Ahrefs Finds Heavy AI Use Correlates With Weaker Google Performance
Ahrefs published a July 27 study of Google search results finding that pages with more detector-flagged AI text generally ranked lower, were indexed less often and received fewer impressions. The pattern was gradual rather than absolute: 5.3% of pages in positions one through three were scored as fully AI-generated, and heavily flagged pages appeared throughout the top 10.
Ahrefs published a new analysis on July 27 examining how pages with different levels of detector-flagged AI text performed in Google Search. The company found a gradual relationship between heavier AI signals and weaker ranking, indexation and impression outcomes, but no evidence of a hard rule that blocks AI-heavy pages.
What the study measured
For the ranking analysis, Ahrefs drew one million pages from the top 10 results for 100,000 searches conducted in June 2026. About 300,000 pages were present in its crawler database, and roughly 150,000 contained at least 350 words, the minimum used by its AI-content detector.
Among pages ranking in positions one through three, 5.3% received a 100% AI score and 9% scored at least 80% AI. Pages below 50% AI accounted for 82.2% of those top-three results. Across the full first page, the share of pages in the very-high bucket rose from 8.4% at position one to 11.7% at position 10.
The study also sampled pages to examine indexation. Ahrefs counted a page as indexed when it had a detected organic ranking, a Search Console impression since January 2026, or an exact-URL result in a site search. By that measure, 49.28% of low-AI pages were indexed, compared with 40.35% of very-high-AI pages.
Correlation is not a Google penalty
The findings do not show that Google detects and penalizes AI-written text. Ahrefs used its own probabilistic detector, acknowledged that detectors are imperfect and said its system is not necessarily comparable to anything Google may use. The data also cannot isolate AI use from other differences among sites and pages, such as authority, originality, usefulness or editorial quality.
Search Engine Journal independently reviewed the analysis and emphasized the same boundary: AI-heavy pages appeared at every top-10 position, while higher detector scores were associated with slightly weaker outcomes on average.
Google's published guidance focuses on accuracy, quality, relevance and added value. It warns that generating many pages without value may violate the scaled-content-abuse policy, but it does not prohibit generative AI as a production tool.
Why it matters
For data and content teams, the study argues against treating an AI-detector score as a direct ranking signal or a reliable verdict on an individual page. The more defensible operational lesson is to measure the attributes that can be improved: factual accuracy, original evidence, expert judgment, usefulness and sustained search performance. Ahrefs' results are consistent with those qualities mattering more than whether AI assisted the writing process.
Key Points
- 1Ahrefs analyzed roughly 150,000 sufficiently long pages drawn from the top 10 results for 100,000 Google searches conducted in June 2026.
- 2Pages with less than 50% detector-flagged AI text accounted for 82.2% of top-three rankings, while 5.3% of top-three pages scored as fully AI-generated.
- 3Indexation was 49.28% for the low-AI bucket and 40.35% for the very-high-AI bucket under Ahrefs' three-signal indexation method.
- 4The study shows correlations, not proof that Google uses AI detection as a ranking penalty; Ahrefs says its detector is probabilistic and differs from any system Google may use.
Scoring Rationale
A large first-party search dataset has direct implications for teams using AI in publishing, but the vendor's proprietary detector and observational design limit causal conclusions about Google's ranking systems.
Sources
Primary source and supporting public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
