Does the Citation Matter? Quality, Grounding, and Physician-Rubric Accuracy of Google's AI Overviews for Medical Information

Sam Agnoli; Brigid Reilly; Lily Barna; Emőke-Ágnes Horvát

cs.IROct 29, 2026

Does the Citation Matter? Quality, Grounding, and Physician-Rubric Accuracy of Google's AI Overviews for Medical Information

Authors: Sam Agnoli, Brigid Reilly, Lily Barna, Emőke-Ágnes Horvát

Organizations: Northwestern University, Evanston, Illinois, USA

Abstract

Google's AI Overviews (AIOs) are the first result users encounter when searching the web, shifting the work of selecting and evaluating sources from the user to Gemini. This may be especially consequential for health information, given that Google's traditional algorithm was designed to prioritize quality sources, and inaccurate information carries health risks. We evaluated 3,611 benchmark health questions, comparing each AIO against that question's top five traditional results on source selection and quality. We also evaluated the AIOs' statement grounding in their cited sources and how accurate AIOs were against physician-written rubrics. Overviews appeared for 96.4% of questions but drew on substantially different evidence than traditional search: only 44.4% of the top five results were cited by the corresponding AIO, which instead overrepresented commercial health publishers (e.g., WebMD, Healthline) at the expense of academic and government sources (e.g., Cleveland Clinic, National Health Service). AIO-cited pages scored lower than traditional results on four of six well-established credibility criteria. The grounding of AIO statements to their sources was poor: after correcting for classifier error, an estimated 56.7% of cited statements were supported by their cited page, and a further 20.8% of all statements were factual claims with no citation at all. When AIOs were graded against physicians' rubrics, they earned 38.2% of all available points (e.g., accuracy 54.2%, completeness 21.6%), and among the subset of AIOs whose physician rubric anticipated a specific inaccurate or potentially harmful claim, 22.2% of AIOs contained the harmful content. Notably, the accuracy of AIOs was unrelated to the grounding of their statements. We recommend AIO literacy interventions for users and holding AIOs at least to Google's own health-ranking standards.

Figures & tables

Classification	Low complexity	Medium complexity	High complexity	Total
Low risk	29	148	1,102	1,279
Medium risk	207	226	1,179	1,612
High risk	143	24	553	720
Total	379	398	2,834	3,611

Table 1. HealthBench Questions by Complexity and Risk

Measure	n	M	SD	Range
AI Overview length (words)	3,481	283.2	121.7	7–1,983
Sources per question: all AI Overview	3,481	7.70	4.60	0–38
Sources per question: AI Overview, citing overviews only	3,154	8.50	4.07	1–38
Sources per question: traditional	3,481	4.84	0.81	0–5

Table 2. Descriptive Statistics for the Analytic Sample. n counts questions that returned an AI Overview. Traditional results are capped at the top five by design. AI Overview citations are captured uncapped.

Panel of Figure 1 — Figure 1. Proportion of AI Overviews Citing No Sources by (A) Rubric Complexity and (B) Question Risk Level. N = 3,481 overviews. Error bars are 95% confidence intervals. Brackets mark pairwise (Holm-adjusted) chi-square tests within each panel. * p < .05. ** p < .01. *** p < .001. Proportion of AI Overviews Citing No Sources by (A) Rubric Complexity and (B) Question Risk Level

Figure 2. Paired Effect Sizes for JAMA and HONcode Measures. Points show Cohen’s d z for AI Overview citations minus traditional results, with 95% confidence intervals, across 3,103 paired questions. * p < .05. ** p < .01. *** p < .001, Benjamini–Hochberg adjusted. Paired Effect Sizes for JAMA and HONcode Measures

Figure 3. Composition of AI Overview Statements. All 24,060 statements across 3,474 overviews. Only supported and not supported statements ( n = 12,175) enter the support rate. Indeterminate statements had at least one cited page that could not be read. Composition of AI Overview Statements

