AI-Generated Content and Google Rankings: The Evidence

Google's written position on AI-generated content has not changed in substance since February 8, 2023: content is judged on quality, however it was produced, and using automation to generate content primarily to manipulate rankings is spam. What has changed is the enforcement text around it, the scaled content abuse policy of March 2024, and the guidance page, which Google updated on October 1, 2026. Meanwhile five published studies, read on November 5, 2026, have measured how much AI content actually ranks. They agree that it ranks and that fully AI-generated pages are rare at the top. They disagree on almost every number, because each used a different detector.

This page is a reader's guide, not a study. It quotes Google's pages with their dates, then sets out each study with its sample size, collection date, detector and headline finding, then says which conclusions survive across all of them. Where two studies cannot be compared, it says so rather than averaging them.

The foundational text is a Search Central blog post dated February 8, 2023, posted by Danny Sullivan and Chris Nelson on behalf of the Google Search Quality team. Its argument runs in three steps. Google's ranking systems aim to reward original, high-quality content, and that focus on quality rather than on how content is produced has been core since the company began. Using automation, including AI, to generate content with the primary purpose of manipulating ranking is a violation of the spam policies. And not all use of automation, including AI generation, is spam.

The post then points creators at the qualities Google calls E-E-A-T and at a "Who, How and Why" framework added to the helpful-content page the same day. On disclosure it is measured rather than mandatory: AI or automation disclosures are useful where a reader might reasonably wonder how content was made, accurate author bylines are advised where readers would expect them, and giving AI itself a byline is described as probably not the best way to tell readers AI was involved. That framework is still on the helpful-content page today, which asks whether extensive automation is being used to produce content on many topics, and whether the use of automation is self-evident to visitors through disclosures or in other ways.

The failure case in reading this text is treating the first half without the second. The sentence quoted above is the one most often repeated, and it is true. The sentence that follows it in the same post, that automation used primarily to manipulate rankings is spam, is the one that became a named policy thirteen months later.

On March 5, 2024 Google announced a core update and three new spam policies. The Search Central post that day, by Chris Nelson, defines the one that matters here: scaled content abuse is when many pages are generated for the primary purpose of manipulating search rankings and not helping users, and the use of automation, including generative AI, is spam if the primary purpose is manipulating ranking in search results. Google's companion post the same day, by Elizabeth Tucker, said the combination of the update and the policies was expected to reduce low-quality, unoriginal content in results by 40%; a note added to that post on April 26, 2024 revised the observed figure to 45%. Both numbers are Google's own, measured by Google.

The spam policies page, as read on November 5, 2026, carries the current wording. Scaled content abuse "is when many pages are generated for the primary purpose of manipulating search rankings and not helping users", typically "creating large amounts of unoriginal content that provides little to no value to users, no matter how it's created". The first listed example is "Using generative AI tools or other similar tools to generate many pages without adding value for users". The other examples are scraping and transforming feeds, stitching content from other pages, creating multiple sites to hide the scaled nature of the content, and keyword-stuffed pages that make little sense to a reader. The page's closing instruction is that a site hosting such content should exclude it from Search.

Google's documentation page on using generative AI content was updated on October 1, 2026; the changelog entry says the update brought in information from the Search Quality Rater guidelines. The page as read on November 5 opens by saying generative AI can be particularly useful when researching a topic and to add structure to original content, then draws the line in the same place as the spam policy: using generative AI tools to generate many pages without adding value for users may violate the scaled content abuse policy.

Three instructions on the page are more specific than anything in the 2023 post. "It is critical to manually factcheck and review all AI-generated content for accuracy and trustworthiness before publishing." For ecommerce sites, it adds that Google Merchant Center requires AI-generated images to carry IPTC DigitalSourceType metadata, and AI-generated product data, such as title and description attributes, to be specified separately and labelled as AI-generated. On raters, the page is careful to say that the guidelines are not a guide to ranking first and that rater ratings do not directly influence ranking. On disclosure it keeps the 2023 tone: sharing information about how a piece of content was created can help give readers more context.

