Does AI-generated content rank worse? The data says it's effort, not origin.

What Google actually says

Google's own developer documentation, last updated 1 October 2026, does not say AI-written content ranks worse. It says the opposite kind of thing matters: content built with generative AI should still meet Search Essentials and the spam policies, with a direct warning that "using generative AI tools or other similar tools to generate many pages without adding value for users may violate" the scaled content abuse policy. The instruction to creators is to "focus on accuracy, quality, and relevance, especially when automatically generating the content," and to "manually factcheck and review all AI-generated content for accuracy and trustworthiness before publishing," including the title, meta description, structured data and alt text, since all of those can surface in Search on their own. The page points to Quality Rater Guidelines sections 4.6.5, scaled content abuse, and 4.6.6, main content built with "little to no effort, little to no originality, and little to no added value," and notes plainly that rater guidelines do not directly set rankings. Google's policy is about mass production of low-value pages, not about the tool used to write them.

The scoreboard: 331,000 pages, and AI content is still winning position one

Ahrefs tested that policy against live results. Ryan Law's study, published 27 July 2026 and updated 14 September 2026, drew on 331,000 pages across three separate checks: a sample of the top 10 results for 100,000 SERPs in June 2026, an indexation check, and a longer performance window tracked through Search Console. Positions 1 through 3 were not clean of AI content. 5.3% were entirely AI-generated and 9% were at least 80% AI, while pages under 50% AI content still made up 82.2% of the top three. Average AI share barely moved by rank, 27.1% at position 1 against 30.9% at position 10. If Google were filtering AI text out of the top of the page, that line would not be nearly flat.

Where the gap shows up is underneath the rankings, not inside them. Indexation fell from 49.28% for pages Ahrefs' detector scored low-AI to 40.35% for pages scored very-high-AI, and impressions for the low and moderate buckets ran 2 to 3 times higher than for the high and very-high buckets, a gap that held steady rather than widening across the two periods Law tracked. Roughly 40% of the heaviest AI pages were still indexed and still showing up, just less often. Law's own read is that AI use correlates with a quality gap rather than causing a penalty on its own, since newer or lower-authority sites tend to lean on more AI and the sample skews toward pages Google already ranks at all.

The same pattern, now with citations: ChatGPT and Perplexity

Graphite's Five Percent research team ran the equivalent test on answer engines. The study, led by Jose Luis Paredes with Gregory Druck and Ethan Smith, published and last modified 14 October 2025, scored 31,493 keywords across 10 categories. In Google's own results, 86% of articles came back human-written and 14% AI-generated by their detector, but only 7% of number-one results were AI, half of what you would expect if AI made no difference at all. A paired-keyword comparison found human pages ranked statistically significantly higher, a Wilcoxon signed-rank result with a p-value under 1 in a million.

Citations told a looser story. Articles cited by ChatGPT came back 82% human and 18% AI. Perplexity landed on the identical 82/18 split. That 18% is higher, not lower, than AI's 14% share of what ranks in Google, which argues against either model filtering out AI-assisted sources. Nothing here shows an AI detector sitting in the citation pipeline. It shows roughly the same quality sorting Ahrefs found in classic search, playing out again once the retrieval moves from a results page to a generated answer.

Why the numbers move without a stated penalty

Put Google's own policy next to both studies and the shape holds together. Nobody involved, not Google, not Ahrefs, not Graphite, claims the platforms run a line that reads "if AI-written, demote." What both research teams describe instead is a correlation: content produced with less editorial effort tends to use more AI, and that lower-effort content underperforms for reasons that would show up with or without a generative model involved. Law states this directly as his conclusion. Paredes, Druck and Smith go further and separate the two variables the public debate usually blends together, writing that heavily human-edited AI-assisted content was not evaluated in their study and "may be more effective in search and answer engines," a claim they are careful to flag as untested rather than proven. The variable worth managing is editorial effort. AI is one of several ways that variable gets skipped, not the cause of the drop on its own.

What we could not verify

Both detectors are probabilistic, not exact. Ahrefs' tool only scores pages with at least 350 words, and its own methodology notes it may not match whatever Google uses internally, if Google scores this at all. Graphite's classifier carries a stated 4.2% false positive rate and a 0.6% false negative rate, and the 2024-to-2025 comparison in that study used two different detector versions, so the rise from 12% to 14% AI share year over year is not a clean trend line. Neither study isolates cause from correlation. Both samples skew toward content that already ranks or gets cited, and neither team can rule out that lower-authority, newer sites simply use more AI for unrelated reasons. Nobody has published a study that scores heavily-edited AI-assisted drafts as their own category rather than lumping them in with raw, unedited output. That is the comparison a working content team actually needs, and it does not exist yet.

A checklist before you publish anything AI-assisted

1. Write for the policy Google actually enforces. The rule is scaled, low-value production, not point of origin. A single well-researched, fact-checked piece is not what either the spam policy or the quality raters are aimed at.
2. Fact-check and edit every draft before it ships, title, meta description, schema and alt text included, exactly as Google's documentation instructs, not just the body paragraphs.
3. Do not mass-produce near-duplicate pages from a model with no added research or editing. That is the specific behavior named in Google's scaled content abuse policy, and it is also the pattern both independent studies associate with weaker indexation and impression numbers.
4. If you run a mixed content program, track indexation and impressions by draft type. Both studies found a real, measurable gap at scale even without a formal penalty. Your own pages are the only dataset that tells you whether that gap exists in your work specifically. See is AI search traffic worth it for how to read that data once you have it.

None of this changes what actually earns a citation once a page is live. That part of the job is still about where AI engines quote from on the page, and about keeping a publishing pace your editorial process can actually fact-check at, rather than one a model alone could sustain.

Common questions

Sources: Google Search Central, Google Search's guidance on using generative AI content on your website, last updated 1 October 2026. Ryan Law, Google Doesn't Punish AI Content, It Punishes Bad Content, Ahrefs, published 27 July 2026, updated 14 September 2026. Jose Luis Paredes, Gregory Druck and Ethan Smith, How Does AI-Generated Content Perform in Search and Answer Engines?, Graphite (Five Percent), published and updated 14 October 2025.