Panel of Figure 4 — Figure 4. Proportion of Cited Statements Supported by a Cited Page, by (A) Rubric Complexity and (B) Risk Level. n = scorable cited statements in each level. Error bars are 95% confidence intervals. Brackets mark pairwise (Holm-adjusted) chi-square tests within each panel. * p < .05. ** p < .01. *** p < .001. Proportion of Cited Statements Supported by a Cited Page, by (A) Rubric Complexity and (B) Risk Level

Measure	n	% of available points
Score (% of available points)
All criteria	3,474	38.2
Accuracy and completeness (clinical score)	3,460	39.3
Accuracy	3,016	54.2
Completeness	2,922	21.6
Context awareness	2,401	32.3

Table 3. HealthBench Rubric Scores of AI Overviews. Scores are mean percentages of available points (points earned divided by maximum positive points per AIO); per-AIO scores can be negative, and group means are clipped to 0–100% as in HealthBench. Axis and consensus rows include only AIOs whose rubric had such criteria. Rates include only AIOs whose rubric contained the relevant criterion. n = AIOs.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Stage	Questions	AIO-cited pages	Traditional results
HealthBench questions submitted	3,611	–	–
Search result captured	3,610	–	–
AIO returned	3,481	26,805	16,833
Gate 1: overview text of at least 200 characters	3,474	26,780	16,803
Gate 2: readable pages only	–	23,657	14,443
At least one readable AIO-cited page	3,137	–	–

Appendix

Table A1. Questions and Pages at Each Filter Stage

Figure A1. Distribution of HealthBench Question Length. N = 3,611 questions. Bars are 10-word bins. The dashed line is the median (31 words) and the dotted line is the mean (43.6 words). Nine questions were 300 words or longer and are not shown. Right-skewed histogram of question length in words, with most questions under 100 words.

Criterion	Scored 1 (passed)
Authorship	A credentialed person is named as author, contributor, or medical reviewer, or the page is on a trusted health domain (MedlinePlus list, including subdomains).
Attribution	The page has a reference-section heading, at least one numbered citation link, or at least two distinct external links in the article prose.
Disclosure	The page carries a disclosure statement (a sentence of at least six words with a disclosure cue) or a named policy link.
Currency	A date the reader can see: a labeled date anywhere on the page, an unlabeled date near the top of the article, or a publication date the scraper read from the visible page text.

Appendix

Table A2. JAMA Benchmark Coding Rules

Criterion	Judgments	Agreement (%)	$\kappa$
JAMA Authorship	29	90	.75
JAMA Attribution	29	79	.56
JAMA Disclosure	29	90	.66
JAMA Currency	29	83	.66
HON Authority	29	76	.50
HON Complementarity	29	86	.26

Appendix

Table A3. Agreement Between Automated and Human Coding

Measure	Full sample	Single-turn, $\leq$ 44 words
AI Overviews citing no sources (%)	9.4 [8.5, 10.4]	5.8 [4.8, 7.1]
Low / medium / high complexity	4.2 / 6.0 / 10.6; $\chi^{2}$ (1) = 20.96***	3.4 / 4.0 / 6.5; $\chi^{2}$ (1) = 4.27*
Low / medium / high risk	6.4 / 8.1 / 17.6; $\chi^{2}$ (1) = 55.57***	4.3 / 6.5 / 7.5; $\chi^{2}$ (1) = 4.62*
Source overlap (%)	44.4 [43.4, 45.4]	46.7 [45.3, 48.2]
Source quality, $d_{z}$ (AI Overview $-$ traditional)
JAMA Attribution	$-$ 0.06 [ $-$ 0.09, $-$ 0.02]**	0.00 [ $-$ 0.05, 0.05]

Appendix

Table A4. Main Results in the Full Sample and in Single-Turn Questions of 44 Words or Fewer

