Every marketing team has now had the same meeting. Someone points out that a writer costs money and a model costs cents, someone else says Google will penalise it, and nobody in the room can quote the actual policy. The result is usually either a cautious ban or an enthusiastic content dump — and both are decisions made without reading the rule.
The takeaway up front: Google's position is that it rewards high-quality content however it is produced, and that its spam policies target scaled content abuse — producing many pages primarily to manipulate rankings rather than to help people. The method of production is not the trigger. The intent and the outcome are. A page written by a model that genuinely answers the query is fine. A thousand thin pages assembled to catch keywords are a problem whether a human or a machine typed them.
That distinction is the whole article, but the practical consequences are worth spelling out, because the failure mode most sites hit is not a penalty. It is being ignored.
What the policy actually says
In March 2024, Google updated its spam policies and named three practices explicitly:
- Scaled content abuse. Generating many pages primarily to manipulate search rankings rather than to help users. This replaced the older "spammy automatically-generated content" wording, and the rewrite mattered: the old phrasing let people argue about automation, while the new one puts the emphasis on scale plus intent, regardless of how the text was produced.
- Site reputation abuse. Publishing third-party content on an established domain mainly to exploit that domain's ranking signals — the coupon-section-on-a-newspaper pattern, often called parasite SEO.
- Expired domain abuse. Buying a domain with existing authority and repurposing it for unrelated content to inherit its signals.
At the same time, the helpful content system stopped being a separate periodic update and was folded into Google's core ranking systems. That change is easy to miss and important: "helpfulness" is no longer something you get judged on every few months. It is assessed continuously as part of ordinary ranking.
Google's own guidance on AI-generated content, published a year earlier, said the same thing in plainer terms: focus on quality, not the means of production. It also noted that using automation to generate content for the primary purpose of manipulating rankings has always been against its policies — and always will be.
So: nobody gets demoted for using a model. Sites get demoted for the pattern of behaviour the model made cheap.
Why most AI content programmes stall anyway
Penalties are rare. The common outcome is duller and more expensive: pages that get crawled, sometimes indexed, and never rank. Four causes account for nearly all of it.
It says what forty other pages already say
A language model produces the consensus of its training data. That is exactly the text already sitting in the top ten results. Publishing a competent restatement of the consensus gives a search engine no reason to prefer your page. This is not a duplicate-content penalty — that particular myth is dealt with in our piece on whether duplicate content hurts SEO — it is simply an absence of any differentiating value.
It has no first-hand experience in it
Google's Quality Rater Guidelines describe E-E-A-T: Experience, Expertise, Authoritativeness and Trustworthiness. The first E was added deliberately, and it is the one a model structurally cannot supply. It has not installed the product, priced the job, handled the complaint or run the migration. Original photographs, real numbers from your own operations, named customers, a process you actually follow — these are the components that separate a page from its competitors, and every one of them has to come from the business.
It gets crawled and dropped
Publish a few hundred near-identical pages and a common result is a swelling "Crawled – currently not indexed" or "Discovered – currently not indexed" count in Search Console. Google is allowed to decide a page is not worth storing. Low-value pages also consume crawl budget on larger sites, which slows the discovery of pages you actually care about. The reasons pages fail to get indexed are worth understanding before you multiply your page count.
It was never aimed at a real query
Volume tempts teams to skip the research step. A page written to a topic rather than to a demonstrated query, with a clear intent behind it, has no audience to arrive from. Cheap production makes this worse, not better, because the cost of publishing something pointless approaches zero.
A workflow that survives contact with Google
The point of using AI in content is to remove the mechanical parts of writing, not the thinking. In practice that means front-loading human judgement and keeping human accountability at the end.
1. Research and intent, done properly. Every commissioned page starts from a query with real demand and a classified intent — informational, commercial, transactional, navigational. If you cannot say what the searcher wants and what they should be able to do afterwards, the page is not ready to be written.
2. A brief that carries what the model cannot know. This is where the leverage is. A brief containing your pricing model, your process, your constraints, the objections your sales team hears, and the two or three claims only you can make will produce a materially better draft than a topic and a word count. Our guide to writing a content brief that produces a page that ranks covers the structure.
3. Draft with assistance, then add the things that are yours. Use the model for structure, first drafts and tightening. Then add the original material: your data, your screenshots, your worked example, your named expert.
4. Verify every checkable claim. Models produce confident, wrong statistics and cite studies that do not exist. Every number, date, regulation and citation gets checked against a primary source or removed. This is not optional; a single fabricated figure discovered by a prospect costs more than the page ever earned.
5. A named human signs it off. Someone with relevant expertise reviews, corrects and takes responsibility for accuracy. That is what an author byline is supposed to mean.
6. Publish at a pace your review capacity supports. The binding constraint on a content programme is not writing throughput — it has not been for a while. It is expert review, original inputs and internal linking. Scale the output to the constraint and you get compounding. Scale past it and you get an archive of pages nobody reads.
7. Measure at the page level. Track impressions, clicks and average position per URL, and check indexation status. Pages that gather impressions but no clicks usually have a title or intent-match problem. Pages that gather nothing after a couple of months were probably never differentiated. Prune or rewrite them rather than leaving them to dilute the site.
Where the honest line sits
There is no threshold number of AI-assisted pages that trips a filter, and anyone selling you one is guessing. The workable test is a question you can answer about your own archive: would this page still be worth publishing if search engines did not exist? If yes, production method is irrelevant. If no, no amount of editing rescues it.
That test also explains why the "AI content is fine, actually" and "AI content is doomed" camps both keep finding evidence. They are looking at different archives.
FAQ
Will Google penalise my site for using AI to write articles?
Not for the use of AI itself. Google has stated that it rewards quality content regardless of how it is produced. What its spam policies target is scaled content abuse — producing pages primarily to manipulate rankings rather than to help people. A small number of genuinely useful AI-assisted pages sits well within the policy.
Can Google detect AI-written content?
Detection is unreliable, and third-party detectors produce false positives on ordinary human writing. More to the point, it is the wrong question: ranking systems assess whether a page is useful, original and trustworthy. A page that fails those tests does poorly whoever wrote it, and detection never has to enter into it.
Do I need to disclose that content was AI-assisted?
There is no ranking requirement to do so. Google's guidance is that disclosure is helpful where a reader would reasonably wonder how something was produced. Many publishers add a short editorial-standards note covering how AI is used and who reviews the output; it costs nothing and builds trust.
How much content can I safely publish per month?
There is no safe number, because the risk is not volume in isolation. The practical ceiling is how many pages you can supply with original input and put through genuine expert review. Publish to that limit and the volume takes care of itself.
What should I do with the thin AI pages I already published?
Audit them by performance and quality. Rewrite the ones on topics that matter to the business, adding original material and expertise. Consolidate near-duplicates into one stronger page with redirects. Remove the rest. A smaller archive of pages that earn their place is easier to rank than a large one that does not.
Do it properly, or don't do it at scale
AI has made drafting cheap and left everything that actually ranks — original experience, verified facts, expert review, real research — exactly as expensive as it was. The teams getting results are the ones that reinvested the saving into those parts rather than pocketing it as volume. The wider ordering of that work is covered in our SEO strategy guide.
If you would rather the research, briefing, production and review sat with someone else, see how WeSEO's fixed-price packages and transparent work log work — fixed scope, visible deliverables, and no ranking promises.
