info.link logo
Go back
Trust & Provenance

How accurate are AI search answers? Most claims hold, attribution often fails

Google AI Overviews left 11% of claims unsupported by the pages they cited, mostly through omission. Naming the source of a news excerpt fails far more often: eight AI search engines got over 60% of attribution queries wrong. Citations raise user trust even when they are wrong.

AI search answers match their sources on most claims and fail often on attribution, and the error rate depends on the task. Google AI Overviews left 11.0% of 98,020 atomic claims unsupported by the pages they cited arXiv · 2026. Asked to name the source of a news excerpt, eight AI search engines got more than 60% of 1,600 queries wrong Tow Center for Digital Journalism, Columbia · 2025. A citation raises user trust whether or not it is correct arXiv · 2025. The sections below take claim fidelity, attribution, deep research, user trust and brand facts in turn.

AI Overviews accuracy: 11% of claims lack support in the cited pages

Claim fidelity measures whether each statement in an AI answer is backed by the page it cites. Researchers at Washington University in St. Louis ran 55,393 trending queries over 40 days in March and April 2026 arXiv · 2026. They split the resulting AI Overviews into 98,020 atomic claims and found 11.0% unsupported by the cited pages.

Omission was the dominant failure mode in that audit arXiv · 2026. Read the other way, about 89% of claims were supported. The figure covers one Google surface, trending queries and a single 40-day window. It describes AI Overviews in spring 2026, not AI search in general.

AI search engines get source attribution wrong more than 60% of the time

Attribution is a harder test: naming the headline, publisher, date and URL behind a quoted passage. The Tow Center at Columbia ran 1,600 such queries, 200 per engine, across eight AI search engines Tow Center for Digital Journalism, Columbia · 2025. More than 60% of answers were incorrect, from 37% on Perplexity to 94% on Grok 3.

The engines rarely signalled doubt. ChatGPT misidentified 134 of 200 articles and flagged uncertainty only 15 times Tow Center for Digital Journalism, Columbia · 2025. Paid tiers did not help. Perplexity Pro and Grok 3 answered more queries correctly than free versions but had higher error rates, because they rarely declined.

Broken links were common on some engines. More than half of the responses from Gemini and Grok 3 cited fabricated or broken URLs Tow Center for Digital Journalism, Columbia · 2025. On Grok 3, 154 of 200 citations led to error pages. The report ran each excerpt once, in early 2025, on models since replaced. Naming a news source is also narrower than most questions buyers ask.

Deep research agents keep links valid but get facts wrong

In deep research agents, links mostly work and facts often do not. A May 2026 preprint scored inline citations from 14 language models on three checks: a working link, relevant content and factual consistency arXiv · 2026. The strongest frontier models kept link validity above 94% and relevance above 80%.

Factual accuracy for those same models ranged from 39% to 77% arXiv · 2026. More searching made it worse. As tool calls rose from 2 to 150, fact-check accuracy fell by about 42% on average across two frontier models.

Earlier work found the same gap in consumer answer engines. A peer-reviewed FAccT 2025 study of You.com, Perplexity and BingChat found hallucination and inaccurate citation common Salesforce AI Research et al. · 2025. Its 21-participant user study identified 16 limitations of answer engines compared with traditional search.

Citations raise trust in AI answers even when the citations are wrong

Readers treat a citation as a sign of accuracy. In a preregistered randomised experiment on a US-representative sample, reference links significantly raised trust in generative AI search arXiv · 2025. The effect held even when the links were incorrect or hallucinated. On average, participants trusted generative AI search less than traditional search.

Showing the system's own confidence lowered trust and sharing, whether that confidence was high or low arXiv · 2025. Higher trust predicted more clicks and less time spent evaluating results. Set against the fidelity figures above, a citation is a weak signal that the claim beside it is correct.

Field data adds a counterweight. In a March 2026 field experiment, users assigned to Google AI Mode reported lower trust, satisfaction and usefulness than the control group Wang et al. · University of Pennsylvania; Northeastern University · 2026. That arm imposed AI Mode as the default for seven days, so it measures assigned use, not chosen use.

Source credibility does not predict AI answer accuracy

A credible cited domain does not make the claim drawn from it more faithful. In the AI Overviews audit, source quality and claim fidelity were largely independent of each other arXiv · 2026. Cited domains were more credible on average than the first-page results shown beside them.

One self-reported experiment tests the other direction, a false source that gets retrieved. A Peec AI researcher built about 50,000 pages of invented statistics for roughly $50 and reports about 72,000 ChatGPT-User requests in 30 days Yesilyurt · Peec AI · 2026. The figures come from the author's own analytics, and whether ChatGPT used the numbers is unverified. It is a single unreplicated site; treat it as provisional.

What AI accuracy studies mean for brand facts

The accuracy studies above test news excerpts, trending queries and research reports, not brand facts. The available brand data shows who supplies the material instead. Across almost 100,000 outside-source citations in AI answers about brands, about 40% pointed to the brand's direct competitors Khallad · SurfacedBy · 2026. The vendor calls its brand sample small and mixed, so the direction is firmer than the decimals.

