Free audit

For CRM and growth teams

Know which of your A/B winners were real.

AheadWise re-checks every push, WhatsApp, SMS and email test your platform called, says how sure each winner is, and sizes the next test so it can actually tell. Where a sealed back-test on your own history earns it, it also ranks copy before you send.

Audit your past tests, free
  • 9 of 47recorded winners in a public sample are probably the worse copy (likely 4 to 15), and nobody can say which. Only 3 of the 47 were proven.
  • 0files uploaded. The audit runs entirely in your browser.
  • 64%of all decided tests, and 76% of clear winners, picked in a held-out rehearsal on 300 public US headline tests. The best public models score about the same. Sealed back-tests on client copy so far: 0.

Winner check

Is your A/B winner real?

VariantSentClickedRate
AFree delivery ends tonight –
BYour cart is waiting –
Unproven

Unproven.

Same statistics as the full audit. Check a whole export

For YouTube creators

Send me your YouTube title tests and I'll pick the winners, free.

See how it works

The problem

Your platform sends a winner. Nobody checks it.

CleverTap, WebEngage, Netcore and Insider One's email tests send whichever variant has the higher rate when the test window closes, with no significance test. Braze, Klaviyo, MoEngage and Airship do test, but Braze still sends and labels a "Winner" that failed its test, and Klaviyo sends its winner when the window ends. We picked the higher rate in one headline pair from each of 300 public Upworthy tests: two in three of those picks could not be told apart from chance.

Sending the higher rate is a fair bet. Writing it into a playbook as a lesson is not: unproven wins teach your team noise. Your own tests may be better or worse; the free audit tells you in two minutes.

How far from chance is each winner?

  • Real
  • Borderline
  • Unproven
  • Uneven split
Too close to call Real 0 4 6 8 1.96 2.5 ← can't tell apart |z|, distance from chance

Each dot is one of 48 real tests in the Upworthy Research Archive, the sample the audit opens with: the winning headline against the runner-up, as a platform would pick it. Further right means less likely to be chance. Left of 1.96, the counts can't tell the two apart, though the higher rate is still the better bet.

See which of your winners were real

How it works

Prove it on your own history first. Pay after.

  1. 1Free

    Audit your past tests

    Upload the A/B export from your engagement platform. Every test's winner is re-checked from its own counts, entirely in your browser.

  2. 2Free

    Run a blind back-test

    A pre-flight first reads only your counts and says whether a back-test can give a clear answer. Then AheadWise seals a prediction for each past test before reading any result, and scores every call against what actually won.

  3. 3Paid

    Rank copy before every send

    Once your back-test passes, send it your variants and get a ranking with a label on every call: leaning, or test it live, and confident only once your own history has earned it. Every test you run sharpens the next call.

Evidence

Tested on public data. Not yet on yours.

Every accuracy figure here is held out: AheadWise learned from one part of the data and was scored on another it never saw. Held out is not sealed: we ran these ourselves, on public data. Sealed back-tests on client copy so far: 0. We will publish every one, pass or fail.

Clear winners picked

A random 400 of 2,526 held-out US headline A/B tests (AheadWise on all 2,526: 76.1%). Every judge saw the same 400 pairs; each Claude model is shown at its best set-up.

AheadWise78.2%
Claude Opus 5.5shown 80 examples70.3%
Claude Fable 5.166.9%
Claude Opus 564.4%
Claude Sonnet 553.4%
Claude Haiku 4.549.9%
Every gap to a Claude model shown is statistically significant. The best public models are level: Gemini 3.1 Pro picked 79.5% on a different 1,788 pairs from the same archive.
64%/76%
of all decided tests, and of clear winners, in a full product rehearsal on 300 held-out tests. We always quote both.
79%
of its most confident calls (the top 30%) were right across all decided tests, against 64% for all its calls. On the clear-winner tests alone that rises to 97% (the best Claude model's most confident: 80–83%).
81%
of the time the observed winner was among the 3 variants it shortlisted, against 65% for 3 picked at random, across 1,500 held-out tests averaging 4.7 variants. On tests with a statistically clear best: 93%.
1 in 80
identical A/B tests is called "real" by the audit, across 240,000 simulations read once at their planned length (about 1 in 33 if checked daily and stopped when it looks good). That isn't the share of "real" verdicts that are wrong: "real" also needs the winner to be the better copy at least 19 times in 20, given how much your versions usually differ. Counts that repeat people, and Meta ads, are set aside, not judged.

