AI content and the Google guidelines.

The documentation on generative AI is strikingly short and formulates neither a ban nor a licence, but a condition. Two sections carry the whole message: "Focus on accuracy, quality, and relevance" and "Give users context".

On where it may be used, Google writes that generative AI "can be particularly useful when researching a topic, and to add structure to original content". That is the only positive purpose Google names, and it is remarkably narrow: research and structure, not producing the content itself.

Right next to it stands the limit: "using generative AI tools or other similar tools to generate many pages without adding value for users may violate Google's spam policy on scaled content abuse". And as an overarching requirement: "make sure your work meets the standards of the Search Essentials and our spam policies".

The helpful-content guidance is even more direct. It contains the sentence against which every content decision can be measured: "If you use automation, including AI-generation, to produce content for the primary purpose of manipulating search rankings, that's a violation of our spam policies." The criterion is the primary purpose. A page that exists because a search volume exists, rather than because someone wanted to answer a question, is exactly the case the rule is aimed at.

Since the March 2024 core update the term has replaced the earlier rule on automatically generated spam content, and it is deliberately worded to be technology-neutral. Google's definition: "Scaled content abuse is when many pages are generated for the primary purpose of manipulating search rankings and not helping users." The documentation gives five examples:

• "Using generative AI tools or other similar tools to generate many pages without adding value" — the case at issue here.
• "Scraping feeds, search results, or other content to generate many pages … where little value is provided" — harvesting third-party sources automatically and republishing them.
• "Stitching or combining content from different web pages without adding value" — assembling blocks of text from several sources.
• "Creating multiple sites with the intent of hiding the scaled nature of the content" — spreading the same pattern across several domains.
• "Creating many pages where the content makes little or no sense to a reader but contains search keywords" — text that exists only for the search engine.

What is missing from that list is notable: a number. Google names no page count at which things become critical, because volume on its own says nothing. An online shop with three thousand product pages full of real data is unproblematic. Twenty location pages with identical text and a swapped-in municipality name are not — and they additionally fall under the doorway rule: "Doorway abuse is when sites or pages are created to rank for specific, similar search queries."

That distinction is why we build every location page on this website with its own content, its own references and its own wording instead of generating it from a template. What that looks like is visible, for instance, on the page for Winterthur compared with the one for Zug. The effort is considerably higher — and it is the only route that holds up over years.

E-E-A-T is not a ranking factor you can switch on. It is the framework Google's quality raters use to assess pages — and for AI-assisted content the first E is the critical one. Experience, Expertise, Authoritativeness and Trustworthiness do not sit side by side as equals; the documentation singles out trustworthiness as the most important aspect.

Experience is the property a language model structurally cannot supply. It has never written a quotation, never watched a project fail on a misunderstood requirement, never negotiated a price with a client. Yet that material is exactly what decides whether a text is useful. The quality rater guidelines state the flip side as a criterion for the lowest rating: main content created with "little to no effort, little to no originality, and little to no added value".

In practice this translates into four checkable requirements: visible authorship with a plausible connection to the subject; verifiable statements instead of assertions; clear accountability in the legal notice, including address and contact details; and thematic fit between what a website claims and what the company actually does. The definition of the term is in the glossary under E-E-A-T; further background is collected in our glossary.

One more piece of context: with the March 2024 core update, Google folded the previously separate helpful content system into its core ranking systems. Since then there is no separate switch to be flipped — the assessment of usefulness and originality sits inside normal ranking.

The dividing line does not run between human and machine. It runs between content with an own contribution and content without one. In our own work the following split has proven itself.

Unproblematic: research and source review with AI; drafting outlines and lines of argument; producing first drafts that are then enriched and checked by a subject expert; variants for titles, meta descriptions and subheadings; translations followed by editing from someone who commands the target language; summaries of your own longer documents; checking an existing text for gaps, contradictions and missing internal links.

Risky to non-compliant: publishing whole pages unchecked; generating pages from a template that differ only in one swapped variable; rewriting content from other websites and passing it off as your own; producing large volumes of articles because a tool spat out a keyword list; inventing testimonials, reviews or case numbers; publishing health or finance topics without specialist review.