The studies below are the ones with a published page stating how the work was done. Each was read on November 5, 2026. Three are by commercial SEO tool vendors and two of those used the vendor's own detector, which is a conflict worth holding in mind rather than a reason to discard the data. Graphite, a content agency, used a third-party detector and is the most explicit about limitations. Every figure below is that study's own, and none is comparable to another's without checking the detector column.

Ahrefs' study, published July 27, 2026 under Ryan Law's byline, is the largest by pages sampled and the only one with three separate samples: ranking pages, an indexation sample and a traffic sample. Its ranking finding is that AI content level barely moves with position. The average AI level rises from 27.1% at position 1 to 30.9% at position 10, and the median from 17.1% to 19.5%. The study does not report a correlation coefficient; it shows the position table and says AI use seems to be correlated with lower performance, while noting there are many reasons AI content might earn fewer impressions.

The two other samples are where the study earns its title. On indexation, drawn from a million pages at one page per domain, the share of pages Google had indexed fell as AI level rose: 49.28% of low-AI pages were indexed against 40.35% of very-high-AI pages, which the study reads as a modest penalty at the gate rather than at the ranking stage, noting that even in the very-high bucket 40% of pages were indexed. On traffic, from Search Console data on live URLs in four buckets of roughly 13,000 to 29,000 pages, low and moderate AI-level content received two to three times the impressions of high or very-high AI content, with no sharp cliff between the two upper buckets.

The study's own limitations are the ones to carry forward: detection is not perfect, Ahrefs' method of detecting AI content differs from whatever Google does, the ranking sample only contains pages Google already chose to index, and the traffic window is short. Its author's closing line, that marketers should be afraid of spam and scaled content rather than of AI content creation, is a conclusion rather than a measurement, but it is the same line Google's policy draws.

Put the headline numbers side by side and they do not agree. Originality AI's tracker says 17.31% of top-20 results were AI in September 2025. Graphite says 14% of search results in June 2025. Semrush's December 2024 study says about 8% of URLs were likely AI. Ahrefs does not report a single share at all, preferring a distribution in which 54.7% of top pages score under 20% AI and 9.0% score 80% or more. These are four different detectors, four different thresholds for calling a page AI, four different keyword sets and three different years. Graphite says it plainly of its own two studies: the 2024 figure of 12% and the 2025 figure of 14% "are not directly comparable because they are computed using different AI detectors".

The top-position question is contested in a subtler way. Semrush's April 2026 study puts the probability that a position-1 blog page is human-written at 80.5% against 10% for AI-generated, while its own December 2024 study found near parity in the top 10, 57% against 58%. Graphite found 7% of number-one articles AI-generated. Ahrefs found 5.3% of positions 1–3 at a 100% AI score and an average AI level at position 1 of 27.1%. Those figures can all be true at once: they measure different things, a page classified as AI by a binary threshold versus a continuous AI level, on different page types. What they share is the direction: fully AI pages are a small minority of top results, and AI-assisted pages are common everywhere.

One more limitation applies to all five. A detector reads the finished text. Semrush's April 2026 write-up says so directly: the detector does not care about the process, it reads the finished product. A page drafted by a model and rewritten by an editor may score as human; a page written by a person in a flat style may score as AI. Graphite notes it did not evaluate AI-assisted content with heavy human editing at all. So none of these studies can answer the question most teams actually ask, which is whether a human-edited AI draft ranks differently from a human draft.

Separating the two is the point of the page. A claim holds if Google's text and every study that measured it point the same way. A claim is contested if it rests on one detector, one vendor or one year. The grid is the short version; the prose below it gives the evidence for each cell.

The practical reading is the one Google's own text supports and none of the studies contradicts. Use the tools; keep the editor; never publish at a scale the value cannot justify; fact-check before publishing; label images and product data as the guidance now requires; disclose where a reader would want to know. The routing below sorts the common situations. What happens to a site that gets this wrong during a core or spam update, and what to measure afterwards, is in our August 2026 spam update post.

If you want help building a content operation that uses models without tripping the scaled-content line, with the editorial gates and the labelling in place, it is part of what our content engine service and our agentic SEO work cover; the scoring we apply before anything ships is in our 12-point content quality rubric.