Your landing page has no error message

A language model can only rearrange the specifics it was given. Ask it for a landing page for “an invoicing tool for freelancers” and it has two facts to work with, so it fills the rest with the safest sentences it knows: streamline your invoicing, get paid faster, built for modern freelancers. Each one is grammatical, confident and true of every invoicing tool ever made.

The trap is that the result looks finished. Polish is the signal most of us use to judge whether writing is done: no typos, good rhythm, a headline that sounds like a headline. A model produces that polish for free, so the copy passes the check you were actually running, which was “does this look professional?” rather than “does this tell a stranger something they did not know?”

You are also the worst-placed person to notice. You know what the product does, so when you read “Get paid faster” your head silently supplies the rest: the reminder emails, the currency handling, the thing that makes yours different. A visitor has none of that. They get the five words.

There is no single metric for “the copy is not working”, but there are signs worth taking seriously:

• Most visitors leave within seconds, and the ones who stay mostly came from people who already knew you: friends, followers, a community where you explained the product in your own words.
• People ask you in messages what the product does, when the page supposedly answers that.
• When someone describes your product back to you, they use words that appear nowhere on the page.
• Sign-ups arrive and then never use the thing, because they signed up for something else they thought it was.

None of these proves the copy is the problem. Together they are about as close to an error message as copy gets.

The fastest test is a comparison. Collect the words the people you are selling to actually use, and put your first screen next to them.

1. Gather raw language: support emails, direct messages, notes from calls, threads in communities where your users complain about the problem, reviews of the tools they use now, and the search queries in Google Search Console if you have any search traffic.
2. Highlight the nouns and verbs they use for the problem and for the result they want. Not your category name, their words.
3. Read your headline, the line under it and your main button. Does any of their language appear there? If your page says “streamline” and nobody you have ever talked to said “streamline”, you have found the mismatch.
4. Run a five-second test. Show the first screen to someone who fits your audience but has not seen the product, for five seconds, then hide it and ask what it does and who it is for. If their answer is vaguer than you hoped, or confidently wrong, that is the error message you were not getting.

It also helps to strip out the phrases that carry no information before you judge what is left. The free vibe coded website checker lists stock phrases, vague buttons and stock headline shapes on any public page, with no AI involved. It is a heuristic, not a verdict, but it is a quick way to see how much of your first screen could belong to anyone.

Most AI-written copy goes wrong because of the order it was written in. The model writes first, and you edit. By the time you edit, the frame is already the model's frame, which is the average, and you are polishing a generic page into a slightly less generic one. Reverse it.

First, talk to five people who have the problem. Five is not a statistical number. It is enough for the same phrases to start repeating, and few enough that you will actually do it this week. Ask about the last time they dealt with the problem, not about your product: what they were trying to do, what they tried, what annoyed them. Write their words down exactly, including the clumsy ones.

Second, write the ugly sentence yourself. One sentence that says what the product does, for whom, with what goes in and what comes out. It will be long and graceless. For an invented example, a bookkeeping tool: “You upload the CSV your bank exports and get back a categorised spreadsheet your accountant will accept, without retyping anything.” Nobody would put that on a billboard. It is still more useful than “Streamline your finances with AI”, because every part of it is specific and checkable.

Third, let the model polish. Now the model has something to work with. Give it the ugly sentence and the customer quotes, and constrain it: keep every concrete noun, add no adjectives, make no claim that is not in the input, and keep it under a set length. Ask for several versions and pick one. Models are good at compression and rhythm. They are bad at knowing what is true about your customers, and this order keeps them on the job they are good at.

Continuing the invented example, the polished line might come out as “Bank export in, accountant-ready spreadsheet out” with “No retyping, no bank login” underneath. It is shorter and better, and it still says what the ugly sentence said.

You cannot give copy a stack trace, but you can add a few cheap places where the failure becomes visible:

• Ask one question after sign-up: “What almost stopped you from signing up?” or “How would you describe this to a colleague?” The answers are free customer language for your next draft.
• Watch a handful of session recordings of the first screen, if your analytics tool offers them and your consent setup allows it. Where people stop, scroll back up or leave is as close to a bug report as copy gets.
• Read your numbers by traffic source rather than as one average. A first screen that works for people who searched for you can fail completely for people who tapped a link in a feed. There is a separate post on reading traffic by source.
• See how your headline would need to change for different arrivals. The free headline rewriter rewrites a page's headline, subhead and button for visitors from search, a social post, Product Hunt and Hacker News. Treat its output as a prompt for your own thinking, not as final copy.

None of this means you should stop using AI to write copy. It means the model is the last step, not the first. The specific material, the words your customers use and the one sentence only you can write about your product, has to come from you, and no tool can generate it from nothing.

For transparency: we build IntentGrid, which rewrites a page's copy for each visitor based on how they arrived. It has exactly the limit described above. It can only rearrange what is already on your page, so it does not replace the five conversations, and it measures nothing, so it will not tell you whether your copy works either. It is in early access.