The most expensive mistake on that list is not the guideline breach but the invented fact. A language model phrases a non-existent study just as convincingly as a real one. When a client discovers an error like that, it costs more trust than ten extra pages could ever bring in visibility. How to catch this in day-to-day editorial work is described in the article on AI content strategy for Swiss SMEs.

Google does not require a disclosure but explicitly recommends one — and other rulebooks may well create an obligation. The helpful-content guidance asks, under the heading "How": "Is the use of automation, including AI-generation, self-evident to visitors through disclosures?" The guidance on generative AI adds that information about how a piece of content came about can give readers additional context.

Strictly separate from that are obligations from other sources. Contracts with clients, editorial guidelines, university examination regulations and tender conditions each settle the question themselves. On top of that comes European regulation: Article 50 of the EU AI Act requires providers of generative systems to mark synthetic output in a machine-readable way and has applied since 2 August 2026. Switzerland is not bound by it; Swiss companies with EU customers effectively are. That duty falls primarily on the providers of the systems, not on every website — but it explains why watermarks have been appearing in AI text output since August 2026, as we set out in detail on cleaning up AI text.

Our recommendation is pragmatic: a short, honest note about your editorial process on a page of its own is worth more than a disclaimer under every article. It answers Google's "How" question, it builds trust, and it forces an internal decision that has to be made anyway.

Use AI before and after the writing, not instead of the writing — and anchor three control points that are not negotiable. The workflow we run ourselves has six steps.

• 1. Check demand before a page exists. Does the question really get asked, and is nobody answering it well yet? If the answer to the second part is no, you are about to create a redundant page. This check is the most effective protection against scaled content abuse, because it caps the volume from the start.
• 2. Collect substance. Figures, price ranges, one real example, one typical mistake. Ten minutes with the person who owns the subject is usually enough. Without this step you get interchangeable text, no matter who writes it.
• 3. Produce a draft, fix the structure. This is where AI plays to its strengths: outline, completeness, alternative phrasings.
• 4. Enrich by hand. The collected substance goes into the text, adjectives give way to facts, and one section names the limits of the service. Passages like that are quoted by language models at an above-average rate.
• 5. Fact-check against the original source. Every figure, every study, every name. No exceptions, no shortcuts.
• 6. Assign accountability. One person per page who would defend the content in a client meeting. That is also the most practical implementation of the "Who" in Google's test questions.

The three non-negotiable control points are steps 1, 5 and 6: no page without a real question, no figure without a source, no content without someone accountable. Keep those three and you can automate the rest of the process as you like without putting visibility at risk. What the same work adds in terms of citability in AI answers is described under GEO Agency Switzerland; the ongoing implementation is what we handle within ongoing SEO support.

Google's documentation answers fewer questions than it is often credited with — and knowing the gaps prevents false confidence.

• They are not legal advice. Copyright, data protection, advertising law and sector-specific rules apply regardless of what Google writes about ranking. For a binding assessment you need a lawyer, not an SEO agency.
• They say nothing about disclosure duties towards third parties. What a client, a university or an editorial team requires follows their rules. That question belongs settled before delivery.
• They guarantee no visibility. Compliance is the precondition, not the result. A flawless page in a crowded competitive field can still sit on page three.
• They are no protection against your own mistakes. The greatest damage comes not from a guideline breach but from wrong statements on your own website — an invented fact costs trust, not positions.
• They change. The wordings on this page reflect the position as of 19 August 2026. Google has rephrased several times between 2024 and 2026, most recently in May 2026 with the clarification that the spam policies apply to the whole of Search, generative answers included. When in doubt, check the original documentation.

And one honest observation to close: for most smaller companies the guideline question is not the real problem. The real problem is that too little content gets created and what exists is too generic. Anyone with ten good pages instead of a hundred mediocre ones need not worry about scaled content abuse — and usually ranks better. Further groundwork is collected in our knowledge section and in the glossary.