See which draft would win, before you send it.
- ControlWinter Sale: 30% Off All OuterwearSimulating…
- Draft 1Your coat's been waiting. 30% off ends SundaySimulating…
- Draft 2The jacket you almost bought is 30% offSimulating…
- Draft 3Winter savings are hereSimulating…
Predicted clicks vs the control. Simulated by Neural 2.1, trained on years of real send results.
A/B Simulate runs your A/B test in simulation, inside Claude, ChatGPT or another AI assistant. Your assistant sends it 2 to 20 versions of a subject line, SMS or ad headline, your current copy first as the control. Neural 2.1 ranks them head to head and predicts each one's lift over the control. You send the winner.Built on Neural 2.1, trained on billions of real sends.
Just good copy.
Every send gets an A/B test, in simulation, including the ones you'd never have tested. The predicted losers stay in the chat, real tests only carry variants worth sending, and the sends you can't test still go out with the winner.
The winner you send- 01
Run the test you'd otherwise pay for.
A real A/B test spends half your list on the losing line, then makes you wait for the answer. A simulation costs one credit, answers in the chat, and the predicted loser never reaches an inbox.
Winter savings are here#4 of 4 · −6% vs control - 02
Send the winner, every time.
Every campaign gets the same head-to-head, including the ones written at 5pm on a Friday.
- 03
Only test what's worth testing.
When you do run a real test, put your top two in it. They've already beaten the rest in simulation, so both halves of your list get a line worth sending.
- 04
Get a read when you can't test.
Most copy never gets a fair test. The list is too small, the send is a one-off, or the headline is live before results could come back. A/B Simulate runs the test anyway, in simulation.
Small listsOne-off sendsSMS and pushWeb and ad headlinesLast-minute sends - 05
Decide in the chat, not next week.
No waiting on results to pick a line. The predicted winner turns up where you're already writing.
Skip the test. Or make it count.
A real A/B test is only as good as the lines in it. Put a weak one in, and part of your list gets the weaker email so you can learn what a simulation would have predicted. Simulate first, then test your top two. Or skip the test and send the winner.
Half the list gets the weaker line, and you wait a week to find out which half.
Both halves get one of your two predicted best lines, and the real test settles which is better.
Ask. Simulate. Send the winner.
A/B Simulate lives inside the assistant you already write with. There's nothing to install on your email platform, and nothing new to learn. Developers can run the same simulation through our REST API.
- #1
Ask for a few versions
"Give me five subject lines for the spring drop, and test them against last year's." Your assistant writes the variants and puts your current copy first, as the control.
- #2
A/B Simulate runs the test
Neural 2.1, a model trained on years of real send results, simulates an A/B test between 2 and 20 variants. It ranks them head to head and predicts each one's lift over the control.
- #3
Send the winner
Copy it, or have your assistant build the campaign in your email platform. Ask it to share the results afterwards and see how the prediction held up.
Sharing is free on every plan
Every channel. Head to head.
From subject lines to push to homepage headlines: if someone has to click it, A/B Simulate can simulate the test. Sometimes your control wins. That's worth knowing too.
- Email subject linePredicted clicks
- The jacket you almost bought is 30% off+18% vs controlWinner
- Your coat's been waiting. 30% off ends Sunday+9% vs control
- Winter Sale: 30% Off All OuterwearControl
- PreheaderPredicted clicks
- Plus free returns, and your size is still in stock.+7% vs controlWinner
- Shop the sale now.Control
- SMSPredicted clicksProvisional · order only
- Northfold: 30% off outerwear ends at midnight. Reply STOP to opt outWinner
- Northfold Winter Sale: 30% off all outerwear. Reply STOP to opt outControl
- PushPredicted opens
- Back in stock: the Harbor parka, in your size+11% vs controlWinner
- New arrivals are hereControl
- Ad headlinePredicted clicksProvisional
- Warm coats, built to last. Free shipping over $75+6% vs controlWinner
- Shop Winter Coats OnlineControl
- Web headlinePredicted clicksProvisional
- Coats for the long way homeControlWinner
- The Winter Collection Is Here−5% vs control
- Social postPredicted clicksProvisional
- Tag someone who still wears their coat from 2016. It's time.+9% vs controlWinner
- Our winter sale is live.Control
Lift is predicted against the control, the first variant you send. Email subject lines, preheaders and push are calibrated. Other channels are provisional, since Neural 2.1 isn't calibrated for them yet, and SMS shows the order only.
