Why AI written copy reads as generic, and the pass that fixes it

The actual reason it sounds like that

A language model produces the most probable next thing. That is not a criticism, it is the mechanism, and it is why the output is competent. It is also exactly why it is generic: the most probable sentence about a plumbing business is the sentence that appears in the most plumbing websites.

So the failure is not a writing failure. It is an input failure. Asked to write about your business with nothing specific about your business, the model writes the average of every business like yours, and the average of everyone is by definition nobody.

• It reaches for the category, not the case. "We pride ourselves on quality workmanship" is true of every trade that has ever existed, which is what makes it worthless.
• It smooths out anything odd. The specific detail that makes a business memorable is statistically unusual, so it is exactly what gets averaged away.
• It has no opinions. Real voice involves saying one thing rather than another. A model with nothing to go on hedges, because hedging is safer and more probable.
• It has no memory of your customers. It does not know what people actually ask you, so it answers the questions the internet asks instead.

What voice is actually made of

Voice gets talked about as tone, which is why fixing it usually fails. Tone is the last five percent. The other ninety five is information, and information can be collected.

The voice pass, step by step

This takes about ninety minutes the first time and produces something you reuse forever. It is deliberately low technology: a recorder and a document. There is no tool to install and that is the point, because the bottleneck was never the software.

1. Record yourself answering ten real customer questions. Not written, spoken, out loud, as if the customer were in front of you. Use the ten you actually get asked most. Speaking bypasses the part of your brain that writes marketing copy, which is the part causing the problem.
2. Transcribe it and do not tidy it. The value is in the bits you would have edited out: the aside, the caveat, the story about the one job that went wrong. That is the material nobody else has.
3. Pull out the sentences only you would say. Highlight anything that names a specific moment, a real constraint, a genuine opinion. You are looking for twenty to forty sentences. This is your voice sample, and it is worth more than any style guide.
4. Write down what you refuse to do. Jobs you turn down, customers you are wrong for, work you do not take. Nothing establishes a position faster than a clear no, and no competitor's copy will contain it.
5. List the five objections you answer every week. Word them the way customers word them, not the way you wish they would. These become the questions your copy answers, and they are the reason anyone reads past the first line.
6. Name one reader. An actual customer, with their actual situation. Writing to "homeowners in Montreal" produces copy for nobody. Writing to the person who called last Tuesday about the leak produces copy that works on everyone like them.
7. Assemble the brief. One document: the voice sample, the refusals, the objections, the reader. This is what you hand to a model, a copywriter or yourself in six months.

Where the model belongs in the process

After all of that, not before. Used at the end, on real material, it is genuinely good at the part that takes you longest.

• Good use: structure. Turning ninety minutes of rambling transcript into an ordered draft. This is real work and it saves real hours.
• Good use: volume with a spine. Two hundred product descriptions, each built from actual attributes and your own phrasing, is a job a person should not be doing by hand.
• Good use: the second draft. Tightening something you wrote badly is much easier than generating something you never said.
• Bad use: the first draft from nothing. This is where every generic page comes from, without exception.
• Bad use: asking it to "make it sound more human". It will add contractions and a rhetorical question. That is a costume, not a voice, and readers see through it faster than through plain competent prose.

How to tell if it worked

Four tests, in increasing order of how much they hurt.

1. The swap test. Put a competitor's name in place of yours. If the page still reads true, you have written about your category rather than your business. This kills most copy on the first try.
2. The read-aloud test. Say it out loud. Anywhere you would not talk like that to a customer standing in front of you, change it. This catches almost everything the first test misses.
3. The specificity count. Count the sentences containing something checkable: a place, a constraint, a number of days, a named situation. Under one in five, and the page will feel like nothing to read.
4. The disagreement test. Find the sentence someone could argue with. If there is not one anywhere on the page, you have not said anything, and nobody remembers a page that says nothing.

Where this fits with everything else

A voice pass is worth doing before any content project, and it is worth doing whether or not a model is anywhere near the work. Most of its value is that it forces a business to write down what it actually thinks, which almost nobody has done.

If the copy is fine and the problem is that visitors leave before reading any of it, that is a different failure. The four second test covers what a buyer judges before they scroll, and the website test scores a page against it.

If you want the plain explanation of what these systems are actually good and bad at, without the sales pitch, that is what AI automation actually is.