Outside the US it is level, not ahead. On 400 posts from Indian subreddits, AheadWise picked 66.5% of clear winners; Claude Opus 5.5, shown 80 of that audience's past results, picked 65.9%. Arabic, Bahasa, Thai and Vietnamese copy is not measured yet: no public A/B results exist in those languages, so your back-test is the first measurement.

What it's worth

Choosing better is money you already spend.

Every campaign is a choice between versions. Put in your own numbers to see what choosing better is worth on top of the tests you already run, and when your tests are already enough on their own.

How much your versions differ in purchases, not in their best result: the free audit measures it from your own counts. Remembered winning lifts overstate it.

Expected values from a model of your campaigns, not a promise. It assumes AheadWise picks the better of two versions 62% of the time, as it did on public click tests (range shown: 55–68.5%; it is not yet measured on your copy), that a version's click advantage only loosely predicts its purchase advantage, and that your team's own pick is a coin flip. If most of your versions perform alike, expect about half. Your blind back-test measures AheadWise on your own copy before you pay.

In your workflow

Fits the way your team already works.

Starts from the export you already have.

Presets read each platform's own column names. Any other CSV works by picking the columns. Holdout groups, winner sends and total rows are left out and counted, blank or unreadable numbers are never read as zero, and template variables are read as variables, not as copy.

  • MoEngage
  • CleverTap
  • Braze
  • Insider One
  • WebEngage
  • Netcore
  • Klaviyo
  • Airship
  • Meta Ads
Detected: Insider One app push reportExample
TestCampaign IDFound
VariantVariant IDFound
Message textCampaign MessageFound
SentDeliveredFound
ClicksOpensFound
GoalConversionFound
Template variables{{1}} · {{first_name}} · RM0.00Handled

Wherever you send

Made for the markets you send to.

Your platform's export, your customers' languages, and the weeks when they stop behaving normally: Ramadan, Diwali, 11.11, payday.

Reads your platform's export

MoEngage, CleverTap, Braze, Insider One, Klaviyo, Airship, WebEngage, Netcore and Meta Ads exports load as they are. Any other CSV works by picking the columns.

Reads copy as it's written

Arabic, Hindi, Bahasa, Thai and Vietnamese are read correctly, down to Arabic digits, right-to-left marks and the labels on promotional SMS. Accuracy is measured in English so far; your back-test measures it in yours.

Knows your calendar

Tests that ran in Ramadan, Diwali, a national day or a big sale are flagged, and your back-test is scored inside and outside those weeks.

Goals, not just clicks

Where your export has purchases or leads, the audit says whether each click winner also won on them. A version can win the tap and lose the sale.

Pick a market

Calendars for the Gulf, Southeast Asia and India.

Calendars cover 13 markets so far, and we add yours on request; the audit and back-test work anywhere. WhatsApp prices are Meta's rate per marketing message from 1 October 2026; "est." marks dates not yet announced.

Security

Built for teams that must know where their data goes.

  1. 01

    The audit never uploads

    Your file is read by code in the page and never sent. The page's security policy blocks uploads, fetches, beacons, form posts and anything loaded from another site; you can check that in your browser's developer tools.

  2. 02

    A few columns, no personal data

    AheadWise reads each variant's text and its sent and click counts, and, when your export has them, the send date, a goal count such as purchases, and the image name. Never user IDs, phone numbers or events.

  3. 03

    One named processor

    In back-tests and predictions, each variant's text and a one-line channel and audience description go to TypeSafe's hosted API (the Jev model) for scoring and guardrails, even when AheadWise runs in your own cloud. User data never does. The audit sends nothing to anyone. An optional second opinion from Claude, blended in while AheadWise is still learning your audience, or a comparison with Claude, only with your written agreement, also sends copy to Anthropic.

  4. 04

    In your region, on your terms

    Templates and counts hold no personal data, so personal-data laws such as GDPR, the Saudi and UAE PDPLs, Singapore's PDPA and India's DPDP Act should not reach what AheadWise reads; your counsel confirms. Banks also answer to their own outsourcing rules: AheadWise can run in your own cloud account, in your region, and we provide what your review needs. Scoring still calls TypeSafe's API; Claude stays off unless you ask for it.

  5. 05

    Delete means delete

    One request removes your tests and everything learned from them.

What AheadWise reads
Test or campaign nameRead
Variant nameRead
Message textRead
Sent and clicked countsRead
Send date, goal count, image nameIf present
User IDs, phone numbers, emailsNever
Events and user journeysNever
<!-- audit.html -->
Content-Security-Policy: default-src 'self'; connect-src 'none'; form-action 'none'; worker-src 'none'

Pricing

Agree the pass mark first. Pay only if it's met.

