Table of Contents
TL;DR
- Generic AI copy comes from generic AI briefs — the prompt is the brief, and a thin prompt returns the statistical middle of the training data.
- A usable brief carries six things: the exact reader, their awareness level, the emotion to move, the mechanism, the proof order, and a named structural frame.
- The minimum substrate for copy that belongs to one client is three competitor pages pasted in full, two verbatim customer quotes, and one sentence of real differentiation.
- A refusal list — "high-quality," "innovative," "next-level," "seamless experience," passive voice — prevents the model from reaching for its defaults.
- Twenty to forty minutes of brief-writing replaces hours of editing generic drafts. Pick one poorly-converting page this week and rewrite its brief before touching the copy.
Generic AI copy is not an AI model problem. It is a brief problem.
Of course the draft reads generic.
The AI model has been asked to produce the average of every SaaS landing page it has ever seen. That is what the prompt describes. That is what comes back.
The surprising part is not that the copy is generic. The surprising part is how often the user blames the AI model.
Think about the difference between asking someone for "a meal" and asking them for grandma’s Sunday stew. The one with bay leaves, paprika, and pork that sat in buttermilk since Friday night. One request gets you whatever is easiest. The other gets you something that could only come out of one specific kitchen.
The brief is the prompt. The prompt is the brief. And the specificity of what goes in is the entire story of what comes out.
Why does my AI draft keep coming back generic?
Generic AI copy comes from generic AI briefs. Telling an AI model to "write landing copy for a SaaS" returns the average of a thousand landing pages.
The fix is substrate. Three competitor pages, two verbatim customer quotes, one sentence of real differentiation. That brief produces copy that can only belong to this client.
The brief is the prompt, not an input to one.
This is the shift most writers miss. In the agency era, the brief went to a human who filled in the context gap with judgment and experience. A half-finished brief still produced usable copy because the writer carried the missing half in their head.
The AI model does not carry any of that. It has read the entire internet and remembers none of your client specifically. Whatever you leave out of the brief, the AI model fills with the statistical middle of its training data. The middle of the training data is, by construction, average prose.
Rob Palmer, a direct-response copywriter who has tested AI prompts daily across $523M in tracked campaigns, puts the scale of the problem plainly. "Ninety percent of the AI copywriting prompts you find online are useless. They produce copy that reads like it was written by a chatbot — because it was."
The fix is not a better AI model. The fix is a brief that carries what the AI model cannot invent.
What does a brief that actually changes the output look like?
It carries customer language the AI model could not invent. It carries competitor pages the AI model can contrast against. It carries a refusal list that forbids the phrases the AI model defaults to.
A brief that reads like a project kickoff doc produces copy that reads like a kickoff doc.
A brief that reads like a persuasion argument produces copy that can sell.
Here is the test I use. Read your brief out loud. If it describes the product, you wrote a kickoff doc. If it describes what has to change in the reader’s mind between paragraph one and the call to action, you wrote a brief.
The difference is not length. A strong brief can be a page. A weak brief can be four pages. The difference is whether the brief commits to a specific reader, a specific emotional movement, and a specific structural frame.
Without those commitments, the AI model reaches for general-purpose copywriting muscle memory. It has a lot of that. Almost none of it is good.
What has to be in the brief for the copy to belong to this client and no one else?
Palmer names six components that separate usable output from chatbot prose.
- The exact reader — not "small business owners" but "B2B SaaS founders at $1M to $5M ARR who have failed with content marketing agencies."
- The awareness level of that reader, on the spectrum from unaware to most-aware.
- The emotion the copy has to move: fear, frustration, aspiration, envy, shame, or hope.
- The mechanism — the specific reason the offer works, not a generic claim that it does.
- The proof order, so the AI model knows whether to lead with a case study or a credential.
- The structural frame: PAS, AIDA, Star-Story-Solution, something named.
Miss any one of those, and the AI model fills the gap with statistical average prose.
This is why "write a landing page for my tool" returns something that could be about any tool. The brief named a tool. It named nothing else.
The six-item checklist looks like overhead until you read a draft produced against it. The first reading usually ends with "wait, this sounds like us." That sound is the brief doing its job. The copy is not hallucinating voice. It is reflecting the voice the brief carried in.
What is the three-plus-two-plus-one rule?
Three competitor pages pasted in full. Two verbatim customer quotes, pulled from a real conversation or review. One sentence of what this offer does that none of the three competitors do.
That is the minimum substrate for a prompt that will not produce generic copy. A fuller variable-by-variable template lives in the step-by-step guide for writing a brief for ChatGPT, with each prompt variable as its own slot.
Below it, the AI model averages. At or above it, the AI model has enough constraint to produce copy that can only be about this client.
Why these specific numbers? Three competitors gives the AI model a contrast field. One competitor looks like a target — three competitors look like a category.
Two customer quotes give the AI model idiom — actual words real buyers use, not paraphrase. One sentence of differentiation forces you, the brief-writer, to commit to a position. A brief that cannot name what is different in one sentence has not actually decided what the offer is.
