Auditing Google's AI Overviews and Featured Snippets: A Case Study on Baby Care and Pregnancy

AI summary

Overview

Research area: Algorithmic auditing of AI-generated search results — specifically Google's AI Overviews (AIO) and Featured Snippets (FS) — in the high-stakes health domain of baby care and pregnancy.

Technical level: Intermediate. The paper is a measurement/audit study rather than a modeling paper; it combines automated crawling with a heavily manual annotation protocol and standard statistical tests (chi-square, Krippendorff's Alpha), so readers need comfort with empirical methods but not with model architecture.

One-sentence scope: A systematic audit of 1,508 baby care and pregnancy queries that measures how often AIO and FS appear, how consistent and relevant they are, whether they carry medical safeguards, what sources they cite, and whether query sentiment biases their answers.

What This Paper Is About

Google Search now places directly generated answers — AI Overviews and Featured Snippets — at the top of results for health queries, and users often treat these as standalone answers without checking with a doctor. The authors note that 85% of pregnant women surveyed by Giacometti et al. (2024) turned to Google for pregnancy-related information in 2024, that over 80% said online information had medium to very high influence on their pregnancy decision-making, and that only 42% reported validating that information with their doctors. The paper's goal is to measure the quality and safety of what these two features actually display in that domain.

Key Contributions

1. A first systematic audit of AIO and FS in baby care and pregnancy. The authors built a pipeline that collected 1,508 well-formed questions and crawled Google Search Engine Result Pages (SERPs) from a fixed US-based IP address, producing the first large-scale comparative evaluation of the two features in this topic.

2. A transferable multi-dimensional evaluation framework. Five manually annotated metrics — consistency, relevance, safeguard, domain category, and sentiment valence — plus a published codebook, designed so the framework can be reused for other high-stakes domains such as legal or political information seeking.

3. Detection and analysis of "suppressed" AI Overviews. The authors wrote a parser that distinguishes AIOs visible to users from AIOs present in the underlying HTML but deliberately hidden in the interface, and found 157 of the latter.

4. Evidence on co-occurring features and query-level factors. The study is the first to measure consistency between AIO and FS when they appear on the same page, and to test how question type and query sentiment relate to appearance, consistency, relevance, safeguards, source categories, and response sentiment.

Main Findings

• AIO appears far more often than FS. Visible AIOs appeared for 84% (1,272) of queries versus 32.5% (490) for FS, with the two co-occurring on 22% (322) of queries. 92.5% (1,177) of AIO responses and 88% (429) of FS responses contained a highlighted section.

• Co-occurring answers frequently disagree. Among the 322 queries showing both, 32.3% of full answer pairs and 40.7% of highlighted-text pairs were inconsistent (the abstract states inconsistency in 33% of cases). For full answers, the breakdown was 1.8% binary contradiction, 18.7% numeric mismatch, and 11.8% other problematic mismatch; for highlighted text, 2.6%, 23.2%, and 14.9% respectively. Examples ranged from feta cheese and ice cream consumption to baby acne duration and sleeping positions for pregnant women.

• Question type drives both prevalence and inconsistency. Question type was significantly associated with FS frequency (p=0.015) and AIO frequency (p<0.01). FS was least common for "Why" questions (15.69%) and most common for "When" questions (35.48%), while AIO stayed at 80–92% across all types. Consistency differed significantly by question type (p<0.001): only binary questions produced binary contradictions (11.1%), while numeric mismatches were 34.2% (full) and 48.7% (highlighted) for "How + adjective/adverb" questions and 57.4% and 62.5% for "When" questions.

• Query sentiment matters for appearance but not for consistency. FS appeared nearly twice as often for negative (48.41%) as for neutral (24.84%) queries (p<0.001). AIO also differed significantly by sentiment (p=0.0265), appearing most for negative (86.62%) and neutral (85.99%) queries and less for positive ones (76.43%). Consistency labels showed no statistically significant difference by sentiment (p=0.052).

• Relevance is high. 96.6% of AIO and 88.7% of FS answers were rated highly relevant, and only about 1% of AIO/FS were rated low relevance. In 95.5% of AIO cases and 89.5% of FS cases the highlighted text was the most relevant part of the answer.

• Medical safeguards are largely missing. Safeguard cues appeared in only 11% of AIO and 7% of FS responses. Question type was significantly associated with safeguard presence in AIO only.

• Source profiles differ sharply. Both AIO and FS cite more health-related sources than the regular ten blue links, but FS drew disproportionately from business and shopping sites. A manual credibility assessment of the 144 unique domains in the top 10% most frequently cited found that low- and medium-credibility sources account for nearly half of AIO/FS citations. Source categories were associated with question type but not sentiment.

• Suppressed AIOs look worse than visible ones. About 1% of visible AIOs were of low relevance, versus 16% of suppressed ones. Only 31% of suppressed whole AIO answers were consistent with the co-displayed FS, versus 67% of visible whole AIO answers. Qualitative review found suppressed AIOs often contained pro-life/anti-abortion or anti-vaccination content.

• No evidence of confirmation bias in AIO. When query topic and question type were controlled, the authors found no evidence that query sentiment was reflected in AIO answer sentiment (RQ4). FS did not have enough cases for the same analysis.

