The Google AI content guidelines are shorter than most of the articles written about them. The position has not really moved since Google published it in February 2023: automation is fine, content made primarily to manipulate rankings is spam, and the quality bar is the same bar everything else has to clear. What nobody hands you is the practical bit. How do you take that guidance and run an actual draft through it on a Tuesday afternoon before it goes live?

That is what this piece is. Not another paraphrase of the policy, but the policy treated as a specification, with the checks we run every week on our own pipeline.

What Google's AI content guidelines actually say

Google's guidance is a purpose test, not a production test. Read the Search Central post on AI-generated content and the operative phrase is "using automation, including AI, to generate content with the primary purpose of manipulating ranking in search results". Primary purpose. Not method, not tooling, not percentage of machine-written words.

So the useful question about a draft is never "was this written by AI". It is "does this page contain anything that required knowing something". If the answer is no, the page has a problem whether a human typed it or not.

The three documents that matter

People talk about "the guidelines"as if there is one page. There are three, and they do different jobs.

• The AI-generated content guidance. The short one. It sets the purpose test and confirms that appropriate use of automation is not against Google's policies.

• Creating helpful, reliable, people-first content. The self-assessment questions, including the one that does most of the work: does the content provide substantial value when compared to other pages in search results? This is the quality specification.

• The spam policies. Where scaled content abuse and site reputation abuse are named explicitly. This is the enforcement specification, and it is the one with teeth.

Quality problems cost you rankings. Spam policy breaches can cost you the site. Keeping the two separate in your head stops a lot of panic.

Three things people get wrong

Disclosure is not mandatory. Google has not required you to label AI assistance for search purposes, and nothing in the guidelines turns an unlabelled page into a violation. Disclose if your readers would want to know, or if your sector has its own rules. Do not do it because you think a byline note buys you safety.

There is no detection score. Google has not published an AI-detection signal, and third-party detectors measure predictability of phrasing, not policy compliance. A detector flagging 90% tells you the prose is generic. That is worth knowing, but it is a style report, not a verdict.

And a "human edit"satisfies nothing on its own. Fixing commas in a draft that says the same thing as results one to ten leaves you with a tidier version of the same problem. We have covered what actually gets flagged in more detail in our piece on whether AI content is penalised by Google.

Why E-E-A-T sits underneath all of it

Experience is the first E, and it is the hardest element to fake at volume. A model can approximate expertise by assembling what has already been written. It cannot tell you what happened when a customer's boiler warranty claim was rejected on a technicality, or what the quote actually came back at. That is the gap AI drafts fall into most reliably, and it is also the gap readers notice first.

The gap between allowed and rankable

Permission is not promotion. Passing the AI content guidelines means your page is not automatically a policy breach. It does not mean the page deserves a position, and treating compliance as a ranking strategy is how teams end up with forty clean, compliant, invisible articles.

Scaled content abuse hinges on purpose and volume together. Consider two local trades businesses with the same annual word count. One publishes forty near-identical "[service] in [town]"pages in a single week, each with the place name swapped and the rest untouched. The other publishes one researched article a week for ten weeks, each answering a question customers actually ask. Same output on a spreadsheet. Completely different risk profile, and completely different reader experience. If you want the full list of what sets this off, our breakdown of the spam update triggers for AI content goes through them one by one.

Originality in Google's self-assessment is competitive, not absolute. "Substantial value compared to other pages in search results"is measured against the live top ten for that query, today. Which means originality is whatever remains after you strip out everything the current results already cover. We plan every brief against the actual top ten for exactly this reason: it tells us what is already said, so the article can be the part that is missing.

Most thin AI output fails long before anyone reaches for a policy. It restates the same eight points as the existing results, in a slightly different order, in prose that sounds like nobody in particular.

A pre-publish checklist for any AI-assisted draft

Four checks, one draft, roughly 15 to 20 minutes once you are used to it. Run them in this order.

1. The specificity pass

Go section by section and look for one fact, number, named standard, price, date or first-hand observation per section that could not have been written without knowing this business or this sector. Not one per article. Per section.

Sections that fail this pass are usually the ones hedging: "there are several factors to consider", "costs vary depending on your requirements". Either replace the sentence with the actual factors and the actual range, or delete the section. This single check catches more policy risk than any detection tool, because it is testing the same thing Google's guidance is testing.

2. The duplication pass

Cannibalisation is the quiet failure mode of AI publishing. A team generates five posts on overlapping topics in a month, each one perfectly decent, and splits its own rankings five ways. Before writing, search your own domain for the target query and the two nearest variants. If an existing page already ranks for it, you are updating that page, not writing a new one.

Checking takes two minutes. Merging those five duplicate posts a year later takes an afternoon and loses you the links.

3. The claims pass

Every statistic, percentage, price and date is either sourced in the same sentence or cut. Language models produce plausible-looking figures with great confidence, and "a recent study found"with no study attached is the most common fabrication in AI drafts. Use UK sources for UK readers: ONS, Ofgem, HMRC, the relevant trade body. A number without a source is worse than no number, because it is the thing a sceptical reader checks first.

4. The structure and schema pass

Read the first screen on a phone. Does it answer the query, or does it clear its throat for 300 words? Pages with impressions and almost no clicks are usually failing here, not in the body copy.

