ObsurfableGoogle AI Overviews are not a footnote on search results anymore. They are the answer for a growing share of queries — including, as of late September 2026, many brand-name searches third-party trackers follow daily.
A landmark independent study from Washington University in St. Louis puts numbers on what marketers have felt anecdotally: citations can look trustworthy while claims drift from the sources they point to. The work was presented at the October 2026 ACM Internet Measurement Conference and is available as arXiv:2605.14021 (Measuring Google AI Overviews: Activation, Source Quality, Claim Fidelity, and Publisher Impact).
Authors Haofei Xu, Umar Iqbal, and Jacob M. Montgomery issued 55,393 Google searches on trending topics over 40 days (March 13–April 21, 2026), collecting each AI Overview, its citations, co-displayed first-page results, and cited page text.
Four findings should reshape how you think about being cited vs being represented accurately.
Finding 1: Activation is selective — questions dominate
Overall, AI Overviews appeared on 13.7% of trending queries in the sample. That headline rate hides sharp structure:
| Query shape | AIO activation rate |
|---|---|
| All trending queries | 13.7% |
| Question-form queries | 64.7% |
| Non-question queries | 9.5% |
Question-shaped prompts triggered overviews at 6.8× the rate of non-questions. Politically sensitive categories showed markedly lower activation — the authors read that as evidence of undisclosed editorial discretion in triggering logic, not a neutral "always summarize" rule.
Implication for content strategy: Pages that answer a question in the heading — pricing, eligibility, comparisons, how-to — compete for a surface that activates far more often than corporate boilerplate. This aligns with question-intent citation patterns seen in multi-engine monitoring studies.
Finding 2: Cited domains are credible — but not the same as page-one organic
AI Overview citations skew more credible than co-displayed first-page organic results in the authors' source-quality metrics — pushing back on earlier work suggesting the opposite.
Yet 29.8% of AIO-cited domains did not appear in the first-page organic results shown alongside the overview for the same query.
That is the smoking gun for SEO teams who equate rank with AI eligibility:
- Google's overview retrieval uses a distinct selection path from the classic ten blue links
- Your page can be cited without ranking for the visible query — likely via query fan-out and sub-queries, consistent with Ahrefs' declining top-10 overlap measurements
Implication: Track citation presence, not rank alone. Optimize narrow pages that win sub-questions, not only head terms.
Finding 3: 11% of claims fail claim–citation fidelity
The authors decomposed overviews into 98,020 atomic claims and checked each against the text of cited pages they could fetch.
| Outcome | Share of claims |
|---|---|
| Supported (clear or broad) | ~89% |
| Unsupported (total) | 11.0% |
Unsupported breaks down roughly as:
- Omission — claim not found in cited text (dominant mode)
- Contradiction — cited text conflicts with the claim
- Conflict — multiple cited sources disagree
Critically, source quality and claim fidelity were largely independent. A high-trust domain in the citation list does not guarantee the sentence beside it was extracted faithfully.
The 5.3%–11.0% range
The team could not crawl major social and video hosts (Facebook, Instagram, Reddit, TikTok, X, YouTube). When those citations are treated as unknowable rather than unsupported, the authors argue a generous lower bound around 5.3% unsupported claims.
Read the public number as one in ten claims may not hold up to the cited page in the strict audit — and at least one in twenty even under favorable assumptions.
Implication for brands: Being cited is reputationally double-edged. If Google's summary misstates your pricing, guarantee, or medical claim, the link to your domain may still look like an endorsement. Accuracy reviews matter as much as citation counts.
Finding 4: Publisher economics — ads on cited pages, clicks lost
Well over half of AIO-cited pages carried display advertising in the study. When overviews satisfy the query without a click, publishers eat the viewability loss — while Google's own ads can still appear on the SERP.
That finding fueled October 2026 news coverage (Phys.org summary, Digital Information World) as Google tests a limited AI Contribution publisher pilot. Payment experiments do not change claim fidelity; they acknowledge revenue tension.
What marketers should do differently
1. Separate KPIs: citation, mention, fidelity
- Citation rate — is your URL in the card?
- Mention rate — is your brand named in the prose?
- Fidelity — does the prose match your canonical facts?
Tools that only count links miss the 11% failure mode — and the source-to-mention gap where sources appear without brand credit.
2. Write for atomic claim extraction
The audit rewards pages where one sentence = one verifiable fact, with the same numbers and qualifiers repeated in schema-visible fields (title, meta, FAQ, tables). Vague marketing copy gives the synthesizer room to omit or soften until the claim no longer matches.
3. Monitor question queries hardest
At 64.7% activation, question-form SERPs are where wrong summaries scale. Map your top 50 buyer questions in Search Console and run a monthly claim check against live overviews.
4. Do not treat third-party citations as "someone else's problem"
When overviews cite Reddit or a review site, the brand claim may be about you. Earned media strategy is partly fact stewardship on properties you do not own.
How this interacts with September's branded-overview surge
DemandSphere reported AI Overviews on ~80–90% of tracked brand-name queries by late September 2026 (analysis). The WashU study did not focus on navigational brand queries, but the fidelity math travels: if buyers read an AI summary before your site, unsupported claims hit branded intent too.
How Obsurfable fits
Independent audits describe the platform. Your job is to track your category on the prompts buyers actually run. Obsurfable records observations across engines — answers, mentions, citations — in a public corpus you can browse in Explorer and match to your company in the App.
Use repeated runs to measure whether corrections on source pages change mentions — not single screenshots. Volatility is expected; see why rank-style tracking fails.
FAQ
Is 11% "bad" compared to human editors?
The authors frame it as material for a system shown to billions, not a product grade. Your bar may be stricter for YMYL categories.
Did the study include AI Mode?
The paper targets AI Overviews on Google Search during the March–April 2026 window. AI Mode behavior can differ; treat this as a lower bound on Google's generative search stack, not the whole product map.
If 30% of cited domains are off page one, should we ignore SEO?
No. SEO still drives crawlability and many citations. The point is non-ranking pages can still be in the overview card — so long-tail answer pages are not "wasted" effort.
Does higher source quality mean we are safe?
No. Quality and fidelity were independent in the study. Continue manual branded SERP reviews.
Bottom line
Across 55,393 queries and 98,020 claims, WashU researchers show AI Overviews are selective, retrieval-distinct from blue links, and imperfect at matching citations to sentences. About 11% of claims did not hold up to cited pages in their crawl — with a stated ~5.3% floor. For brands, the lesson is not "avoid citations." It is measure mentions, read the prose, and fix the sources — because the link icon is not a fact-check.