Methodology in Plain English

The authors started from a public dataset of 10.4 million distinct Bing search queries (ORCAS). They filtered it in three stages: keeping queries containing "baby," "babies," "pregnant," or "pregnancy"; keeping queries of at least three words and removing near-duplicates; and using gpt-4o-mini to remove queries about animals, plants, celebrities, or non-informational content. The model's classifications matched manual annotations on a 50-query validation subset 100% of the time. This left 9,516 queries.

From those, they selected well-formed questions and grouped them into six question types (binary, wh*, when, how-to, how + adjective/adverb, and why), sampling 100 per type per topic where possible — the paper reports this yielded 1,037 queries, while Table 1 reports question-type distributions totaling n=838 for baby care and n=1,012 for pregnancy. Separately, they took 157 binary queries (23 baby care, 134 pregnancy) that could be naturally rewritten in neutral, positive, and negative forms, producing 471 sentiment-varied queries. In total, 1,508 queries entered the pipeline.

Selenium agents then ran those queries on Google Chrome from a fixed New York City IP address without logging into a Google account, crawling SERPs on April 8–9, 2025. The authors also crawled from five different locations and report no significant differences in FS or AIO appearance or content by location. They parsed AIO and FS out of the saved HTML, including a custom parser to flag AIOs that exist in the HTML but are hidden from users.

For quality assessment the authors used full manual annotation for every metric except source category. Three coders scored consistency (Krippendorff's Alpha 0.74 on 100 samples); two coders scored relevance (0.98 on 100 samples), sentiment (0.81 on 33 samples), and safeguards (0.84 on 30 samples). They chose manual annotation over LLM-as-a-judge because prior work reports only moderate agreement between the two — citing a Pearson correlation of 0.62 in a similar search setting. Source domains were categorized with FortiGuard's website categories, and then the top 10% most-cited domains (144 unique domains) were manually re-graded as high, medium/contextually-dependent, or low credibility.

Why This Matters

Impact on research: The paper supplies a reusable audit methodology and codebook for generative search features, extends algorithmic auditing into AIO-FS co-occurrence that prior work had not examined, and shows how to control for topic and question type when testing sentiment bias — a control that earlier studies (Venkit et al. 2024; Li and Sinnamon 2024) did not apply. It also introduces safeguards as a measurable dimension of AI answer quality.

Real-world applications:

• Expectant parents and caregivers can understand that AIO and FS on the same page can contradict each other, and that explicit advice to consult a clinician appears in a small minority of answers.
• Clinicians and public health communicators can see where patients are likely to encounter incomplete or contradictory guidance, and which question phrasings (notably "When" and "How much/How often" questions) produce numeric mismatches.
• Search platforms and regulators can use the consistency, safeguard, and source-credibility metrics as concrete quality benchmarks for health-related AI answers.
• Health information consumers and journalists can use the finding that nearly half of top-cited domains for AIO/FS fall into low- or medium-credibility categories as a prompt to check sources.

Industry relevance: The results bear directly on Google's YMYL (Your Money or Your Life) policy commitments and on the design of direct-answer features that reached over 200 countries and 2+ billion monthly users by mid-2025. They also speak to the practice of suppressing generated answers: the suppressed AIOs were both less relevant and more inconsistent than visible ones, which suggests suppression, where it happens, correlates with lower quality.

Future Directions

• Establishing safeguard norms. The authors state that there is no established ground truth for which queries should carry safeguards, and explicitly frame safeguard presence as an open challenge and a first step for future research.

• Extending the framework to other high-stakes domains. The paper argues its metrics transfer to audits of legal or political information seeking, which would test whether the patterns found here hold outside health.

• Investigating suppression as a quality signal. Only 157 suppressed AIOs were found here; larger-scale work could test whether suppression reliably tracks low relevance and inconsistency, and what content gets suppressed.

• Broadening beyond a single domain and a two-day crawl. The study is deliberately deep on one topic and collected SERPs on April 8–9, 2025; work covering more topics, longer time windows, and personalization factors would test how stable these findings are.

Target Audience

This paper is most useful to researchers in algorithmic auditing, information retrieval, and health informatics; to platform policy and trust-and-safety teams; to clinicians and public health communicators who advise patients relying on search; and to anyone studying how generative AI features shape access to high-stakes information. It requires no machine learning background, only willingness to follow a detailed annotation protocol and its statistical results.

Authors’ abstract

Google Search increasingly surfaces AI-generated content through features like AI Overviews (AIO) and Featured Snippets (FS), which users frequently rely on despite having no control over their presentation. Through a systematic algorithm audit of 1,508 real baby care and pregnancy-related queries, we evaluate the quality and consistency of these information displays. Our robust evaluation framework assesses multiple quality dimensions, including answer consistency, relevance, presence of medical safeguards, source categories, and sentiment alignment. Our results reveal concerning gaps in information consistency, with information in AIO and FS displayed on the same search result page being inconsistent with each other in 33% of cases. Despite high relevance scores, both features critically lack medical safeguards (present in just 11% of AIO and 7% of FS responses). While health and wellness websites dominate source categories for both, AIO and FS, FS also often link to commercial sources. These findings have important implications for public health information access and demonstrate the need for stronger quality controls in AI-mediated health information. Our methodology provides a transferable framework for auditing AI systems across high-stakes domains where information quality directly impacts user well-being.