Domain	AI Overview citations n (%)	Traditional results n (%)	t (3102)	p	$d_{z}$
nih.gov	2,965 (12.5%)	1,748 (12.1%)	$-$ 1.46	.172	$-$ 0.03
clevelandclinic.org	1,119 (4.7%)	1,158 (8.0%)	$-$ 15.09	$<$ .001	$-$ 0.27
mayoclinic.org	929 (3.9%)	839 (5.8%)	$-$ 7.63	$<$ .001	$-$ 0.14
cdc.gov	689 (2.9%)	495 (3.4%)	$-$ 2.61	.012	$-$ 0.05
nhs.uk	364 (1.5%)	442 (3.1%)	$-$ 10.72	$<$ .001	$-$ 0.19
webmd.com	471 (2.0%)	200 (1.4%)	5.48	$<$ .001	0.10

Appendix

Table A5. Most Frequent Domains Among AI Overview Citations and Traditional Results

Measure	M	SD	M	SD	t (3102)	p	$d_{z}$
	AI Overview citations		Traditional top 5
JAMA Authorship	0.50	0.24	0.50	0.26	$-$ 0.19	.849	0.00
JAMA Attribution	0.48	0.23	0.49	0.26	$-$ 3.17	.002	$-$ 0.06
JAMA Disclosure	0.32	0.22	0.35	0.24	$-$ 6.68	$<$ .001	$-$ 0.12
JAMA Currency	0.63	0.23	0.65	0.25	$-$ 4.87	$<$ .001	$-$ 0.09
HON Authority	0.59	0.24	0.64	0.26	$-$ 10.45	$<$ .001	$-$ 0.19

Appendix

Table A6. JAMA and HONcode Scores of Pages Cited by AI Overviews and of the Traditional Top Five

Risk level	Theme	n	Clinical score (%)	Inaccurate or harmful (%)
Low	Health data tasks	66	35.8	16.4 (55)
	Context seeking	495	35.2	14.3 (335)
	Global health	662	19.5	18.3 (410)
Medium	Responding under uncertainty	821	47.1	27.5 (625)
	Expertise-tailored communication	737	39.3	23.7 (333)
High	Emergency referrals	431	62.0	13.0 (146)

Appendix

Table A7. Clinical Scores and Inaccurate or Harmful Content by HealthBench Theme

Medical QA

Medical question answering develops and evaluates AI systems that answer clinical, patient-facing, and health-information questions.

See all 86 papers

May 13, 2026

Measuring Google AI Overviews: Activation, Source Quality, Claim Fidelity, and Publisher Impact

Google AI Overviews (AIOs) are arguably the most widely encountered deployment of generative AI, reaching over 2 billion users who may not realize the answers they see are AI-generated. Where search engines have traditionally surfaced ranked sources and left users to evaluate them, AIOs synthesize and deliver a single answer - giving Google unprecedented editorial control over what users read and know. We present a large-scale longitudinal measurement study, issuing 55,393 trending queries across 19 topical categories over a 40-day window (March 13 - April 21, 2026). We report four main findings. First, overall AIO activation is 13.7%, rising to 64.7% for question-form queries, while politically sensitive topics see markedly lower rates. Second, AIO-cited domains are more credible than co-displayed first-page results, yet nearly 30% do not appear in those results at all, indicating a source selection mechanism distinct from Google's ranking algorithm. Third, decomposing responses into 98,020 atomic claims, 11.0% are unsupported by the cited pages - with omission the dominant failure mode - and source quality and claim fidelity are largely independent. Fourth, well over half of AIO-cited pages carry display advertising, meaning publishers lose revenue when AIOs suppress the click-through, even as Google's own sponsored ads continue to appear on the same page. Together, these findings document a rapid transformation of the online information ecosystem whose consequences for epistemic security remain poorly understood.

