Quick answer

Review AI content ranking studies from Ahrefs and Semrush. Understand detector labels, sample limits and Google policy before planning publication.

AI-generated content can appear in Google rankings. That does not tell you whether an unedited draft will rank, how much human editing helps, or what publishing volume is safe. Google's guidance permits appropriate AI use, while published ranking studies describe particular samples of pages. Those are different kinds of evidence.

This review separates Google's policy from observations reported by Ahrefs and Semrush. Use it to evaluate a research claim before turning that claim into a content budget or publishing target.

What does Google say about AI content?

Google does not prohibit appropriate use of AI or automation. Its AI-content guidance distinguishes helpful publishing from generating content primarily to manipulate rankings. Permission to use a tool is not a promise that the resulting finished page will perform.

The current generative-AI guidance calls for manual accuracy and trustworthiness checks, including metadata, structured data and image alternative text. The scaled-content-abuse policy concerns large amounts of unoriginal, low-value content made primarily to manipulate search, regardless of how it was created.

These documents provide no monthly article quota that makes AI publishing safe. A publishing schedule should follow your review capacity and the usefulness of each proposed page. Reducing volume does not repair a misleading article; increasing volume does not establish abuse by itself.

AI detectors measure text patterns, not the writing process

A detector assigns a label or probability to the finished text. It does not observe who researched a page, whether an editor checked its claims, or whether a real practitioner supplied the examples. An article written from an AI outline may leave no detectable AI wording. A human-written passage may receive an AI label.

That distinction affects the language used to report research. “Classified as mixed by this detector” describes a measurement. “Written and reviewed by humans using AI” describes a production workflow. You need different evidence for each statement.

Keep the detector and its categories attached to any percentage you cite. A proportion from one classifier should not become an estimate of verified authorship across the whole web.

What the Ahrefs ranking sample shows

The Ahrefs report published in July 2025 used 100,000 random keywords, extracted the top 20 URLs and analyzed 600,000 pages for which it had content. Its own detector classified the text into pure-AI, pure-human and mixed categories.

The chart below shows that classification split. It is a composition chart, not a comparison of ranking improvements. Ahrefs cautions that detectors can produce false positives and reports no clear relationship between estimated AI content and ranking position in its sample.

Ahrefs detector labels in its reported ranking-page sample Shares of 600,000 pages with available content sampled from top-20 results: pure AI 4.6 percent, pure human 13.5 percent, mixed 81.9 percent. These are detector classifications, not verified authorship. Horizontal axis: percent of sampled pages, from zero to one hundred. 0 25 50 75 100 Pure AI4.6% Pure human13.5% Mixed81.9% Share of sampled pages (%)
Ahrefs detector classifications among 600,000 sampled ranking pages with available content. Source: July 2025 report, checked October 12, 2026. Exact collection dates are not stated in the opened methods. Classification is not verified authorship.
Accessible chart data: reported shares of the Ahrefs sample
Detector categoryShare of sampled pages
Pure AI4.6%
Pure human13.5%
Mixed81.9%

The large mixed category makes a simple “AI versus human” headline incomplete. These labels do not tell you which pages were accurate, reviewed by specialists or useful to readers. Nor does observing classified AI text among ranked pages establish its odds of succeeding when newly published.

Why the Semrush findings need their own context

Semrush published ranking analyses in December 2024 and April 2026. Their samples and presentation differ. Treating both reports as one experiment would hide those differences.

Compare the population and measurement before comparing conclusions
ReportRanking sample and detectorDate and scope limits
Ahrefs, July 2025Available page content from top-20 URLs; Ahrefs detector.Exact collection dates, language and geography are not established in the opened methods.
Semrush, December 202420,000 blog articles from top-20 results for 20,000 random keywords; GPTZero.The article mentions an October analysis; exact dates, language and geography remain unestablished.
Semrush, April 202642,000 pages filtered from top-10 URLs using the /blog/ path; GPTZero.Collected November 2025. The URL filter excludes articles published under other path patterns.

The later report also surveys 224 SEO professionals. That survey records respondents' views; it is not a measurement of the ranking pages' actual writing process. The ranking analysis and the survey cannot independently validate one another's authorship labels.

