The real workflow behind AI-drafted landing page copy: what Claude gets right, where it needs heavy edits, and the $0.19 cost per client project.
A single client landing page draft costs me about $0.19 in Claude API calls and 40 minutes of editing, down from the 3 hours I used to spend staring at a blank headline field. AI writes the first draft. I do the part that decides whether the page converts: making it sound like the client instead of every other SaaS page on the internet.
That is the whole trick. Using AI to draft client landing page copy works when you treat the model as a fast intern who has read a million pages and remembers none of them well. It fails the moment you paste its output straight into production.
The workflow is brief in, variants out, human edits for voice, then a structured A/B test. I run four stages and never skip the third.
The model is good at generating twelve doors. It is bad at knowing which one the client's customer will actually walk through. That judgment is the job, and it does not automate.
Here is the prompt skeleton I reuse. It lives in a text file, not my head.
You are drafting landing page copy for [CLIENT], who sells [OUTCOME]
to [AUDIENCE]. Their voice is [3 adjectives + one they are NOT].
Rules: headline under 10 words. No em dashes. No antithesis phrasing.
No "revolutionize", "seamless", "unlock".
Return 6 headline + subheadline pairs, labeled by angle
(literal / benefit / edge). One sentence each on who each is for.
AI reliably nails structure, volume, and the boring-but-correct 80% of a page. Ask for 8 variants and you get 8 usable starting points in about 9 seconds.
It is genuinely strong at:
That coverage is why I keep using it. The draft stopped being the bottleneck. On a recent build for a Pylonworks client, the model produced the entire below-the-fold section in one pass and I changed maybe six words.
AI copy needs the heaviest revision on brand voice, specific claims, and anything with real stakes. That accounts for roughly 90% of my editing time.
The failure modes are consistent enough that I keep a checklist:
| What the model produces | Why it fails | What I do |
|---|---|---|
| "Revolutionize your workflow" | Generic, unprovable, every competitor says it | Replace with the specific outcome and a number |
| Antithesis phrasing (the "X, then Y reversal" pattern) | Formulaic, reads as AI-written | Rewrite as a plain declarative |
| Invented stats ("trusted by 10,000 teams") | Fabricated, legal and trust risk | Delete or replace with a real, sourced number |
| Uniform enthusiasm across every line | No human writes at one energy level | Vary the register, let some lines be flat |
| Five adjectives where one works | Padding | Cut to the strongest one |
The fabricated-numbers problem is the dangerous one. Models will happily assert social proof that does not exist, and per the FTC's rules on endorsements and fake reviews, invented testimonials or usage stats carry penalties that can reach tens of thousands of dollars per violation. I treat every number the model outputs as false until I trace it to the client. If the client has 340 users, the page says 340, never "thousands."
Voice is the other big lift. The model defaults to a confident, mid-Atlantic SaaS register. A plumbing-supply client and a fintech client come back sounding identical unless I intervene hard. That intervention does not compress into a prompt. I have tried.
A full landing page draft runs about $0.19 in tokens and 40 to 50 minutes of my time end to end. Here is the breakdown for a typical 5-section page.
The token cost is a rounding error. The value is the reclaimed 2 hours and 20 minutes per page, multiplied across every client project. That is the actual line item. The model is cheap. My attention is not, and this workflow spends less of it (see How Pylonworks Prices Web Development Projects in 2026).
Skip AI when the page hinges on a single specific claim, a legal statement, or a founder's personal story. Those need a human from the first word.
Regulated claims, pricing guarantees, and anything a lawyer will read do not get an AI first draft. Neither does founder narrative. The model cannot invent the texture of a real decision the founder made at 2am, and a landing page that leads with a fabricated origin story reads exactly as hollow as it is. For those I write cold and use the model only to pressure-test, asking it where a skeptical reader would bounce.
No, not by itself. Search engines rank for relevance and usefulness, not authorship. The risk is that unedited AI copy is generic, and generic copy ranks poorly because it says nothing specific. Edit for specificity and it competes fine.
For landing pages I use Claude Sonnet 4.5 because it follows negative constraints ("no em dashes, no antithesis phrasing") more reliably than most alternatives I have tested. The best model is the one whose default voice sits closest to plain, because you are editing away from its defaults either way.
Generate 6 to 8, ship 2. More than two live variants at once splits your traffic too thin to reach significance. With median landing page conversion near 4.6%, you need a few hundred sessions per variant before the winner is real and not noise.
Start with the prompt skeleton above, run one real client brief through it this week, and time yourself editing the output. If the edit takes longer than writing from scratch, your brief was too thin. Fix the brief, not the model.
Tired of re-keying the same data between tools? Pylonworks builds custom automation and internal tools for businesses without a developer, on a fixed quote you approve up front. Tell us what's eating your time