AI-written articles and SEO: what Google's guidance actually requires

What the guidance says

Google's documentation on AI-generated content takes a position that surprises many content teams: automation is not, by itself, against its guidelines. The published guidance says the ranking systems reward content that is original, helpful and written for people, however it was produced. What it names as a violation is using automation, AI or otherwise, primarily to manipulate rankings.

The same page points readers to the "people-first" criteria: does the content show first-hand expertise, does it leave the reader feeling they learned enough, would it be published without search as a motive. None of those questions mention the author. All of them are about the text and what stands behind it.

Where AI articles fail the test

The failures the trade press reports are not about writing quality. Search Engine Land's coverage of the March 2024 core update describes the target as "scaled content abuse": many pages produced primarily to rank, whatever the method. The Verge documented sites that had published thousands of near-identical explainers and were removed from results within weeks of the update.

The pattern in those cases is the same. Text generated from a model's memory, with no sources behind it, on topics chosen from keyword volume rather than from anything the publisher knew. Press Gazette reports that publishers who lost traffic in 2024 were disproportionately those relying on unsourced explainer content^†. That figure has one outlet behind it in this source set, so it is attributed rather than stated.

Sourcing as a signal, indirectly

Search engines do not read a source list and award points. But almost everything they do reward follows from having one. An article written from five outlets covering the same event contains specifics a generic draft does not: dates, names, figures with attribution. It links out to primary documents. It gets cited by other pages because it is checkable. Those are the behaviours the helpful-content guidance describes, produced as a side effect of the method rather than as an optimisation.

The Reuters Institute's 2025 report adds the reader's side: audiences say they are more comfortable with AI-assisted news when a human has checked it and when they can see where it came from. A visible source list and a human byline are trust signals for people first, and search engines have spent a decade trying to measure what people trust.

A checklist before publishing

Before an AI-drafted article goes live: every factual claim traces to a named source, and single-source claims are attributed in the sentence. A person has read it and their name is on the approval. The title answers a question someone actually asks, not a keyword string. Headings follow the argument, so a table of contents reads like a summary. Structured data declares the article, its author organisation, its dates and its citations. If any of those is missing, the problem is not that a model wrote it. Press Gazette puts the median traffic loss for publishers hit by the 2024 update at 30 percent^†, and that number, too, is attributed rather than stated.

Frequently asked questions