in sample sheet shows what a finished evaluation looks like before you start.
| Row1 | Message | Persado score | Clicks · delivered |
|---|---|---|---|
| 2 | Spring sale reminder | ||
| 3 | Your spring picks are 20% off until Sunday | 1 | you enter these |
| 4 | Spring savings end Sunday: shop now | 2 | you enter these |
| 5 | Don't miss our biggest spring sale | 3 | you enter these |
| 6 | Webinar invite | ||
| 7 | Save your seat for Thursday's live session | 1 | you enter these |
| 8 | Join us live on Thursday at 1pm | 2 | you enter these |
Outcomes stay with you. We never ask for them.Rank shown in place of score
On 2,086 A/B-tested campaigns, the model identified the observed winner in 87.5% of cases
Flagship general-purpose LLMs, given the same messages and the same number of runs, identified it in 51% of cases. Your evaluation follows the same protocol and reports the same three metrics.
Top-1 accuracy: the variant ranked first was the observed winner
Top-1 accuracy on the same campaigns and runs
Top-1 accuracy
Share of campaigns where the variant ranked first was the observed winner.
Spearman rank correlation
Agreement between the predicted order and the observed order of all variants.
Run-to-run stability
Share of campaigns given an identical ranking on every repeat run.
PersadoGeneral-purpose LLMs
- Dataset
- 2,086 campaigns, each with two or more variants and a recorded outcome
- Period
- 31 May 2025 to 15 May 2026
- Comparison
- Gemini 3.1 Pro, GPT 5 and Claude Opus 4.8, scoring the same content over the same runs
- Published
- September 2026
Run the same evaluation on your platform's tests.
Download the workbookAn accurate ranking moves the decision from after the send to before it
An A/B test identifies the stronger variant only after part of the audience has received the weaker one. A model that ranks variants accurately when they are written gives your users that answer earlier.
Ranking at the point of creation
Users see a predicted order for their variants inside the tool they already use, before anything is sent.
Fewer sends spent on testing
Users can keep testing where variants are close and spend fewer sends confirming clear differences.
A measure of content you don't see today
Most content sent through your platform is written elsewhere. A ranking gives you a quality signal on it, and gives your users feedback.
What we receive, and what we do with it
Campaign names and message text, plus the account, channel and industry you enter at the top of the workbook. No recipient data and no results.
- Used only for your scores
The content is not used for any other purpose.
- Not used for training
Submitted content does not enter our training data.
- Deleted after 30 days
Or sooner, on request.
- Outcomes stay with you
You enter them only after the scored workbook is back in your hands.
Questions about the evaluation
Other questions can go in the email you send the workbook with.
Download the workbookThe scores have to be produced without them. If we saw the outcomes, the comparison would show nothing. You keep them and run the comparison yourself, in the workbook.
Tests where the variants ran at the same time, to comparable audiences. There is no need to filter out close calls: the workbook runs a two-proportion test on each campaign’s top two messages and sets aside any campaign where they can’t be separated.
Send what you have. Below 40 campaigns the metrics are less stable, so treat them as indicative. 40 or more gives a more reliable estimate, and 50 or more is better.
Only after the scored workbook is back with you: unique clicks and messages delivered (or impressions) for each message. The workbook computes click-through rate the same way for every campaign. Use the same basis for every variant in a campaign, and avoid open rates, which automatic tracking makes unreliable.
No. Each workbook holds one channel, such as email subject lines, and one industry, so the results are comparable. Send a second workbook for another channel or industry.
Reply to the person at Persado who reached out to you, with the completed workbook attached. Preparing 40 tests takes about 15 minutes if the results are already in a report, and the scored workbook comes back within two business days.
Evaluate the model on your own data
Download the workbook, add 40 past A/B tests without their results, and send it back to the person at Persado who shared this page.