I needed 10 headline angles for one landing page. Triage, not generation, was the hard part.

I had a landing page that converted at a number I didn't trust. The copy was fine, but the headline had been written once, by me, on a tired afternoon. The fix everyone recommends is A/B testing, and the first step of that is having variants. So I built a tiny workflow that produces ten headlines across five angles — benefit, curiosity, list, how-to, story — in a single run.

The generation took minutes. The surprising part was that generating wasn't the work at all.

Why ten, and why by angle

A single model call that says "give me 10 headlines" tends to return ten near-identical sentences with shuffled words. They test the same thing, so you learn one thing when you ship them. Bucketed by angle, you get coverage: a benefit line tests the value prop, a curiosity line tests the open loop, a how-to line tests intent match. You're not testing headlines, you're testing hypotheses about what the reader cares about.

When I used HeadlineGen, the point wasn't the raw count. It was that the five angles are explicit, so I could throw away the three I didn't want to test and keep the two that mapped to a real question. That's a smaller, sharper test.

The triage step nobody warns you about

Once you have ten lines, you face a second problem: you can't A/B test ten headlines on a page that gets 40 visits a day. You'll never reach significance. So the actual job is elimination.

My rule became: drop anything that isn't a complete, legible sentence on its own; drop anything that needs the subhead to make sense; drop anything that only differs from another by an adjective. That usually leaves three. Three is testable on a small traffic page over a couple of weeks.

This is the part I underestimated. The tool removes the writing friction, which is real, but it shifts the bottleneck to judgement. If you're not willing to cut, you'll sit on ten headlines and test none.

What the model is bad at here

Headlines are a place where the model's eagerness hurts. It wants to please, so it over-promises in the copy: "the only tool you'll ever need," "guaranteed to convert." Those are not just tacky, they're claims you can't back and shouldn't ship. I started post-filtering for superlatives and removing them by hand, because the generator has no incentive to be restrained.

The other failure is sameness. Without the angle constraint, the fifth "variant" is the first one with synonyms. The structure matters more than the model does.

What I'd do differently next time

I'd start the test smaller and more honest. One benefit headline against one curiosity headline tells you whether your visitors want a payoff or a question. That's a real fork, and you can read it in a week on modest traffic. Then iterate within the winning angle.

If you want to see the angle-bucketed output shape, HeadlineGen is where I ran these. It gives ten per run across the five angles and copy-ready text, but treat the output as candidates, not answers. The cutting is the job.

FAQ

Do I need ten headlines?
No. Ten is a starting pool. Most pages should test two or three at a time. The value of generating many is having enough to cut down to the few that matter.

Can a generator replace a copywriter?
For a first draft under time pressure, it's useful. It won't know your audience or your claims, so you still own the cuts and the honesty of what ships.

Why bucket by angle instead of just asking for variety?
Bucketing forces coverage of different hypotheses. Unstructured "variety" collapses into synonyms that test the same thing.

Will better headlines fix a bad product page?
No. A headline test reveals message fit, not product fit. If the page converts poorly after the headline is right, the problem is downstream.