The practical brief
The finding
Across 7,562 AI Overviews, the mean share of cited domains absent from the accompanying first page was 29.8%; those off-page references were not lower quality under the study’s measures. [c1] [c4] [c7] [c12] [c19] [c20]
Why it matters
Ranking reports alone miss part of AI citation visibility, including sources with higher measured domain credibility.
| Measure | Observed result | Qualitative interpretation |
|---|---|---|
| Citation overlap | 29.8% mean share of cited domains absent from the corresponding first page across 7,562 AI Overviews | Rankings miss some citation visibility; off-page references were not lower quality among credibility-scored references. [c4] [c19] [c22] |
| Source agreement | 11.03% of 98,020 claims from 7,491 verifiable AI Overviews classified as inconsistent with retrieved citation text | A pipeline-derived source-agreement estimate, not an independently established factual-error rate. [c5] [c10] [c15] [c16] |
| Publisher advertising | Visible ads on 30,994 of 61,212 cited reference URLs: 50.63%, an observed lower bound | Potential exposure for ad-supported publishers, not measured traffic or revenue loss. [c7] |
What to try
BLURSOR’s practical interpretation
Our interpretation: track citations separately from rankings, check how answers represent your content, and assess business outcomes with your own analytics.
Study boundary
This U.S.-localized trending-query audit did not test optimization tactics or measure traffic, conversions or revenue effects.
Rankings alone leave a visibility blind spot
If you use first-page rankings to judge whether customers can discover your business in Google’s AI answers, your report is incomplete. Across 7,562 AI Overviews—the generated summaries shown in search—the mean share of cited domains absent from the accompanying first page was 29.8%. This is domain overlap, not the percentage of individual pages lacking an organic ranking.
The missing sources were not simply lower-quality alternatives. Among references covered by the study’s domain-credibility score, off-page references scored higher on average than on-page references. They also had a smaller share from the study’s defined user-generated-content platforms. The credibility comparison excludes unscored domains, so it cannot establish the quality of every citation.
What the researchers actually measured
Researchers collected 55,393 query executions across 19 trending-topic categories from March 13 to April 21, 2026. A Chrome crawler in Northern Virginia captured Google results, Overviews and cited sources. Overviews appeared in 7,583 executions, or 13.7%.
Question-form queries triggered Overviews in 64.7% of executions, versus 9.5% for non-questions. That association does not show that question-shaped website copy wins citations. Sports supplied 51.4% of executions, and queries could recur.
Giving each unique query equal total weight left the substantive conclusions unchanged: activation rose, while fidelity and cited-page advertising prevalence changed only slightly. Repetition therefore does not explain away the principal findings.
Credibility and faithful representation are different
Cited domains scored higher than the unweighted full first page, but that advantage reversed when comparisons focused on top-ranked results. A separate graph-based reliability measure did not reproduce a statistically significant advantage. Domain credibility is not answer accuracy.
An automated verifier classified 11.03% of 98,020 claims as inconsistent with retrieved citation text. This measures fidelity—agreement with sources—not independent truth. The authors describe the narrower 9.64% omitted-or-contradicted subset as a conservative ceiling on substantive unfaithfulness, not a precise error estimate: real-time answers may use structured feeds, while later-captured pages show different values.
The pipeline also missed facts. Extraction validation found 83.23% recall across 100 sampled Overviews; verifier validation audited only 100 verdicts. Partial paywalls and excluded social content can create apparent failures.
Separate a citation from a business outcome
Visible ads appeared on 30,994 of 61,212 cited reference URLs, or 50.63%. That lower-bound observation identifies exposure for ad-supported publishers, not measured lost visits or revenue.
Our interpretation is to keep three reporting questions separate:
- Discoverability: is your domain cited independently of its ordinary ranking?
- Representation: does the answer preserve the cited page’s claims and qualifications?
- Business outcome: do your own analytics show useful visits or conversions?
Treat placements as snapshots, not guarantees
The audit does not represent all searches, countries or personalized experiences. Citation findings also concern eligible content: noindex and nosnippet controls could restrict Overview use while affecting traditional Search.
When adjacent cross-day observations of the same query both produced an Overview, exact cited-URL-set agreement was only 0.24%. Repeated monitoring is therefore a reasonable interpretation, not a proven growth tactic. The study cannot tell you which edits increase citations or whether a citation changes purchases.
The source and evidence
Measuring Google AI Overviews: Activation, Source Quality, Claim Fidelity, and Publisher Impact
Study limitations and disclosures
- This single-vantage-point, 40-day audit sampled U.S.-localized trending queries, not all Google traffic or personalized experiences. Topic mix and repetition affect aggregate rates, but giving each unique query equal total weight did not change the substantive conclusions.
