AI-Generated Content and Google Rankings: Where the Line Actually Sits

Every prospective client asks us some version of this question, usually in the first ten minutes of a call, and usually because they have read two contradictory things that morning: one post insisting Google punishes anything a model touched, another insisting rankings are a solved problem if you buy the right prompt pack. Both are wrong, and both are wrong in ways Google has documented in public, dated policy pages. This piece walks through those documents, the demotion data from 2024 and 2025, and the three conditions that separate AI-assisted pages that rank from AI-generated pages that get buried. We run an AI-heavy content operation ourselves, so we have a direct commercial interest in getting this answer right rather than comfortable.

Google's written policy: production method is not a ranking factor

Google's position has been on the record since February 2023, when the Search Central blog published its guidance on AI-generated content. The core statement: Google rewards high-quality content "however it is produced," and appropriate use of AI or automation does not violate its guidelines. The same document draws the boundary in the next breath: using automation, including AI, to generate content "with the primary purpose of manipulating ranking in search results" is a violation of its spam policies.

Read carefully, that is not a permission slip. It is a shift of the test from how was this made to why was this made and what does it give the reader. Google evaluates the page against E-E-A-T, its shorthand for experience, expertise, authoritativeness, and trust, and against the helpful content signals that now run inside the core ranking systems. A model can produce text that passes those tests. A model prompted with nothing but a keyword cannot, because it has no experience to report and nothing original to add.

March 2024: the policy grew teeth

The March 2024 core update and its accompanying spam policy changes are the moment the abstract principle became enforcement. Google formalized scaled content abuse as a named spam policy: producing many pages, by any method, whose purpose is manipulating rankings rather than helping people. The wording was deliberately method-neutral. Human content farms and AI content farms fall under the same clause.

The enforcement was not subtle. Google stated the combined changes would reduce low-quality, unoriginal results by about 45 percent, and site owners watched entire domains deindexed within days of the announcement. Follow-up industry analyses through 2024 and 2025 found that the sites removed or demoted shared a profile: high publishing velocity, near-identical page structure, no identifiable authors, and text that summarized existing results without adding anything. Authorship by AI was common in that group, but it was the pattern, not the tool, that Google named in the policy.

January 2025: the rater guidelines put it in writing

Google's Search Quality Rater Guidelines are the training manual for the roughly 16,000 external raters who score sample results so Google can measure whether its algorithms are working. Ratings do not move individual pages up or down; they calibrate the systems that do. That makes the guidelines the closest thing to a published spec of what Google is tuning toward, and the January 2025 revision addressed generative AI directly for the first time.

Two additions matter here. First, the guidelines define generative AI plainly and call it "a useful tool" while noting it can be misused. Second, and more consequentially, the Lowest page quality rating now explicitly applies when all or almost all of a page's main content is AI-generated "with little to no effort, little to no originality, and little to no added value" for visitors. The rater overview document pairs this with a long-standing principle: main content is high quality when it shows significant effort, originality, and talent or skill. Effort is the operative word in both directions.

We keep a full breakdown of the rater framework, including the E-E-A-T scoring process and the YMYL standard, in our quality rater guidelines explainer. The short version for this question: raters are not asked to detect AI. They are asked to judge effort, originality, and added value, and to bury pages that fail all three regardless of who or what wrote them.

What the ranking data shows, independent of policy

Policy tells you Google's intent. Crawl data tells you the outcome, and two large 2024-2025 studies point the same direction from different angles.

Ahrefs analyzed 600,000 pages in April 2025 and found essentially no correlation between the detected share of AI-generated text and ranking position across the top 20 results. Pure AI pages were rare in the sample; lightly and heavily AI-assisted pages ranked everywhere. Semrush's study of roughly 42,000 blog posts, published in late 2024, adds the asterisk that matters commercially: human-written content held the number one position about 80 percent of the time, against 9 percent for purely AI-generated posts, with the gap narrowing from position five onward.

Put together, the two findings are not contradictory. AI assistance does not suppress a page. Pure generation rarely wins position one, because position one usually demands something the model was never given: original data, first-hand experience, a real point of view. The studies measure the same boundary the rater guidelines describe, from the outside.

What the demotions actually hit

Our own research team spent early 2026 dissecting the post-2024 demotion pattern, and the findings are published in our anti-slop research report. The sites that lost visibility shared traits that are easy to list and expensive to fix: hundreds of pages shipped per week, interchangeable intros, headings lifted from whatever already ranked, no named authors, no sources, no numbers a reader could check. Several were technically flawless. Schema validated, Core Web Vitals green, internal linking tidy. None of it mattered, because the text itself was a paraphrase of the existing top ten.

This is worth stating bluntly because the industry keeps misdiagnosing it. Sites are not being demoted for using AI. They are being demoted for publishing commodity summaries at industrial volume, which AI merely made cheap. The same demotion hit human-written content farms a decade earlier under Panda. The tool changed; the failure mode did not.

