The Study: 55,393 Searches Over 40 Days
Researchers at Washington University in St. Louis ran 55,393 Google searches on trending topics over 40 days, from March 13 through April 21, 2026. The study, led by Jacob Montgomery (political science), Umar Iqbal (engineering), and Haofei Xu (computational and data sciences), will be presented at the October 2026 ACM Internet Measurement Conference.
Their goal was to answer basic questions that Google has not publicly addressed: When does an AI Overview appear? Which sources does it draw on? Are the summaries accurate? And what happens to the rest of the web when people no longer click through?
The results challenge several prevailing assumptions about AI Overviews.
AI Overviews Appear Less Than Expected
The study found AI Overviews appeared for 13.7% of all queries. This is far lower than the 97% figures reported by some industry analyses, though higher than Google's own characterization of the feature as selective.
The activation rate varied sharply based on how users worded their search. Nearly 65% of question-form queries (starting with "how," "what," "why," etc.) produced an overview, compared with just 9.5% of non-question searches. Longer queries also triggered more overviews: the rate rose from 9.9% for one-word searches to 38.7% for queries of six words or more.
Topic also mattered significantly:
| Topic Category | AI Overview Rate |
|---|---|
| Hobbies & leisure | 46.1% |
| Science | 39.9% |
| Political | 7.5% |
| Law & government | 9.6% |
This pattern suggests Google applies editorial discretion in deciding when to generate an AI summary, avoiding sensitive political and legal topics while embracing informational and leisure queries.
Citation Quality: Credible Sources but Gaps Remain
The researchers analyzed 61,212 citations across the overviews. The domains cited by AI Overviews received higher average credibility scores than the traditional first-page results shown for the same searches, and the overviews relied less on user-generated content.
However, 29.8% of the cited domains did not appear in the first-page results. This indicates that Google uses a source-selection process for its AI summaries that is distinct from its conventional search-ranking algorithm. An average overview cited about eight sources.
The most concerning finding involves accuracy. The team separated overviews into 98,020 verifiable claims and checked each one against the cited webpages. About 89% were clearly or broadly supported by the sources. The remaining 11% included:
- 7% of claims where support was not found in the cited text
- 4% of claims that directly contradicted the cited source
Only 41.9% of overviews were "fully grounded," meaning every verifiable claim was supported by the available cited text. The researchers noted this figure is likely conservative because some cited pages could not be fully captured.
As Montgomery put it: "Google's summaries generally draw on credible websites, which is good news. But quality sources are not enough. In a meaningful share of cases, we could not find support for a claim in the pages Google cited. Readers still need to check what they're told, even when the citations look trustworthy."
The Revenue Problem for Publishers
The study also examined financial stakes for publishers. When an AI Overview answers a question directly, publishers may lose the visit that generates advertising, subscription, or sales revenue. More than half of the cited webpages in the study displayed advertising.
The researchers found that 2.2% of search pages containing an AI Overview also displayed a Google-sponsored ad, including 39 pages where an ad appeared above the overview. That arrangement, they wrote, "may preserve Google's ad revenue while giving users less reason to visit the publishers whose content informed the answer."
The researchers did not directly measure traffic or revenue losses, but Montgomery warned: "Google's AI Overviews are built on content that publishers pay to produce, and many of those sites depend on advertising revenue. If people get their answer from the overview and never click through, those publishers lose the visit and the revenue."
What This Means for SEO and GEO Strategy
The WashU audit offers several practical insights for businesses optimizing for AI search:
Optimize for question-form queries. Since 65% of question-form queries trigger AI Overviews (versus 9.5% for keyword searches), structuring content to answer natural-language questions significantly increases the chance of AI Overview inclusion. FAQ sections, Q&A formats, and clear answer-first paragraphs matter more than ever.
Focus on informational and scientific topics. With 46.1% of hobby/leisure and 39.9% of science queries producing overviews, these categories offer the highest probability of AI search visibility. Commercial and legal queries are far less likely to trigger overviews.
Credible sources are necessary but not sufficient. The 11% unsupported claim rate means Google's AI can misrepresent even trusted sources. This means structured data, clear factual statements, and source-attributed claims in your content can help Google's model accurately represent your information. Avoid ambiguous phrasing that an AI summary could misinterpret.
Internal linking and authority signals matter differently now. Since 29.8% of cited domains were not in the first-page results, Google's AI source selection is not purely a ranking continuation. Building topical authority through depth and accuracy may matter more than traditional ranking signals for AI Overview inclusion.
Publisher revenue is an industry-wide challenge. The finding that Google places its own ads alongside AI Overviews while publishers lose traffic highlights the structural tension. With Google's new AI contribution pilot paying publishers 0.1% of ad revenue for AI-cited content, the economic model for content creation in the AI search era is being renegotiated in real time.
Want to improve your visibility in AI-powered search results? Contact Vivid for a GEO audit.