Publication year is another important distinction. A report released in 2026 may describe results collected in 2025. Calling it a snapshot of current rankings would change the claim.

The opened 2026 report gives differing position-one percentages in its introductory summary and detailed section. This review therefore does not reproduce those figures as a precise benchmark. The methodology remains useful for understanding the sample; the conflicting figures need clarification from the publisher.

A ranked-page sample cannot predict your chance of ranking

Studies of pages already in search results answer a conditional question: what do the selected results look like? They do not automatically answer how often all newly published AI or human articles reach those results.

Consider a report describing which detector categories appear at the first position. Its denominator is the selected first-position results. To estimate the probability that a new article reaches first position, you would instead need a defined population of published articles, including failures, and a consistent observation period. Those denominators are not interchangeable.

The same caution applies when reading a near-equal top-10 comparison. Similar proportions within an already-ranked sample do not prove that changing the authoring method leaves a page's ranking unchanged. The original pages may differ in topic, website reputation, age, internal links or the quality of the answer.

This is a reason to avoid precise forecasts from broad research headlines. A useful study can establish that AI-classified text appears in rankings while leaving the effect of your editorial process unresolved.

Do the studies prove a particular editing benefit?

The reviewed methods do not establish a universal percentage improvement from human editing, because they do not measure a defined editing intervention. They classify finished pages or collect opinions, rather than assigning comparable drafts to editorial treatments and tracking the resulting change in rankings.

A claim such as “editing closes the ranking gap” needs evidence that identifies the treatment. Grammar corrections, source verification, a new example and a complete structural rewrite are different interventions. Combining them under “human edited” makes it hard to learn which change mattered.

There is still a practical reason to edit: readers need reliable answers. Verify factual claims, remove invented experience, clarify ambiguous instructions and check whether the page addresses the intended question. Those are editorial responsibilities even when a ranking improvement has not been measured.

For the verification task, follow the claim-by-claim AI-content fact-checking process. For choosing tools and organizing the review, use the separate fact-checking workflow and tool guide.

Evaluate an AI-assisted workflow on your own site

If you want to compare production methods, record the work and outcomes consistently. A local experiment can inform your team's decisions, although it will not establish a universal rule about Google's algorithm.

  1. Define the decision. Decide whether you are comparing drafting time, editing effort, factual accuracy or search performance. Use the same outcome definition for both groups.
  2. Record the treatment. Document which tools were used, the review tasks completed and material changes made before publication. A detector score cannot replace these records.
  3. Keep the comparison interpretable. Record differences in topics, search intent, publication dates and website context. Avoid presenting unmatched groups as a controlled experiment.
  4. Separate production and search measures. Editing hours and publication dates describe operations. Search Console impressions, clicks and query-level positions describe search visibility. Enquiries require their own reliable tracking.
  5. Report limits alongside results. Show the observation period, missing data and other changes that could explain an outcome. A traffic increase after publication is not proof that AI caused it.

Set the review standard before comparing speed. If one method produces drafts that require unresolved factual corrections, a lower drafting time does not represent the cost of a finished article. Keep operational savings separate from an untested prediction of rankings or revenue.

Scope and method of this review

This is an AI-assisted editorial review of public research and documentation, checked October 12, 2026. It is not an original theStacc ranking experiment, a replication of the providers' raw datasets or independent verification of the actual authorship of their sampled pages.

The source pages were opened and their stated methods compared. Source publication dates are identified where relevant; missing collection or geographic details remain unknown. The chart reproduces compatible published category shares without combining samples. Conflicting source figures are excluded rather than averaged or silently corrected.

Use these findings to ask better questions of a content workflow: what was measured, what was reviewed, and what uncertainty remains? A defensible publishing decision needs those answers more than it needs a universal “AI wins” or “humans win” statistic.

Sources & references

From the theStacc product Explore the Blog & SEO module →

Plan topics, review article drafts, and publish to a connected CMS.

Weekly local SEO teardowns

One practical email a week. Map Pack, GBP, AI Overviews. No fluff. Unsubscribe anytime.