The Everworker team frames a similar discipline as the CARE framework — Context, Ask, Rules, Examples. Their required variables overlap with Palmer’s.
- Persona with objection themes.
- A primary KPI with funnel stage.
- Banned-phrase lists.
- Two on-brand examples plus one "don’t" sample.
- Source material with an inline-citation requirement.
Two practitioner traditions. One conclusion. Constraint density is the thing.
What goes wrong when the brief is just a job title and a word count?
The AI model fills the vacuum with the middle of its training data.
You get "high-quality." You get "innovative." You get "next-level." You get "seamless experience." You get passive sentences that describe nothing. You get three paragraphs that could be swapped into a competitor’s page without a single word changing.
A Toolient case study tracked this effect across a full product catalog drafted by AI with thin briefs. "Every page sounded identical," the team observed.
Click-through rate held steady. Add-to-cart rate did not. The traffic was arriving, but the copy was failing to convert once the visitor landed.
The fix is a refusal list. Ban the cliché set in the brief itself. Force the AI model into trades like these. These same defaults are what makes press-release voice in AI copy the model’s factory setting — the brief’s refusal list and the prompt’s banned-phrase block do the same work from two sides.
- "Lightweight design" becomes "carry it all day without shoulder fatigue."
- "High-quality materials" becomes "lasts three years under daily use."
- "Enhance productivity" becomes "finish your daily workflow in under two hours."
Same product. Same AI model. Different brief.
The second brief refused the default phrases and demanded specific outcomes. The copy had no choice but to become specific.
The brief left no fingerprint, so the copy has none. Leave a fingerprint.
How long should this kind of brief take to write?
Twenty minutes to forty minutes for a landing page. Longer for a full sales page.
That feels like overhead until you count the hours saved on rewriting generic drafts.
Writers I talk to who resist brief-writing are usually optimizing for the wrong metric. They are counting minutes to first draft. The number that matters is minutes to usable copy. A thirty-minute brief and a ten-minute draft beats a five-minute brief and four hours of editing, every time.
Palmer is also explicit that even a six-variable brief does not produce final copy in one pass. "AI output requires three to five iterations and professional editorial judgment to become truly high-converting copy." The brief closes most of the gap. The last yards are human editing. That is still a massive improvement over editing starts-from-scratch generic prose.
A thorough brief is not extra work. It is the work moved upstream, where one decision replaces ten rounds of editing.
Other questions worth answering
How do you tell a real voice-of-customer phrase from paraphrase?
Two tests. First, the phrase has to use words the buyer chose, not words the writer would translate them into. Second, the phrase has to name a specific pain or outcome, not a category.
‘It crashes every Friday afternoon’ passes both. ‘Reliability issues’ fails both. Per Everworker’s March 2026 framework, persona objection themes need this fidelity to count as Context inside CARE.
Does using Claude versus ChatGPT versus Gemini meaningfully shift how thick the substrate must be?
Less than people expect. The choke point sits in the prompt, not the engine. Per Rob Palmer’s February 2026 article, the six-variable brief carries the constraint density the engine cannot supply.
Tone polish drifts a little between Claude and ChatGPT. Substrate density decides whether the first pass is usable.
Why do banned-phrase lists do more work than positive style guidelines?
Because the engine defaults faster to a cliché it knows than to a positive example it has to interpret. A refusal list closes the easy paths. Positive style notes open new ones, but only if the engine can match the abstraction. Per Toolient’s March 2026 guidance, the swap from ‘high-quality’ to ‘lasts three years under daily use’ came from banning the adjective.
Can you reuse the same prompt skeleton across multiple landing pages?
Yes, with one swap each page. The structural frame, refusal list, and voice constraints stay constant. The customer quotes, competitor URLs, and one-sentence differentiation change every time.
Per Everworker’s March 2026 framework, the Rules block is durable across pages, and the Context block is page-specific. Reusing both blocks together is how Toolient’s March 2026 catalog ended up sounding identical.
What should you put in your next AI brief?
Pick one page that converts poorly. Just one.
Write a brief with three things. Three competitor URLs. Two verbatim customer quotes pulled from a real conversation or review. One sentence of what this offer does that the three competitors do not.
Feed it to your AI model. Compare the first draft to whatever the page says now.
If the new draft still reads generic, the brief is still too thin. Add constraint until the draft reads like it could only be about this client.
A banned-phrase list. A named structural frame. A specific emotional movement you want the copy to produce.
This week is not about writing better copy. It is about writing a better brief. The copy follows automatically.
The stew is not about the pan. The stew is about what you chopped before you turned the heat on.
If you rewrote a brief and are unsure whether the draft is specific enough to belong to this client, you can contact me here. Paste the brief and the draft.
I will tell you which parts the AI model actually used and which it ignored. I will show where the copy still reaches for the statistical middle. No pitch.