Pricing answers follow the same pattern. Across about 7,600 responses to a Cloud 100 pricing prompt, company pricing pages were cited in 46% of answers but first in only 12% Poyar · Profound; Growth Unhinged · 2026. Vendr, Reddit and G2 each appeared in 15.9% to 18.7% of responses. The study used one prompt template and covers B2B cloud companies only.

Structured owned pages raised accuracy in one test harness. Across 349 queries, an enhanced entity page lifted LLM-judged accuracy from 3.62 for plain HTML to 4.69 on a five-point scale Volpini et al. · WordLift · 2026. WordLift, which sells that page format, ran the study with an LLM as judge and no live engine.

FAQ
Frequently Asked Questions

Sources

Sources are tiered per our methodology & sources page.

Tier A — Strongest evidenceRead source

AI in Search Reduces Publisher Referrals Without Improving User Experience: Experimental Evidence (arXiv 2608.18352)

University of Pennsylvania; Northeastern University · Stephanie T. Wang et al. · 2026

Key finding

Assignment to AI Mode cut external click-through by 18.8 percentage points (95% CI -22.2 to -15.3) and daily search sessions by 0.92, while lowering trust, satisfaction and usefulness. Hiding AI Overviews raised click-through by 8.8 points. Users searched rival engines 11.2 points more often. Caveat: forced seven-day AI Mode default, not voluntary use; baseline AI Mode use was 0.6% of searches.

Methodology note

arXiv preprint 2608.18352 (cs.IR), submitted 2026-08-18, by Wang, Gleason, Bart, Wilson and Metaxa. Preregistered browser-extension field experiment, March 2026: three baseline days then seven days in one of three arms; 1,100 behavioural, 956 survey. Hide arm reached only 51.1% of Overviews after a Google HTML change, reported as LATE. HTML full text read directly.

arXiv·Accessed
Key finding

Evaluates source attribution in LLM deep research agents (such as ChatGPT, Perplexity, and Gemini deep research modes). Finds that cited URLs are frequently invalid, broken, or unrelated to the claim being attributed. The paper introduces a parser and evaluation framework to measure attribution validity at scale, exposing systematic citation-quality gaps across vendors.

Methodology note

arXiv preprint 2605.06635 (May 2026). Direct fetch returned an empty PDF body; abstract and methodology cross-verified via arxiv API listing. The paper benchmarks attribution validity across multiple commercial deep research agents.

arXiv·Accessed
Key finding

Google AI Overviews appeared on 13.7% of queries overall and 64.7% of question-form queries. Politically sensitive topics saw lower rates. AI Overviews cite domains more credible on average than co-displayed first-page results, but nearly 30% of cited domains do not appear in those results at all. 11% of 98,020 atomic claims were unsupported by the cited pages, with omission the dominant failure mode. Half of cited pages carry display advertising.

Methodology note

Researchers from Washington University in St. Louis issued 55,393 trending queries across 19 topical categories over a 40-day window (March 13 to April 21, 2026), measuring AI Overview activation rates, domain credibility, claim fidelity (decomposing responses into 98,020 atomic claims), and advertising on cited pages.

arXiv·Accessed
Tier A — Strongest evidenceRead source

Structured Linked Data as a Memory Layer for Agent-Orchestrated Retrieval

WordLift · Andrea Volpini et al. · 2026

Key finding

On a 1-5 LLM-judged accuracy scale across 349 queries, JSON-LD moved plain HTML from 3.62 to 3.89. An enhanced entity page reached 4.69. Both headline figures, 29.6% and 29.8%, are measured against the same standard-RAG baseline: against agentic retrieval on plain HTML (4.36) the entity page adds only 0.34. Enhanced+ reaches 4.85/5, but its gain over the base enhanced page is not significant (d=0.08).

Methodology note

arXiv preprint 2603.10700 (v1, not peer-reviewed) by Volpini, Raad, Gamba and Riccitelli at WordLift, whose product is the entity-page format that wins. 158 entities, 349 queries, four verticals, seven conditions, 2,443 evaluations; Gemini 2.5 Flash generates, Gemini 3.0 Flash judges, no human evaluation. Full text read directly 23 September 2026.

arXiv·Accessed
Key finding

Argues that current AI search engines' promise of factual, verifiable, source-cited responses is partly illusory. Empirical analysis across major systems shows frequent unsupported claims even when citations are present, weak ranking of citations by relevance, and citation patterns that systematically advantage well-resourced incumbents. Calls for new evaluation standards before mass deployment.

Methodology note

arXiv preprint 2410.22349 (October 2024). Direct fetch returned the abstract page. The paper provides both critique and empirical evaluation across several commercial AI search systems, with methodology and per-system results in the PDF.

FAccT 2025 (ACM)·Accessed
Tier A — Strongest evidenceRead source

Human Trust in AI Search: A Large-Scale Experiment

arXiv · 2025

Key finding

Across ~12,000 queries in seven countries and a preregistered randomised experiment on a US sample, participants trusted GenAI search less than traditional search on average, but adding reference links and citations to GenAI answers significantly increased trust, even when those citations were incorrect or hallucinated. Uncertainty highlighting reduced trust whether confidence was high or low.

