A client handed me a landing page last March with 3,400 monthly visitors and a 1.8% signup rate. Nothing broken, nothing obviously wrong. The headline was fine. That was the problem: fine. We ran the page through a structured workflow in Anyword and shipped a new hero section nine days later. Signups hit 3.1%. Same offer, same traffic, different words.
What follows is the exact sequence I use now, minus the parts that didn’t work. It takes about an hour for a single page and scales reasonably well to ad campaigns and email sequences.
Step 1: Write the brief like you’re handing it to a freelancer
The biggest mistake people make with Anyword is treating it like a slot machine. Type a product name, pull the lever, hope for something clever. The output is only as sharp as the context you give it, and the context lives in two places: the brand voice settings and the brief box.
Twenty minutes here saves you from sifting through forty useless variants later.
What actually belongs in the brief
- Audience, specifically. Not “small business owners” but “freelance designers who invoice 4-8 clients a month and chase payments by email.”
- The single action. One page, one ask. If you want them to book a demo and download a guide, pick one.
- The objection they’re carrying. For my client it was “I already use a spreadsheet and it’s not that bad.” Naming that changed every headline the tool produced.
- Tone markers with examples. “Direct, slightly dry, no exclamation marks” beats “professional and friendly.”
- Three competitor headlines you want to sound nothing like.
Set the brand voice to match before you generate anything. Anyword will hold that tone across dozens of variants, which is the whole reason you don’t have to rewrite from scratch every time.
Step 2: Build a wide field before you narrow it
Generate 15 to 25 variants, not five. You want enough noise that the good ones stand out. Ask for a mix: benefit-led, curiosity-led, number-led, and one deliberately blunt version that just states the offer.
If the angle list feels thin, that’s not an Anyword problem, it’s an input problem. I often pull in a separate assistant to brainstorm objections, customer phrasings, and adjacent angles first, then paste the best five into the brief. If you want a shortlist of tools that handle that brainstorming step well, there’s a solid rundown of the chatbots worth using in 2025 that covers the practical differences.
One caution: don’t generate 25 variants of the same idea with different adjectives. Vary the structure. A headline that opens with a number behaves differently from one that opens with a question, and the scoring model picks that up.
Step 3: Let the predictive score do the first cut for you
This is the part that separates Anyword from a standard text generator. Each variant gets a performance score, and the tool tells you which ones are likely to outperform your baseline. I usually sort by score and delete anything in the bottom half without reading it too closely.
Scores aren’t prophecy. They’re a filter, and a filter that removes 60% of the noise is worth having. If you want the fuller picture of where those scores come from and how they’re trained, there’s a detailed review of how Anyword predicts conversion that goes deeper into the model than I will here.
Step 4: Read the breakdown, not just the number
A 78 and an 81 are functionally the same score. What matters is why a variant scored where it did, because that tells you what to keep when you start editing by hand.
Audience fit versus channel fit
A headline can score well for your audience and poorly for the channel. Long, explanatory copy plays fine in an email and dies in a display ad. Check both dimensions before you commit. On the client’s page, the top-scoring variant was a warm, narrative line that scored 89 for audience and 63 for the channel. It went into the email sequence instead, where it beat the control by 22%.
Step 5: Hand-edit the survivors
Never ship raw output. The model gets structure and hook right, but it doesn’t know your customer’s exact words, and it will occasionally invent a claim you can’t legally make. My editing pass is short and mechanical:
- Swap in one phrase lifted directly from a real support ticket or review.
- Cut every adjective that isn’t doing work.
- Read it out loud. If you stumble, rewrite.
- Check the claim against the product page. Every time.
The original headline on that page read “Manage Your Money Better.” Scored 61. The rewrite, run through the same tool, came out as “Pay off your smallest card first, then watch the rest move faster.” Scored 84, and it converted at roughly double the old one.
Step 6: Ship it and reset your baseline
Run a real test. One variable, enough traffic to matter, at least two weeks. When you have a winner, paste the winning copy back into Anyword as reference material for the next round. The tool gets better at your voice when you tell it what actually worked, and the scores start lining up more closely with your own results.
Do this three or four times and you’ll notice something useful: you stop arguing about copy in meetings. You have numbers.
A full worked example, start to finish
A five-email onboarding sequence for a project management tool. Here’s how the hour broke down.
Minutes 1-15: Brief written, brand voice set to “plainspoken, technical, no hype.” Objection logged: users had tried two other tools and abandoned both.
Minutes 15-30: Twenty subject lines generated across five angles. Sorted by score. Bottom twelve deleted.
Minutes 30-45: Top eight edited by hand. Two killed for unsupportable claims. One rewritten entirely because the score was high but the phrasing sounded nothing like the brand.
Minutes 45-60: Six subject lines loaded into the ESP for a two-week test across 4,200 subscribers.
Results: the winning line opened at 41.3%, against a 28% baseline. Second place, which had scored four points lower in Anyword, opened at 39.8%. The model got the ranking roughly right, which is the honest expectation. It’s a strong prior, not a crystal ball.
The habit worth building is the loop: brief, generate wide, filter hard, edit by hand, test for real, then feed the result back in. Do that four or five times and your copy stops being a matter of opinion.