What Google actually says
Google’s guidance on AI-generated content, published in February 2023 and unchanged in substance since, says its ranking systems reward original, high-quality content demonstrating expertise, experience, authoritativeness and trustworthiness, “however it is produced”. Using automation, including AI, to generate content with the primary purpose of manipulating rankings is a violation of the spam policies; using it to help create helpful content is not.
In March 2024 Google added a specific spam policy, scaled content abuse, covering the generation of many pages primarily to manipulate rankings and not to help users, and named AI generation as one way that happens. Manual actions under it have been applied to sites since. The policy is about scale and purpose, not tools.
| Penalised? | Source | |
|---|---|---|
| Content written with AI assistance, edited, with first-hand value | No | Google Search Central, AI-generated content guidance, February 2023 |
| Content written entirely by AI, unedited, but genuinely helpful | No, in principle; in practice it rarely is helpful | Same |
| Many pages generated primarily to rank, whatever the tool | Yes: scaled content abuse | Google spam policies, March 2024 |
| Content showing no first-hand experience or expertise | Demoted by helpful-content systems, not a manual penalty | Google, helpful content system documentation, 2022 onward |
| AI-invented facts and figures | Demoted for unreliability; a trust problem before a policy one | E-E-A-T guidance in the Search Quality Rater Guidelines |
Source: Google Search Central documentation as linked in the sources; summarised, not quoted at length.
What actually gets demoted, in practice
- Pages that could have been written about any business. “We are passionate about delivering exceptional results” is the sound of a model that was not told anything.
- Pages with numbers that are not sourced. Models produce plausible statistics; Google’s systems and readers both weigh unsourced claims as noise, and an invented figure discovered is a trust collapse.
- Dozens of near-identical pages: one per town, one per phrasing of a question. Whether a person or a model wrote them, they are thin duplicates.
- Sites where the ratio of generic to first-hand content tips. The helpful-content signal is site-wide; a hundred hollow pages can drag down the twenty good ones.
- Content without a person behind it. No author, no photo, no history is a weak trust signal on any topic where trust matters.
What ranks, and the rules this site uses
Every guide on this site is drafted with AI assistance and none of them is AI content in the sense the policy targets, because each follows six rules that a model cannot satisfy on its own.
- One distinct question per page, chosen because real people search it (the volume and the People Also Ask entries are in each page’s sources).
- A figure from a real source in every piece: Search Console for a named client, browser timing on a live site, search demand data read on a stated date.
- Something only we could know: what went wrong on a job, what a client reported, what our own domain’s Search Console showed.
- A named author, a date, and a source list a reader could check.
- The answer in the first paragraph, written to be quoted rather than to fill space.
- Nothing rewritten from another agency’s page.
And AI answers in the results?
On 8 of 14 UK results we checked on 4 September 2026, Google showed an AI Overview above the links, and on cost and how-to questions it cited small UK sites with clear, dated, sourced answers. The pages being cited are exactly the ones the rules above produce. AI content that says nothing is not cited by AI answers either; the systems generating them need a source with something in it.
