Why does AI-written content all sound the same, and how do you fix it?

The averaging problem

A model asked to write about furnace maintenance draws on everything it has seen about furnace maintenance, which is thousands of pages written by people who mostly had nothing new to say. The output is a competent restatement of that consensus. It is not wrong; it is common.

This is why prompt engineering has limits here. You can change tone, structure and length with instructions. You cannot instruct a model into knowing something that is not in its inputs, and the thing that differentiates your content is precisely what is not in its inputs.

What counts as a real input

• Patterns from your own calls. The four objections that come up constantly, the phrasing customers use, the question that always follows the quote. Call analysis turns this into structured material instead of anecdote.
• Technician observations. What actually fails first on the equipment common in your market, what a symptom usually turns out to mean, what customers consistently misdiagnose.
• Your own operational data. Which job types cluster, what seasonality looks like in your service area, how long things really take.
• Local conditions. Water chemistry, climate load, housing stock, permitting reality. Specific, verifiable, and absent from generic training data.

A structural fix: write the claims first

Instead of asking for an article and editing the result, have the subject matter expert produce a bare list of assertions — five to ten sentences, no prose, just what is true and worth saying. Then have the model expand that outline into readable structure without adding claims.

This inverts the usual workflow and it works because it puts the human where the scarce value is. Fifteen minutes of expert time produces the substance; the system produces everything else. It also makes review trivial, because the reviewer is checking whether the draft still says what the outline said.

Tells you can search for

Keep a list of phrases that signal averaged output and grep for them before publishing: "in today's fast-paced," "unlock," "elevate," "it's important to note," "whether you're a homeowner or a business owner," "delve." None of them are crimes individually and all of them are evidence that no specific claim is being made in that sentence.

A stronger version of the same check: count the sentences on the page that could only have been written by someone in your business. If that count is zero, the page is a restatement of the consensus regardless of how well it reads. Making that count non-zero is the whole job of a content system worth running.