What Google actually says
Worth establishing first, because the amount of confident nonsense in this category is remarkable.
Google's published position is that it rewards high-quality, helpful content regardless of how it is produced, and that using automation to generate content primarily to manipulate rankings is a violation of its spam policies. Those two sentences are doing a lot of work together, and the distinction they draw is about PURPOSE rather than about tooling.
So the useful question is not "will Google know it was AI?" — it is "would this page exist if search engines did not?" A page written to answer a real question that a customer really asks passes that test whatever wrote the first draft. A page that exists because a tool could generate two hundred of them does not, and no amount of humanising will change that.
Where it genuinely helps
In rough order of how much time it actually saves me, which is not the order the marketing suggests.
- Analysis at volume. Clustering a few hundred keywords into intents, spotting near-duplicate phrasings, reading an export and telling you what is in it. The keyword research behind this site involved 883 keywords collapsing into 577 distinct intents — that is a job a machine does in seconds and a person does badly over a day.
- Code and structured data. Generating schema, writing a validation script, drafting a redirect map from two lists of URLs. Deterministic, checkable, and wrong in ways you can see.
- First drafts of things a human will rewrite. Useful precisely because the draft is bad enough to argue with, which is faster than starting from nothing.
- Interrogating your own data. Ask it what changed between two Search Console exports; it will do in a minute what takes an afternoon in a spreadsheet.
- Summarising and reformatting. Turning notes from a client call into a brief, or a long page into the FAQ that should sit at the bottom of it.
- Being the sceptic. Ask it to argue against your own plan. It is unnervingly good at this and it costs nothing.
The pattern: it is good where the raw material already exists and the job is transformation, and bad where the job is to produce something that was not there before.
Where it reliably backfires
- Publishing generated articles at volume. This is the use everyone means and it is the one that fails. A model can only recombine what it has read, so the output contains, by construction, nothing that is not already on ten other pages. It reads fine and ranks for nothing.
- Anything with a number in it. Models produce plausible statistics with total confidence, and a wrong figure on your site is a wrong figure you are accountable for. Every number on this website carries the tool it came from and the date it was read, and there is a whole page explaining why.
- Local content. Ask for a paragraph about doing business in a specific town and you get something that could be about anywhere. This is exactly the thin-content signal that area pages are penalised for.
- Anything requiring first-hand experience. It cannot tell you what a client outcome was, what went wrong on a build, or what you learned in month three. That material is the only thing on your site a competitor cannot copy.
- Meta descriptions at scale, oddly. It will happily write three hundred that are all subtly the same shape, which is worse than the fifty a person writes with attention.
The honest summary is that AI removes the cost of producing average content, which means average content is now worthless. That is bad news for a lot of content strategies and good news for anyone with genuine first-hand material, because the gap between the two just widened.
Getting cited in AI Overviews
A separate question from using AI to make content, and the more commercially interesting one.
On 4 September 2026 I read fourteen live UK search results across the queries this studio cares about. Eight carried an AI Overview — every "what is", "how much" and "…uk" query, and none of the transactional local ones, which carried a local pack instead. On the cost and AI-SEO queries the Overview cited small UK studios rather than the large SEO tools, which is the single most encouraging finding for a business of this size.
What the cited pages have in common, as far as anyone can demonstrate:
- A short, direct, self-contained answer near the top — a paragraph that still makes sense lifted out of its context, because that is exactly what happens to it.
- Figures with their source and date attached in the same sentence. A model repeating a sourced number is repeating something that carries its own defence.
- Genuinely first-hand material. These systems already have everything general; what they do not have is your measurements.
- Consistency about who you are, everywhere you appear.
That is the pattern every guide on this site is written to, which you can verify by reading one. It is also, not coincidentally, what makes a page useful to a person.
A workflow that does not embarrass you
- Do the research with it. Clustering, exports, competitor reading, "what questions does this SERP actually answer". This is where most of the real saving is.
- Decide what the page is for, yourself. If you cannot say what somebody gets from it that they cannot get elsewhere, do not write it — that is the whole editorial decision and no tool can make it.
- Bring the first-hand material. Your numbers, your client outcomes, what went wrong. If you have none for this page, it should not be on your site.
- Draft however you like. Nobody is auditing your process.
- Rewrite every sentence containing a claim, and check every number against its source. This is the step that gets skipped and it is the one that matters.
- Add the sources and the date. If a figure cannot carry one, remove the figure.
- Read it aloud. Generated text has a rhythm — three-item lists, "in today's digital landscape", relentless parallel structure — and reading it aloud finds all of it.
Step five is the whole thing. The tools are excellent assistants and terrible authors, and the difference between the two is whether a person with something to say is accountable for every sentence.