- Citation distributions concern eligible content. Publishers using noindex or nosnippet controls could restrict AI Overview use while also affecting traditional Search, making them absent or underrepresented.
- Credibility scores assess domains, not individual claims. Unscored domains were excluded, coverage differed between source pools, and conclusions depended on ranking weights and the credibility measure. Off-page references scored higher than on-page references only among score-covered references; this does not establish the quality of every off-page citation.
- Fidelity checks compare answers with accessible citation text, not independently established truth. Partial paywalls, excluded social content and later page capture can create apparent inconsistencies. Real-time values may come from structured feeds; the authors therefore call the 9.64% omitted-or-contradicted subset a conservative ceiling on substantive unfaithfulness, distinct from the broader 11.03% inconsistency rate. Extraction validation found 273 missing facts and 83.23% recall across 100 Overviews. Verifier validation covered 100 stratified verdicts; possible sampling or evaluator bias remains.
- Visible-ad prevalence is a lower bound: 14.2% of references led to uncrawled social/video platforms treated as having no ads. The audit does not measure click displacement, revenue loss, conversions, user beliefs or correction-seeking.
- Haofei Xu, Umar Iqbal and Jacob M. Montgomery are affiliated with Washington University in St. Louis. The work received partial U.S. National Science Foundation support under grant CNS-2543517. The verified material does not establish a commercial-interest disclosure; academic affiliation should not be read as evidence that no commercial interests exist.
Evidence behind this briefing
[c1] Measurements used a Puppeteer/Chrome crawler deployed in AWS Northern Virginia. Findings are explicitly scoped to the U.S.-localized experience, not global Google Search behavior.
Section 3.1, Web Crawler · Read the study
[c2] Among 55,393 trending queries across 19 categories during March 13–April 21, 2026, 7,583 returned an AIO (13.7%). Question-form queries activated AIOs at 64.7% (2,716/4,198), versus 9.5% (4,867/51,195) for non-questions, a 6.8-fold association rather than a causal test. Repeated queries and the topic mix affect the aggregate: deduplication increased activation from 13.69% to 16.03%, and Sports supplied 51.4% of queries.
Section 4.1, Query phrasing and length; Tables 3–4; query repetition paragraph · Read the study
[c3] AIO citations had higher average domain credibility in the unweighted full-page comparison, but this advantage was not consistent across ranking specifications: it reversed for the first result, first three results, first five results and strongly top-weighted results. PC1 measures domain credibility, not factual accuracy of individual answers. Comparisons excluded unscored URLs; coverage was 60.4% for AIO references and 48.6% for first-page URLs.
Section 4.2, Source quality by search result ranking; Tables 6–7 · Read the study
[c4] Across 7,562 AIOs in the overlap analysis, the mean share of cited domains absent from the corresponding full first page was 29.8%. This is domain-level overlap, not an estimate that 29.8% of individual cited pages lacked an organic ranking. It supports treating AIO citation visibility as distinct from first-page visibility, without establishing the internal selection mechanism.
Section 4.2, Reference overlap with first-page results · Read the study
[c5] The automated verifier classified 11.03% of 98,020 claims from 7,491 verifiable AIOs as inconsistent with retrieved cited-source text: 6.98% omitted, 2.66% contradicted and 1.39% involving conflicting sources. This measures fidelity to accessible citations, not independent real-world truth; vague support counts as consistent. The analysis excluded 80 AIOs with no extracted claims and 12 with only social-media references.
Section 4.3, Overall fidelity; Table 2 · Read the study
[c6] The content pipeline excluded listed social-media platforms and retained only visible paywalled text. Consequently, some unsupported labels may reflect unavailable source content rather than AIO failures; 262 pages had paywalls, and text retrieval failed for fewer than 1% of attempted pages.
Section 3.2.2, Webpage Content · Read the study
[c7] Across the 40-day window, 30,994 of 61,212 AIO-cited reference URLs (50.63%) displayed visible ads. This is an observed lower bound: 14.2% of references pointed to uncrawled social/video platforms treated as having no ads. The study measured monetization characteristics, not traffic loss or publisher revenue effects.
Section 4.4, Ad prevalence on cited pages; Table 9 · Read the study
[c8] Among 7,583 AIO-bearing search-result pages, 164 (2.16%) also displayed a Google sponsored search ad, and 39 (0.51%) placed one above the AIO. None appeared inside the AIO container. These are layout/co-occurrence observations, not measurements of preserved revenue or causal click displacement; the informational/trending sample underweights commercial-intent searches.
