How to Write AI Copy That Doesn't Sound Like AI

What actually gives the copy away

A language model writes toward the middle. It is predicting a likely next word, and the likeliest word chosen several hundred times in a row produces prose that sits near the average of everything ever published about your subject. Average prose is clear. It is also unplaceable, and unplaceable is the real complaint when somebody says a page sounds like AI.

Writing by a person is lumpier. A nine word sentence lands after a thirty word one. A paragraph opens with “and” because the thought carried over from the last one. A detail shows up that serves no argument and stays in anyway, because it happens to be true. Models iron that flat, and the flatness registers before any single word does.

This is a copy problem, separate from the visual defaults. If the page itself looks generated, that is a different list of fixes and I wrote it up as the six tells and what to do about each one. What follows is only about the words.

The sentence habits to break

Four patterns account for most of it. None of them are errors. They are all defensible choices that a person makes occasionally and a model makes constantly.

Uniform length. Draft copy tends to sit between fifteen and twenty five words a sentence, line after line, and the effect is a metronome. Count instead of reading. If three sentences in a row land within five words of each other, cut one in half or let one run long. A four word sentence after a long one does more for rhythm than any word swap.

Sets of three. Models group things in threes because three sounds finished. Count your actual reasons. If you have two, write two and stop. If you have five, write five and let the list look uneven. A list of four is quietly reassuring for exactly that reason: nobody pads to four.

The hedge in front of the claim. Sentences shaped like “While speed matters, quality matters more” turn up everywhere in generated copy, because balancing is safer than asserting. Pick a side. State what you believe, then give the objection its own sentence instead of smuggling it in as a subordinate clause.

The reversal. “It is not just a website. It is your storefront.” Readers have seen that construction enough times that it now reads as filler. Delete the first half. Often the second half stands up on its own, and sometimes neither half was saying anything.

The restatement loop

Generated paragraphs tend to say one thing three times: a claim, support for the claim, then the claim again in different words. It reads as thorough. It is padding, and it is the habit I catch most often in my own AI drafts.

Try this on any page you have. Delete the final sentence of every paragraph, then read it back. If nothing was lost in nine paragraphs out of ten, those sentences were restatement. Keep the one or two that genuinely added something and leave the rest deleted.

At page level the same thing shows up as a closing section that summarizes what the reader just read. Nobody who reached the bottom of a landing page needs a recap of it. Cut the summary and end on the thing you want them to do.

The substitution that does most of the work

One edit outperforms everything above: replace every category noun with the specific thing it is standing in for.

• “businesses” becomes “two dentists and a dog groomer”
• “solutions” becomes “a booking form that emails you”
• “a wide range of services” becomes the services you sell, named
• “quickly” becomes “same day, usually before noon”
• “affordable” becomes the price

A model cannot average what it has never seen. It does not know your booking flow, your turnaround, or the thing a customer said on the phone last Tuesday. Every specific you add is a sentence no competitor's page can contain, which is also why this edit does more for conversion than it does for tone.

Numbers are the cheapest version of it. Price, turnaround, how many years, how far you travel, how many of these you have shipped. Generated copy reaches for adjectives in those slots because it has no figures to put there. You do.

The pass, in order

Order matters, because the later passes undo the earlier ones if you run them backwards. Rewriting a sentence to make it specific usually deletes its slop word for free, so there is no point hunting words first.

1. Specificity. Highlight every noun that names a category rather than a thing. Replace it or cut it. This is the slow pass and the one that matters.

2. Shape. Read sentence lengths rather than sentences. Break up the runs. Aim for at least one sentence under eight words in every paragraph.

3. Threes. Count every list on the page and adjust it to the real number of items.

4. Restatement. Delete the last sentence of each paragraph. Restore only the ones you miss.

5. Vocabulary. Now run a find for the usual suspects: “delve”, “tapestry”, “realm”, “seamless”, “elevate”, “robust”, “foster”, “holistic”, “transformative”. Add your own as you notice them, and keep the list where your next draft will see it.

6. Read it out loud. All of it, at normal speed, covered next.

On a homepage this runs about forty minutes. On a long sales page, closer to two hours. Either way it is faster than writing from nothing, and it produces a page that could not have come from anyone else.