Before anything is predicted, we agree in writing the score that counts as a pass, and the pre-flight tells you whether your history can reach it. Below the pass mark, you don't pay.

Audit

Free

In your browser, as often as you like

  • Every test's winner re-checked
  • What your tests could and couldn't detect
  • A report to share with your team
Run the audit

Blind back-test Start here

Free

On your past tests, before you commit

  • Predictions sealed before results are read
  • Scored against your real outcomes
  • Accuracy on all decided tests and on clear winners
  • Scored next to the picks your team, or a chat model, would have made

Predictions

After the back-test

Priced on what the back-test shows

  • Ranks variants before every send
  • Confident, leaning or test-it-live on every call
  • Learns from every test you run
  • Runs in your own cloud if needed (scoring calls TypeSafe)

FAQ

What teams ask first.

Why pay, when ChatGPT, Gemini or Claude will pick one for free?
A chat model's pick is often good: on public headline tests the best Gemini picked about 80% of clear winners, and AheadWise 78%. What it can't give you is what the money depends on. It can't tell you which of your past winners were real. It doesn't know how your audience has behaved unless you feed it, and even then it can't say when it's guessing: asked twice, the best Claude models changed 8–18% of their calls, while AheadWise changes about 6%, only toss-ups. And nothing checks it. So don't take our word either: bring the picks your team would have made, from Gemini, ChatGPT or gut feel, and the blind back-test scores them next to ours, on your own past tests, the same way.
Does this replace A/B testing?
No. It tells you which past results were real and how big the next test needs to be. For a one-off campaign, a small test slice followed by sending the leader usually earns more than a test big enough to prove anything, and AheadWise sizes that slice. Where a sealed back-test earns it, it also ranks copy for sends you can't test at all.
Can you check the lift our platform's AI claims?
Yes, when your platform kept a holdout chosen at random before the AI acted (Braze, MoEngage, CleverTap, Klaviyo and Insider can). Export the holdout's counts per day and we say, in orders and money, whether the lift is confirmed, smaller than claimed, or, more often than vendors admit, that the holdout was too small to tell. A matched group, an uploaded list or a before-and-after comparison can't confirm a lift whatever the numbers say, so we don't grade those. No holdout yet? We size one and draw it at random, and say what holding out costs. Check a lift
How many past tests do I need?
The audit works with any number. How many a back-test needs depends on how large your tests were and how much your variants really differ, so before anything is predicted, a pre-flight reads your counts and says what the back-test can show and whether it can conclude at all. With large tests judged on clicks, scoring around 100 is often enough; small or conversion-judged tests can take several hundred. AheadWise learns your audience's own taste from 100 decided tests; until then it predicts with general models and says so.
What counts as a clear winner?
A gap that would turn up less than 1 time in 20 if both versions performed the same (the standard two-proportion test, |z| ≥ 1.96). That is not a 95% chance the winner is better. The audit calls a winner real only at |z| ≥ 2.5, fewer than 1 time in 80, because tests stopped early often land just above 1.96, and only when, given how much your versions usually differ, the winner is the better copy at least 19 times in 20.
Which languages does it handle?
English is measured. Arabic, Bahasa Indonesia and Malay, Thai, Vietnamese, Tagalog, Hindi and mixes such as Arabizi and Taglish are read, but we haven't published accuracy for them: no public A/B results exist in those languages. Your back-test measures it on your own copy, and the pre-flight says beforehand whether your history is large enough to tell.
Does it know about Ramadan, Diwali and 11.11?
Yes, for 13 markets so far: the Gulf, Southeast Asia and India. That means Ramadan and Eid, Diwali, Chinese New Year, national days, the big sales (White Friday, the Dubai Shopping Festival, Prime Day, 9.9 to 12.12, Harbolnas, the Big Billion Days) and published paydays. The audit flags tests that ran in those weeks, because a festive winner may not win in an ordinary week, and the back-test is scored inside and outside them. Tell us your market and we add its calendar. It never changes a verdict.
Where is my data processed?
The audit runs only in your browser. Back-tests run on infrastructure we agree with you, including your own cloud account in your region, and variant text is scored by TypeSafe's Jev model, outside the region. Your data is used only for your back-test and deleted on request.

Find out which of your winners were real.

Two minutes, one export, nothing uploaded.

Run the free audit