Haofei Xu, Umar Iqbal, Jacob M. Montgomery

Washington University in St. Louis

Citation Verification Generative Retrieval Evidence-Grounded Generation IR Evaluation Generative AI Evaluation

Ask PDF

Aug 31, 2026

Sources of Truth: A Multi-Platform, Multilingual Audit of Citations in AI Mental Health Information Queries

Online health information seeking is shifting from keyword search, where users consider a ranked list of links, to conversational systems that compose a single answer and curate its citations. Source evaluation therefore passes from user to platform, yet what these systems surface is poorly characterized. We audited three free consumer products (ChatGPT, Perplexity, Google AI Overview) on twenty English mental health questions under two prompt conditions, with a subset of three also translated into six further languages of varying resource tiers. We recorded 15,942 citations across 1,140 responses and 1,713 unique domains, then classified every citation with a nine-category organizational typology applied by a deterministic classifier validated against human coding. Citations were heavily concentrated: the ten most-cited domains accounted for 43.6% of English citations, and government, commercial health, and academic sources were closely matched at roughly 22% each. Platforms differed little in typical citation volume but sharply in consistency and in the source types they favored. Explicitly requesting sources shifted composition only modestly. Non-English queries surfaced fewer citations and were routed to language-appropriate resources at significantly lower rates. We release the typology, classifier, and annotated corpus as reusable instruments for auditing generative health search.

Phuong Anh Nguyen, Jill Noorily, Matthew Flathers +7

Division of Digital Psychiatry, Beth Israel Deaconess Medical Center, Harvard Medical School, Boston, MA 02446, USA · Psychiatry and Mental Health Department, School of Medicine, Pontificia Universidad Javeriana, Bogotá 110231, Colombia · Tufts University School of Medicine, Tufts University, Boston, MA 02111, USA

LLM Evaluation LLM Auditing Generative AI Evaluation

Ask PDF

Apr 17, 2026

Authority Signals in Claude AI Health Citations: A Descriptive Analysis Using the Authority Signals Framework

This study seeks to determine the authority signals used by Anthropic's Claude AI in its presentation of sources when answering consumer health questions. While there exists a great deal of discourse around the quality of health citations that LLMs produce, there is limited information on the integrity of the sources the citations originate from, and to what extent the sources are, from what health professionals would consider, credible sources. This descriptive cross-sectional study used data from HealthSearchQA, which contains 3,172 consumer health questions curated by Google Research. After exclusions, a final dataset of 3,075 questions yielding 10,038 citations was analyzed. The Authority Signals Framework (Jacques et al., 2026) was applied to examine 10 authority signals across four domains for a disproportionate stratified sample of 542 sources. Established institutional sources accounted for 97.8% of all citations (n = 9,818). Medical Institutions were the most frequently cited organization type (36.5%), followed by Government Resources (31.6%) and Professional Associations (28.4%). Commercial Health Information comprised 2.2% (n = 220). The top 10 organizations accounted for 57.8% of all citations, with Mayo Clinic alone representing 24.7%. Among commercial sources in the focused sample, 86.4% displayed medical review statements, 82.5% used schema markup, and 71.8% had comprehensive content, while traditional institutional sources appeared in Claude's citations with or without these same markers. As Anthropic positions Claude for HIPAA-ready healthcare applications, these findings establish a baseline for Claude's citation behavior and demonstrate the utility of the Authority Signals Framework as a tool for ongoing, cross-platform evaluation of AI-mediated health information.

Erin T. Jacques, Erela Datuowei, Elizabeth Quaye +3

York College, CUNY · Department of Health and Human Performance, York College, The City University of New York, 94-20 Guy R. Brewer Blvd. Jamaica, NY 11451 · Teachers College, Columbia University +4

Healthcare Medical QA

Ask PDF

Abstract

Medical QA

Explore similar work

Measuring Google AI Overviews: Activation, Source Quality, Claim Fidelity, and Publisher Impact

Sources of Truth: A Multi-Platform, Multilingual Audit of Citations in AI Mental Health Information Queries

Authority Signals in Claude AI Health Citations: A Descriptive Analysis Using the Authority Signals Framework