Reading it out loud, and where copy breaks

Reading aloud catches what reading on screen hides. You run out of breath halfway through a sentence that needed to be two. You hear the metronome. You arrive at a line you cannot say with a straight face, and that is almost always a line the model wrote to sound confident.

There is a harsher version of the test. Take a sentence off your page, paste it into a message to a friend, and look at it sitting there. Stripped of the layout and the heading above it, generic copy looks absurd in a chat window. Specific copy reads normally. That gap is the whole test.

If you want the diagnosis before the rewrite, the scored version of this is a ten check audit you run on a site you already have, and three of those checks are copy.

What to put in the prompt so there is less to fix

Editing is downstream. Most of this can be prevented at the prompt, and the highest yield item by a distance is a voice sample.

Paste three hundred words you actually wrote. An email, a forum reply, a voice note you transcribed. Tell the model to match the rhythm and the register, not the topic. One real sample beats any list of adjectives describing your tone.

Give it a banned list. Yours, written down. Mine lives in the project and grows every time a draft annoys me.

Constrain the shape. Vary sentence length. At least one sentence under eight words per paragraph. No lists of three unless there are exactly three things. Mechanical instructions like these get followed far more reliably than “write conversationally”.

Hand over the facts. Price, turnaround, the two objections you hear most, what you will not do. Without those the model fills the gaps with category language, and then you are editing fiction rather than copy.

Name the reader. Not a persona document. One actual person who nearly bought and did not, and why.

Humanizer tools, detectors, and what Google says

Tools that rewrite text to defeat detection do the opposite of the pass above. They swap words and leave the structure alone, so the copy keeps the metronome and the empty claims, now with odd synonyms sitting in it. Readers react to shape. Detectors do not.

A detector scores how predictable the text is, which is a different question from whether the writing is any good or whether a person stands behind it. Writing to beat one makes copy worse. What visitors actually notice, and whether to say anything about how the site was built, is a separate post.

Google's own guidance on this points at helpfulness and at who produced the content, how, and why, rather than at which tool typed it. The page is Creating helpful, reliable, people-first content, and the self assessment questions in it make a better copy brief than most copy briefs.

What to leave alone

Overcorrection has a smell of its own: forced fragments, a one line paragraph every third paragraph for drama, a knowing aside at the end of every section. That reads as performance, and performance is as impersonal as the average it replaced.

Plain, clear sentences are not a tell. A page can be calm and specific and still sound like somebody who has done the work. What you want is copy that could only have been written by you about this, not copy advertising how human it is.

None of it substitutes for the copy doing its job either. Rhythm fixes tone. It does not fix a page that never says what the reader gets, which is a separate craft and the subject of website copy that converts.

---

Pick one page, probably the homepage, since it gets the most readers and was usually drafted in the biggest hurry. Run the passes in order and read the result out loud. You will delete more than you add, the page will get shorter, and what survives will say things a competitor cannot copy without lying. Then put the banned list and the voice sample somewhere your next draft will find them, so tomorrow starts closer to done.

Related: How to Give Claude a Design System to Work From · How to Use Claude Projects to Run a Business · How to Make Claude Check Its Own Work on a Website

Common questions

How do I make AI writing sound more human?

Fix sentence shape and specificity before vocabulary. Break up runs of similar length sentences, count list items instead of defaulting to three, delete the restating sentence at the end of each paragraph, and replace category nouns like businesses or solutions with the actual thing. Then read the page out loud.

Do AI humanizer tools actually work?

They swap words and leave the structure alone, so the copy keeps the uniform rhythm and the generic claims that readers react to, sometimes with odd synonyms added. Editing for specificity and rhythm yourself takes longer and produces a page that reads like a person wrote it.

Will Google rank my page lower because AI wrote the copy?

Google's guidance focuses on whether content is helpful and on who produced it, how, and why, rather than on which tool typed it. Thin, unedited, near duplicate copy is the actual risk, and that risk is the same whether a person or a model produced it.

How long does an editing pass take on a page of copy?

About forty minutes for a homepage and closer to two hours for a long sales page, running the passes in order: specificity, sentence shape, list length, restatement, vocabulary, then reading the whole page out loud at normal speed.

Is it better to fix the prompt or edit the draft?

Both, in that order of effort. Three hundred words of your own writing as a sample, a banned word list, and your real numbers remove most of the problem up front. Every draft still needs a pass afterwards, so prompting reduces the editing rather than replacing it.