What Google actually says
Google’s published guidance is about purpose, not tooling: it rewards helpful, reliable content made for people, and acts against content produced primarily to manipulate rankings. Using automation to mass-produce pages for search is in the second category; using a model to help write something genuinely useful is not.
That is a sensible position and it is worth reading it precisely, because both camps misquote it. It is not “AI content is penalised”, and it is not “anything goes”. It is that the question was never how the words got onto the page.
What a model genuinely does well
- Structure. Turning a messy brain-dump into headings in a sensible order is real work and it does it in seconds.
- The second, third and fourth draft of a sentence you already know is wrong but cannot fix.
- Variations — twelve versions of a heading so you can see which register fits.
- Compression. Cutting 400 words to 150 without losing the argument is a genuinely hard skill and it is good at it.
- The boring necessary things: meta descriptions, image alt text, a FAQ drawn from questions you supply.
Used that way it is a substantial time-saver on the parts of writing that are labour rather than thought. Nobody should feel clever about doing those by hand.
The three things it cannot do
These are not limitations that a better model fixes. They are the same limitation described three ways.
- It does not know your numbers. What you charge, how long you take, what your conversion rate is, what happened on the last job. It will invent a plausible figure if you let it, and a plausible figure is worse than none.
- It has not spoken to your customers. It does not know that everybody asks about parking, or that the objection is always the lead time, or that the phrase your customers actually use is not the industry one.
- It has no stake. It cannot say “I would not do that if I were you”, because it is not you and it does not care what happens next.
Which is why generated copy so often reads as smooth and weightless. Nothing in it could be wrong, because nothing in it is specific enough to be wrong.
The test for any page, however it was written
Could a competitor publish this page unchanged, with their name on it?
If yes, it is not doing anything. That was true long before language models existed — most agency writing has always failed this test — but generated copy fails it faster and at greater volume, which is the actual risk.
The way to pass it is unglamorous: put things on the page your competitors will not. Your prices. Your timescales. What goes wrong and how often. What you would tell somebody not to bother with. None of those can be generated, and all of them are what a customer was searching for.
A workable process
- Write the facts first, badly. Bullet points, no sentences: the prices, the timings, the objections, the things that went wrong, the questions you get every week.
- Have the model turn those into a structure and a first draft. It will do this well because you have given it something to work with.
- Cut everything it added that you did not supply. This is the important step and it is where most of the value is — anything it invented is either generic or wrong.
- Read it aloud. If it does not sound like you talking to a customer, it will not sound like you to a customer.
- Add the one thing only you could have written. There should be at least one sentence per page that no competitor could have published.
Step three is the one people skip, and it is the difference between a useful tool and a page that reads like everybody else’s.
And it matters more now, not less
When we read 31 live UK results on 4 September 2026, an AI Overview sat above 25 of them. A summary can now answer a general question on the results page, which means a page of general answers earns nothing at all.
The encouraging half of the same finding: on the cost and agency queries, those Overviews were citing small UK web studios rather than the large SEO tool blogs. Specific, sourced and first-hand beat big and general.
Which turns the whole question round. The risk is not that you use AI to write your website. It is that your website says nothing an AI could not have said without you.