Section 4.4, Sponsored search ads on AIO-bearing pages · Read the study
[c9] The measured credibility advantage of AIO citations over co-displayed first-page results depended on the credibility metric. Query-paired shared samples supported an advantage under MBFC factuality (7,281 queries; gap +0.106, 95% CI +0.103 to +0.109; p<0.001) and Lasser mean (6,834 queries; +0.013, 95% CI +0.011 to +0.015; p<0.001), but not graph-based Burdisso p-reliability (7,174 queries; -0.007, 95% CI -0.017 to +0.002; p=0.130). Scores use their respective original scales; credibility ratings are not claim-accuracy measurements, and the rating-based checks are not fully independent of PC1.
Appendix B, Sensitivity to Alternative Credibility Measures; Table 10 · Read the study
[c10] Fidelity verification can misclassify real-time claims because cited pages are crawled after AIO generation and structured weather feeds may not match captured page text. The authors therefore describe the combined 9.64% Omitted/Incorrect rate as a conservative ceiling, not a precise substantive-unfaithfulness estimate. Keyword-based adjustments raised Technology to 89.41%, Shopping to 85.90%, and Jobs & Education to 87.71%, but can over- or under-identify real-time claims.
Section 4.3, The Climate and Jobs anomalies; Appendix F, Webpage Changes · Read the study
[c11] The dataset contains 55,393 executions of 32,608 unique normalized queries, with 7,965 queries recurring across dates and accounting for 49.89% of executions. Giving each unique query total weight one changed activation from 13.69% to 16.03%, fidelity from 88.97% to 89.11%, and cited-reference ad prevalence from 50.63% to 52.39%. Among 17,867 cross-day adjacent observation pairs, activation agreed 95.92%; conditional on both triggering AIOs, exact text and citation-set agreement were only 0.06% and 0.24%, respectively. Repeated queries therefore capture changing answers rather than simply identical duplicates.
Appendix E, Identification of repeated queries; Query-balanced robustness analysis; Table 11; Variation across repeated observations · Read the study
[c12] Sampling queries receiving elevated or growing attention limits generalization to the full distribution of Google Search traffic, especially less time-sensitive and long-tail queries. The study does not measure user beliefs, correction-seeking, heterogeneous consequences, or variation across user profiles and personalization states.
Section 5, future-work limitations following Transparency & Accountability · Read the study
[c13] All three authors—Haofei Xu, Umar Iqbal, and Jacob M. Montgomery—are affiliated with Washington University in St. Louis, St. Louis, Missouri, USA.
Author block preceding the abstract · Read the study
[c14] Citation-distribution findings are conditional on eligible content. During the observation window, publishers could restrict AIO use with noindex and nosnippet directives, but these general search controls also affected traditional Search; publishers using them may be absent or underrepresented.
Section 4.2, Source Selection, opening paragraph · Read the study
[c15] Human validation of claim extraction across 100 randomly sampled AIOs found an incomplete claim inventory: 273 missing atomic facts and 83.23% recall. Among 1,379 sampled extracted claims, 24 were wrong; 62 of the 100 AIO samples contained a missing fact.
Section 3.3.3, Pipeline Validation, Claim extraction · Read the study
[c16] Verifier validation covered a stratified sample of only 100 claim-level verdicts, with 20 verdicts from each of five labels. Two annotators independently re-labelled verdicts against the same matched references, rather than exhaustively adjudicating the study's claim inventory.
Section 3.3.3, Pipeline Validation, Claim verification · Read the study
[c17] The authors acknowledge that high inter-annotator agreement and adjudication do not eliminate possible sampling or evaluator bias in verifier validation.
Section 3.3.3, Pipeline Validation, Claim verification, closing sentence · Read the study
[c18] The authors disclose partial NSF support under grant CNS-2543517.
Acknowledgements · Read the study
[c19] Off-page AIO references were not lower quality under the study's domain-level PC1 measure: among scored references, mean credibility was 0.757 for 9,218 off-page references versus 0.724 for 27,677 on-page references (difference +0.033, p<0.001). This comparison applies only to PC1-covered references, not all cited sources.
Section 4.2, Source Selection — Reference overlap with first-page results · Read the study
[c20] Across all AIO references in the overlap comparison, defined UGC platforms accounted for 3.39% of off-page references (590/17,399), versus 18.52% of on-page references (8,081/43,641), a difference of −15.13 percentage points (p<0.001). Unlike the PC1 comparison, this analysis was not restricted to scored references.
Section 4.2, Source Selection — Reference overlap with first-page results · Read the study
[c21] The study defines UGC using a specified set of platforms without strong editorial moderation; Wikipedia is excluded because it has community-based editorial policies and norms.
Section 4.2, Source Selection — UGC prevalence · Read the study
[c22] PC1 coverage was incomplete and differed across source pools. The study does not impute credibility scores to unmatched domains, so credibility comparisons are conditional on scored-domain coverage.
Section 4.2, Source Selection — AIO vs. first-page source quality · Read the study