Where AI Content Genuinely Helps (And Where It Backfires)

The argument about AI-written copy is usually held between two groups who are both wrong. One treats a language model as a printing press for free pages. The other believes any machine involvement triggers a penalty. Neither position survives contact with how search engines actually assess pages, or with what a model is good at.

We use AI in our own writing process every week. We also throw away a lot of what it produces. The useful question is not whether to use it — it is which parts of the job it can be trusted with.

What the rules actually say

Google’s spam policies name a specific offence: scaled content abuse, which its documentation describes as generating many pages primarily to manipulate search rankings, with little or no value added for users. The important clause follows: the policy applies no matter how those pages were created. Scraping, spinning, templating, outsourcing to a content farm and prompting a model all sit inside the same rule.

Google’s public guidance on this has been consistent — its stated focus is the quality of the content rather than the method of production. That cuts both ways. Nobody is going to demote a page because a draft started in a chat window. And nobody is going to spare a page because a human typed it, if it exists only to catch a keyword.

So the line is not human versus machine. It is whether the page was made to be read or made to be counted.

The four jobs it genuinely does well

Structure before substance. Getting from a blank page to a defensible outline is the slowest part of writing for most people. A model is quick at proposing an order of argument and quick at listing the questions a reader will have. You will discard some of them. That is still faster than staring at nothing.

First drafts of things you already know. If you can explain a topic out loud, a model can turn your explanation into paragraphs faster than you can type them. The knowledge is yours; the transcription is the machine’s. This is the single highest-value use and the one least likely to produce something false, because you are the source.

Variations on a line that already works. Fifteen versions of a meta description. Eight headline options for the same offer. Ad copy variants for testing, which is why this pairs neatly with Google Ads work. The original judgement is human; the permutations are mechanical.

Compressing research you already hold. Sales call notes, support tickets, a survey export, a pile of your own past articles. Summarising a document you supplied is the task where models are most reliable, because the answer is in front of them rather than in their weights.

The four places it falls over

Anything requiring first-hand experience. A model has never stood in a client’s warehouse and noticed the thing that changes the recommendation. It can produce a paragraph shaped like experience — “in our work with clients across the sector” — and that paragraph is a lie. Those sentences are the ones readers trust most and the ones you can least afford to fabricate.

Local specificity. Ask for a paragraph about trading conditions in a particular London borough and you will get something plausible, general and slightly wrong. High streets, catchment areas, council rules and which competitors actually operate nearby are all facts a model will confidently approximate. If your page depends on knowing a place, the model does not know it.

Factual claims nobody is going to check. A drafted paragraph will contain numbers, dates, regulations, product specifications and price ranges, delivered in the same even tone as everything around them. Some will be right. The ones that are wrong look identical to the ones that are right. That is the hazard, and it is the reason we treat every drafted claim as something to verify rather than something to publish.

Positioning. Ask a model why a business is different and it will return the same four virtues every competitor also claims. Differentiation is a decision, not a text-generation problem.

How the failure actually shows up

Bad AI content is rarely wrong in an obvious way. It is worse than that — it fails to be wrong at all. It describes the shape of an answer without ever landing on one. Every consideration is acknowledged, no position is taken, and the reader leaves knowing nothing they did not already know.

Practical tests we apply to a draft before it goes anywhere near a website:

• Could a direct competitor publish this paragraph unchanged? If yes, delete it.
• Does it commit to anything a reader could disagree with?
• Is there one specific detail — a number, a name, a situation — that could only have come from someone who has done the work?
• If you cut the first two sentences of each section, does anything of value disappear?

Most drafts fail all four on the first pass. Fixing that is editing, and editing is where the human time goes. The saving is real, but it lands in the drafting hours, not the thinking hours.

Do not use a detector as your gate

There is a market for tools that score text on how machine-written it appears. Published evaluations of these tools keep reporting the same problems: accuracy that varies widely between subject areas, meaningful false-positive rates, and a documented tendency to flag writing by non-native English speakers. Lightly edited text defeats them almost entirely.

A detector score tells you nothing about whether a page is useful, accurate or worth reading. Judge the draft on those things instead. If a piece is specific, correct and says something, its provenance is not the interesting question.

A process that survives scrutiny

1. Decide what the page is for before opening any tool. Which search, which reader, which action at the end.
2. Supply the raw material yourself — notes, transcripts, the actual details of the job.
3. Let the model draft from that material rather than from its own memory.
4. Verify every checkable claim against a source you can name.
5. Add what only you can add: the specific case, the objection you hear every week, the thing you would tell someone in the pub.
6. Cut the passages that could belong to anyone.

That process produces fewer pages than a fully automated one. It also produces pages that hold up when someone reads them properly, which is the only version that keeps earning.

If you want that kind of process running behind your site rather than a volume pipeline, our content work is built around it, and the measurement side tells you whether it is landing.