Then check the markup. Article and FAQ schema where relevant, and only one copy of it. If Yoast, Rank Math or a similar plugin is already emitting Article schema, adding your own block gives you duplicate or conflicting markup. Validate with Google's Rich Results Test rather than assuming.

Finally, read two paragraphs aloud. You will hear the problem before you see it.

How AI drafts give themselves away

Most AI writing has a tell, and it is rhythm more than vocabulary. Every paragraph three sentences long. Every sentence the same shape, subject then verb then tidy qualifier. A conclusion that restates the article three times in slightly different words, because the model was trained to round things off.

Human expert writing is lumpier than that. A one-line paragraph where the point lands. An aside halfway through a section because the writer remembered a case that contradicts the general rule. Sentences of four words next to sentences of forty.

Keep a working banned-phrase list and enforce it. Ours includes the usual suspects, and the discipline matters more than the list: when you cannot write "plays a key role in", you are forced to say what the thing actually does. Cutting a phrase that means nothing does not give you a different empty sentence. It gives you a better one, because you have to fill the gap with information.

The practical risk here is not a manual action. It is readers leaving. People bounce from content that says nothing, and a site full of competent, generic articles trains your audience to ignore you long before any algorithm weighs in.

Editorial controls that make compliance repeatable

Checklists work once. Controls work every week. These are the ones we run in our own pipeline, and you can run versions of them with a spreadsheet and a stubborn editor.

Grade every draft twice. One score for the search job: keyword coverage, structure, internal links, intent match. A separate score for content integrity: accuracy, originality, voice. When the two disagree, integrity wins. A page that games structure perfectly but says nothing fails the helpful content test anyway, so promoting it on the strength of its SEO score is just a slower way of getting the same result. Our editorial process sets this out in full.

Put a sceptical sub-editor in the way. Human or process, the job is one question asked of every claim: how do we find that? Numbers without sources, superlatives without evidence, outcomes stated as certainties. All out.

Profile the brand voice from real writing. Not an adjective like "friendly"or "professional", which tells a writer nothing. Actual sentences from your site, your emails, your proposals. If the site is thin and there is barely anything to learn from, say so and build the profile deliberately instead of pretending the data exists.

Ramp your cadence. Publishing volume is part of compliance. A site that has managed two posts in two years and then drops fifty in a fortnight looks like a scaled content operation regardless of how good the fifty are. Start at one a week. Increase when the pattern is established.

Feed real knowledge into briefs. Service specifics, pricing structures, the three objections you answer on every sales call, the job that went wrong and what you changed. Drafts that start from something only you know end up with something only you could have written.

If you have already published AI content that is not holding up

Triage first, in Search Console, before you touch a single page.

Pages with impressions and near-zero clicks are usually failing on the first screen and the title: Google is showing them, people are not choosing them. Rewrite the opening, sharpen the title, add the specific detail the snippet is missing. Pages with no impressions at all after 90 days are a different problem. More rewriting rarely fixes those; they usually need merging into a stronger page on the same topic, or removing. Pages making unsupported claims go to the top of the queue regardless of traffic, because those are the ones that damage trust.

Merge ruthlessly. Three overlapping posts become one good page with redirects from the other two. Then rebuild coverage properly: a genuine cluster with internal links that reflect how the topics actually relate, rather than another ten articles chasing variations of the same query. If an update has hit you, the sequencing in our guide to recovering from a core update applies here too.

Recovery takes months, not weeks, and it often needs a subsequent update to be reassessed. "Publish more"is almost always the wrong first move, because volume is what created the pattern. Fix what you have, then start publishing again at a credible pace.

Used this way, Google's AI content guidelines stop being something to worry about and become the specification you write to: a purpose test you can answer honestly, section by section, before anything goes live. If you would rather have that run for you every week, that is what our content marketing platform does.

Frequently Asked Questions

Does Google penalise AI-generated content?

No. Google's guidance, first published in February 2023, says appropriate use of automation is not against its policies. What breaches the spam policies is content produced primarily to manipulate rankings, whoever or whatever produced it. Badly written, unoriginal pages lose rankings on quality grounds, which is a separate problem from a policy breach.

Do the Google AI content guidelines require you to disclose AI use?

Not for search purposes. Google has not made AI disclosure a requirement, and omitting it does not create a policy problem. Disclose where your readers would reasonably expect it, such as regulated sectors or editorial sites with a stated policy, and because it is honest, not because you think it protects your rankings.

What counts as scaled content abuse under Google's spam policies?

Generating many pages primarily to game search rankings rather than to help people, whether by automation, by humans, or by a combination. Purpose and pattern matter more than raw word count: forty near-identical location pages published in a week fits the description, while ten researched articles over ten weeks does not. Site reputation abuse is named separately and covers third-party content published on an established domain to exploit its ranking signals.

Can AI-written articles rank on the first page of Google?

Yes, and plenty do. The deciding factor is whether the page offers substantial value compared with the other results for that query, which is a competitive test against the live top ten rather than a fixed standard. Drafts that start from original research, real figures and first-hand detail clear it. Drafts that restate the existing results do not.

How do you check an AI-assisted draft before publishing it?

Run four passes, which takes about 15 to 20 minutes with practice: a specificity check for at least one concrete fact or observation per section, a duplication check against pages you already have, a claims check where every figure is sourced or cut, and a structure and schema check covering the first screen and any duplicate markup from your SEO plugin. Reading two paragraphs aloud catches the rhythm problems the rest miss.