Writing with AI without producing slop
Google does not penalise AI writing; it penalises pages nobody needed. The reason most of it reads badly is simpler than the tooling, and so is the fix.
A model will write you a thousand words on anything in seconds. That is exactly the problem: the cost of producing text collapsed and the cost of producing text worth reading did not move at all.
What Google actually says
This is worth getting from the source rather than from the panic. Google's published position is that AI-generated content is not itself a violation. What its spam policies target is scaled content abuse: producing many pages whose primary purpose is to manipulate rankings rather than to help anyone. Google states the test is the purpose, not the tooling, and applies it whether pages were made by automation, by people, or by both.
Three further things Google says in the same guidance are worth knowing. Using AI gives content no ranking advantage of its own. Output may contain inaccuracies, so creators are expected to check it before publishing. And fabricating creator profiles, with invented names, false credentials or generated headshots, is treated as deception.
Why most AI writing reads badly
Not because a machine wrote it. Because of what it does when it has nothing specific to say, which is to say the general version of everything.
- It restates the question before answering it, so the first paragraph of every section is a paraphrase of the heading.
- It balances every claim into nothing. On the one hand, on the other hand, ultimately it depends on your business.
- It prefers the abstract noun to the concrete example. Leverage your brand equity, rather than put the price on the page.
- It lists three of everything, whether there are three or not.
- It ends sections with a sentence that summarises the section you have just read.
- It never says a number, because a number can be wrong and a generality cannot.
Every one of those is a symptom of the same thing: no specific knowledge was put in, so no specific knowledge came out. The fix is not a better prompt. It is having something to say.
A way of working that produces something worth reading
- Decide the one thing the piece is for, as a sentence. If you cannot write that sentence, there is no article yet, and starting the draft will hide that from you.
- Supply the specifics yourself before asking for a word of prose: the numbers, the example, the thing that went wrong, the sentence a customer actually used. This is the whole job.
- Ask for a draft of the structure before any prose, and argue with it. Structure is cheap to change now and expensive later.
- Draft section by section, not all at once. A whole article arrives smooth and unarguable, and you will accept it because it is finished.
- Cut the first paragraph of every section. It is almost always the restatement.
- Read it out loud. The balanced non-claims are obvious in the mouth in a way they are not on screen.
- Check every specific claim against a source, and link the ones a reader would want to check for themselves.
The test that matters
Give the finished piece to somebody who does the thing for a living and watch for the moment they nod. If they read the whole thing without once recognising their own experience in it, you have written something that is true in general and useless in particular. That happens to AI drafts more than to human ones, and it is the only quality problem here that really matters.
The opposite test is just as useful. If a reader could have got the same value from the heading alone, the words under it were decoration.
Work to do
- Take something you published and delete the first sentence of each section. Decide whether anything was lost.
- Find a paragraph that could appear on a competitor's site unchanged, and either make it specific or cut it.
- Write down the three facts only you know about your subject, and build the next piece around those rather than around the topic.
Sources