Why not just ask your AI which one wins?
You can. It'll pick one and sound sure about it. But it's marking its own homework, and it learned from the internet, not from what subscribers actually click.
"Go with B. It feels warmer, and ‘savings’ sounds friendlier than a hard discount in capitals."
Knows your brief, your offer and your tone. Has never seen your click-through rates.
- Winter Sale: 30% Off All OuterwearControlWinner
- Winter savings are here−6% vs control
Keep your control. B is predicted to get 6% fewer clicks than the line you already send.
Neural 2.1 learned from billions of marketing sends. It has one job: predicting what gets clicked.
Head to head: A/B Simulate's predicted winners against the assistant's own picks, on campaigns the model has never seen. The result goes here once it's independently checked.
Tell it how the send did.
After a campaign goes out, ask your assistant to share the results with A/B Simulate: the clicks and the total sent, from whichever email platform you use. A/B Simulate shows how each line really did next to its simulated rank. Ran a real A/B test on your top two? Share both, and see whether the predicted winner won.
- On every plan, and sharing results doesn't use credits
- Your assistant fetches the numbers. A/B Simulate never connects to your email platform
- Campaign totals only. Never subscriber lists, recipient addresses or individual events
Priced for people who send on a Friday.
Try it on this week's campaigns.
- 10 credits a month
- 1 brand, 1 seat
- Works in Claude and ChatGPT
- Compare predicted winners with real results
- Simulation history and CSV export
For one brand, sending every week.
- 100 credits a month
- 1 brand, 1 seat
- Works in Claude and ChatGPT
- Compare predicted winners with real results
- Simulation history and CSV export
For teams and multi-brand stores.
- 300 credits a month
- 3 brands, 3 seats
- Works in Claude and ChatGPT
- Compare predicted winners with real results
- Simulation history and CSV export
For agencies running client sends.
- 1,000 credits a month
- 15 brands, 10 seats
- Works in Claude and ChatGPT
- Compare predicted winners with real results
- Simulation history and CSV export
Fifteen brands. One login.
Switch brands mid-conversation, keep each client's simulations and results separate, and send them a report that shows your subject lines are backed by data, not vibes.
A/B Simulate for agencies →An A/B test between your drafts. Send 2 to 20 variants of a subject line, SMS, push or headline, with your current copy first. Neural 2.1, trained on billions of real sends, ranks them head to head and predicts each one's lift in clicks over your current copy (opens, for push). You get the predicted winner before anyone on your list sees a thing.
Open our listing in Claude's connector directory, choose Connect, then approve it on A/B Simulate. ChatGPT takes a few more steps, and our How to connect page walks you through both.
2 to 20 in one simulation, and it's 1 credit whatever the number. A single line can't be simulated, because there's nothing to test it against. Your assistant will write a few alternatives and try again, and the refused request doesn't use a credit.
The first variant you send. Put your current or default copy there: the line you'd send if you didn't test. Every other variant's predicted lift is measured against it. Tell your assistant which line you'd normally use and it puts that one first.
Neural 2.1 scores every variant against the control and ranks them by predicted clicks (opens, for push). The winner is the variant with the highest predicted lift, and there's always a single order: if two variants tie exactly, the one you listed first ranks higher. It's a prediction, not a guarantee, so treat a narrow lead as a close call.
1 simulation = 1 credit, whether it has 2 variants or 20. Free includes 10 credits a month, and paid plans include 100, 300 or 1,000. Plan credits reset on the first day of each month (UTC) and don't roll over.
No. Simulations work straight away, and A/B Simulate never connects to your email platform itself. If your assistant can reach it, the assistant can share how each send did. That's free on every plan.
We store the variants you simulate and their predicted rank and lift, plus any aggregate results you choose to share: clicks and total send volume, and opens and revenue if included. We don't ask for subscriber lists, recipient addresses or individual events, and we don't currently use your copy or results to train models.
Simulate the test.
Send the winner.
Add A/B Simulate to Claude or ChatGPT and simulate your next send.