How to Use Anyword’s Predictive Score to Pick Ad Copy
Anyword assigns each generated copy variant a performance prediction on a 0-100 scale, based on a model trained on years of ad performance data. You write a brief, generate 20-30 variants, sort by score, and test the top three against your current control. The score is a filter, not a verdict — variants that score 15 points apart usually do not differ in real results.
Anyword has been selling that prediction since 2013, before generative text was fashionable. The copy generator is now table stakes; the score is the reason people pay.
How is the score calculated?
Anyword does not publish the model, but the inputs are visible in the product: audience persona, channel, content type, plus the text itself. The training set is drawn from real ad performance across millions of impressions.
What that means practically: the score reflects what tends to work for audiences similar to the one you selected, not what will work for your specific account. A score of 82 says “this resembles copy that historically converted well for this persona,” nothing more.
How do you set up a brief that produces useful variants?
Five fields decide the output quality: product description, audience persona, tone, keyword requirements, and the channel.
A brief that works for a paid social ad:
- Product: project-management tool for agencies, $29/user/month
- Audience: agency ops managers who already use spreadsheets
- Tone: direct, no hype
- Keywords: agency, profitability, timesheets
- Channel: Facebook feed, primary text
Without a persona, you get generic marketing prose scored against a generic audience. The persona dropdown is not decoration.
What does a score gap actually mean?
Anyword’s own guidance treats small differences as noise. In practice, variants within about 5 points of each other behave like a tie, and the honest move is to test both if you can afford it.
A working rule:
| Score range | Action |
|---|---|
| Top 3 within 5 points | Test all three, small budget |
| Top 1 leads by 10+ | Make it the primary challenger |
| Everything under 55 | Rewrite the brief, not the copy |
That last row saves the most time. When every variant scores low, the persona or the offer is the problem, and regenerating 20 more variants will produce 20 more low scores.
How do you run the test after scoring?
Score, shortlist, then test for real. The score narrows 30 variants to 3. The ad platform decides the winner.
A budget-conscious structure: your existing control gets 50% of spend, the three challengers split the rest. Run for 4-7 days or until each challenger has at least 1,000 impressions, whichever comes later. Any test that concludes on 200 impressions is measuring randomness.
Keep the winning challenger and retire the control. Then re-brief with the winning language as input — the second round improves more than the first because you are now feeding it evidence.
Where has the prediction been wrong?
New offers with no market history. Scores come from comparison to past copy, so a genuinely novel messaging angle gets scored against things it resembles, not against its own merit. A good idea can score 55.
Regulated industries are a consistent weak point: finance, health and insurance copy that must clear legal review often scores poorly because compliant phrasing resembles the cautious copy that historically underperformed.
And the score says nothing about brand safety, claims substantiation or whether the copy sounds like a human. High-scoring does not mean publishable.
Does the score replace an A/B test?
No, and treating it as one is how people lose money. Anyword’s own framing is prediction, not proof. What the score genuinely buys you is a cheap way to kill the bottom 80% of variants before they reach a test budget.
If you run 5 tests a month, the score saves maybe two of them from wasting spend. That is a modest but real win.
Who is Anyword for?
Performance marketers with a steady ad cadence who write 20+ variants a week and have a defined audience persona. Also teams that need brand-voice enforcement across several writers, since the style controls are part of the same product.
Skip it if you run one campaign a quarter — the per-month price (starts around $39 on the starter tier, higher for the predictive features) does not pay back. Skip it if you cannot test anything, because an untested score is just a number you like.
Quick Summary
- Anyword scores copy variants 0-100 against a persona-specific model trained on ad performance data.
- Treat gaps under 5 points as a tie; regenerate the brief if everything scores below 55.
- Score to filter, test in the ad platform to decide.
- Novel offers and regulated industries score unreliably.
- Worth the price only for teams running frequent, budgeted tests.