The honest answer: yes, under three conditions

So does AI content rank in Google? Yes, and we can be precise about when, because the policy documents, the rater guidelines, and the crawl data converge on the same three conditions.

Fail all three and the answer flips to no, regardless of how good the prose sounds. Commodity AI text, the kind produced from a keyword and nothing else, is the exact substance the Lowest rating was rewritten to catch. There is no prompt that adds information the prompter never supplied.

What accountability looks like in markup

Condition three has a technical expression. Google connects authors to content through structured data, and a page that claims a real accountable author should say so in machine-readable form. Here is the minimal honest pattern, an Article with a named Person whose identity persists across the site:

{
  "@context": "https://schema.org",
  "@type": "Article",
  "headline": "Your article title",
  "datePublished": "2026-07-19",
  "author": {
    "@type": "Person",
    "@id": "https://example.com/#jane-doe",
    "name": "Jane Doe",
    "jobTitle": "Owner, Example Plumbing",
    "knowsAbout": ["residential plumbing", "backflow prevention"]
  },
  "publisher": {
    "@type": "Organization",
    "name": "Example Plumbing",
    "url": "https://example.com"
  }
}

The markup is five minutes of work. The commitment behind it is the point: a person willing to attach their name and job title to the claims on the page. Sites that publish AI text under no name are declaring, in the most machine-readable way possible, that nobody checked it.

How our own engine stays on the right side

Formative Digital publishes AI-assisted content at meaningful volume, so this question is not academic for us. We are exactly the kind of operation the scaled content abuse policy was written to police, and we treat that as a design constraint rather than an insult. Three controls keep our output on the rankable side of the line, and we describe them publicly because a standard you will not publish is a standard you probably do not follow.

Research before drafting. No page begins until live search research establishes what currently ranks, what it says, and what it fails to say. The gap becomes the brief. If the research finds no gap worth filling, the page does not get built; a growing share of our content planning runs end in a decision to skip, because adding a 41st restatement of a settled topic is precisely the low-value pattern Google demotes.

Source grounding. Claims trace to dated, checkable documents: Google's own policy pages, primary research papers, government statistics, or our clients' first-party analytics. Drafting from a model's memory of a topic produces confident paraphrase; drafting from sources produces pages a reader can verify. The difference is auditable line by line.

Corpus-level uniqueness checks. Single pages can each pass a quality gate while the collection quietly converges on shared structure and repeated phrasing, which is the exact fingerprint scaled-abuse detection reads. So we check at the corpus level, comparing new pages against everything already published and rewriting any page that overlaps its siblings. This is the discipline behind Vector 9 in our methodology, Cluster: topical depth only counts when every page in the cluster earns its slot. Our full production standard is documented in our guide to using AI content for SEO without triggering the patterns described above.

Why position one is a different contest

The Semrush finding deserves its own section, because it explains a frustration we hear from businesses that adopted AI writing early: page one is reachable, position one is not. The number one result for a competitive query is usually the page other pages cite, and citation flows to original information. A purely generated page is by construction a recombination of what its training data and its prompt contained, so it can match the field but cannot lead it.

The practical fix is not more generation, it is injection: feed the drafting process material that exists nowhere else. Your service call records, your pricing history, your before-and-after photos, your 20 years of local pattern recognition. In Matt's audits of Ontario service businesses this decade of accumulated operational knowledge is almost always sitting unused in the owner's head while their website recites generic advice. The businesses that win position one are the ones that get it onto the page. Building that depth deliberately, topic by topic, is covered in our guide to building topical authority.

The Canadian small-business angle

For the Ontario businesses we work with, the competitive math is favourable right now. Statistics Canada's 2024 survey data on digital technology adoption shows a minority of Canadian small businesses producing content in any systematic way, and among those that do, commodity AI text dominates because it is what cheap tools produce by default. That leaves the top of most local and regional queries lightly defended. A Brantford contractor or a Hamilton clinic that publishes twelve genuinely researched pages a year, each carrying first-hand detail and a named owner behind it, is competing against near-silence on one side and interchangeable slop on the other.

The window will not stay open. As detection of low-effort patterns improves and more businesses learn the difference, the floor rises. The durable strategy is the one Google has been describing since 2023: publish what only you can publish, and use AI to do it faster rather than to avoid doing it at all.

The verdict, stated plainly

AI content ranks in Google when it stops being AI content in any meaningful sense, when the model is the drafting layer for research, evidence, and experience that a human supplied and a human signs. It fails when the model is the whole supply chain. Google has now said this in a blog post (2023), a spam policy (2024), and its rater guidelines (2025), and the crawl data agrees with all three. The question that decides your rankings was never "did an AI write this." It has always been "did anyone here do the work," and that question has an answer your competitors can read on the page. We publish more findings like this in our research library.