Google does not penalise content for being written with AI. It penalises content made primarily to rank rather than to help — which is a statement about intent, not about who or what typed it. That distinction has been consistent in Google's own guidance for years, and it is the only part of this subject that has not changed with every model release.
So what exactly is against the rules?
Google's term is scaled content abuse: producing many pages primarily to manipulate rankings, with little value to a reader. A person can do that with a template and a spreadsheet, and plenty did for a decade. AI simply made it cheap, which is why the guidance was restated rather than invented.
Nothing in that definition mentions authorship. A thousand thin pages written by hand break it; one carefully-researched page drafted with AI does not.
What gets a site into trouble?
| Practice | Risk | Why |
|---|---|---|
| Publishing forty near-identical pages because it was cheap | High | This is the definition of scaled content abuse |
| Statistics attributed to research that does not exist | High | A reader can check it, and it is a claim about somebody else |
| "We surveyed 1 000 businesses" when nobody surveyed anybody | High | Fabricated first-person research is the least defensible thing on a site |
| Location pages spun from a template with the place name swapped | High | Thin, and obviously so |
| Fluent, confident, subtly wrong detail | Medium | Reads exactly like accurate detail. That is the danger |
| A well-researched page drafted with AI and edited by a person | None | Google has said so repeatedly |
What does Google reward instead?
- First-hand experience. The thing a model cannot produce on its own: what happened when you did this, with your customers, in this market.
- Specificity. Real numbers, real constraints, real trade-offs. Generic advice is the clearest signal that nobody was accountable for the page.
- Sources. Claims that link to where they came from — which matters more than it used to, because it is also how AI answer engines decide what to cite.
- Accountability. A named author or an identifiable organisation standing behind it.
- Currency. A page dated two years ago and never revised loses to one checked last month.
Does the same advice work for AI search?
Mostly, and the overlap is larger than people expect. Answer engines lift passages rather than ranking pages, which changes what a good page looks like in three specific ways: lead with a direct answer instead of building to it, put comparisons in a table rather than in prose, and cite sources so a claim can be checked. All three also make the page better for a human, which is the useful part.
The genuinely new requirement is machine-readable structure — schema markup, a clean heading hierarchy, and files like llms.txt that state plainly what a business does and what it charges.
How should a South African business use AI for content?
As a drafting and structuring tool with a person in front of it and a person behind it:
- Use it for structure. Working out what questions a page should answer, in what order, is where it is genuinely strong.
- Bring the substance yourself. Your prices in rands, POPIA, the eleven official languages, load-shedding, SARS deadlines, how the market actually behaves here. That is the half a general model does not have, and it is the half that gets cited.
- Check every number and every source. If a draft names a report, open the report. If you cannot find it, the claim goes — do not soften it, remove it.
- Do not publish what you have not read. Obvious, and it is the step that gets skipped when the tool makes ten drafts an hour.
- Update rather than multiply. One page kept current beats four covering the same ground.
Fluency is not accuracy, and they read exactly the same. That is the entire risk, and the entire reason a person has to read it before it ships.
What is the shortest version?
Draft with AI, verify everything, publish fewer and better pages, keep them current, and put the local substance in yourself. Do that and the tool is irrelevant to Google — which is precisely what Google has been saying all along.