Methodology note

arXiv preprint 2504.06435 (April 2025) by Haiwen Li and Sinan Aral (MIT Sloan). Preregistered randomised experiment on a US-representative panel, paired with a 12,000-query, 80,000-result global exposure measurement across seven countries. 23 pages, six figures. Direct fetch on arxiv.org confirmed authorship and methodology.

arXiv·Accessed
Tier A — Strongest evidenceRead source

AI Search Has a Citation Problem (Tow Center Report)

Tow Center for Digital Journalism, Columbia · 2025

Key finding

Across eight AI search engines tested, more than 60% of news-attribution queries received incorrect answers. Perplexity got 37% wrong; Grok 3 got 94% wrong. Premium paid models were no more accurate than free ones, and often produced confidently incorrect answers without flagging uncertainty. Several engines retrieved content from publishers that had explicitly blocked their crawlers.

Methodology note

1,600 queries were run across ChatGPT Search, Perplexity, Perplexity Pro, DeepSeek Search, Microsoft Copilot, Grok-2, Grok-3, and Google Gemini. The researchers selected 10 articles from each of 20 publishers, used direct excerpts as queries, and asked each chatbot to identify the headline, publisher, publication date, and URL. Responses were manually graded against six categories.

Columbia Journalism Review·Accessed
Tier B — Citable with caveatsRead source

What AI agents really think about your pricing (Profound x Growth Unhinged)

Profound; Growth Unhinged · Kyle Poyar, Nikolas Laskaris · 2026

Key finding

Across about 7,600 AI responses to a pricing prompt for the Cloud 100 (29 July to 10 August 2026, six engines), company pricing pages were cited in 46% of responses but first in only 12%. Of 77 companies with public pricing pages, 57 were fully readable to bots. Plaid's pricing page was cited first 42% of the time, with billing docs in 70% and FAQs in 50% of answers.

Methodology note

Single-vendor study by Kyle Poyar of Growth Unhinged and Nikolas Laskaris of Profound, published 3 September 2026 and fetched directly. One prompt template, Cloud 100 sample, six engines. Response counts differ slightly (about 7,600 versus 7,675), and the authors say they cannot prove llms.txt causes better retrieval. Partnership disclosed.

Profound Blog·Accessed
Tier B — Citable with caveatsRead source

When AI Describes Your Brand, It Cites Your Competitors (SurfacedBy)

SurfacedBy · Ali Khallad · 2026

Key finding

SurfacedBy analyzed almost 100,000 citations to outside sources behind AI answers about a sample of brands, across ChatGPT, Claude, Gemini, Perplexity, and Google AI Mode over about three months in spring 2026. Around 40% of the outside sources cited about a brand were that brand's direct competitors; the typical brand sat above a third, and nearly nine in ten had at least a quarter. Brand-owned pages are a smaller stream.

Methodology note

First-party study by SurfacedBy, an AI-visibility tracking vendor with commercial interest, published 29 June 2026. Nearly 100,000 outside-source citations resolving to 10,000-plus domains, tagged competitor or third party using each brand's own competitor set, aggregated brand by brand. Authors stress a small, mixed brand sample; decimals provisional. Verified by direct fetch.

SurfacedBy Blog·Accessed
Tier C — Tactical signals onlyRead source

Publish stats, get cited: The tactic that put my data all over the web (Peec AI)

Peec AI · Metehan Yesilyurt · 2026

Key finding

Peec researcher Metehan Yesilyurt reports launching stateglobe.com in March 2026 with about 50,000 invented statistics pages (plus 15,000 listicles later) for roughly $50. He reports 330+ referring domains, over 600,000 bot requests in 30 days (about 72,000 from ChatGPT-User) and second place for visibility on tracked prompts on 7 October. He says the invented figures were cited by an academic paper and a bank; none of these claims is independently verified and the citing institutions are unnamed.

Methodology note

Self-reported single-site experiment by Metehan Yesilyurt, GEO Researcher at Peec AI, an AI search monitoring vendor. Byline date 7 October 2026, page metadata 13 October; fetched directly. Traffic and ranking figures come from the author's own analytics and third-party tools, are unreplicated and are confounded by Google updates.

Peec AI Blog·Accessed

About the author Max Ackermann

Max Ackermann is founder and Managing Director of info.link, the product data platform that makes brands visible in AI search and connects every physical product to the web through GS1 Digital Link. He writes about AI search and generative engine optimization (GEO), AI-powered commerce, and how brands can structure product data for ChatGPT, Gemini, Perplexity, and retailer AI assistants like Amazon Rufus. For the past two years he has built the pipelines that put structured product data into AI answers, and run the experiments that test what actually moves AI citations.

Max has 20+ years of experience building digital products and businesses. He previously led McKinsey's Corporate Venture and Design teams across Europe, and as Managing Director of a leading US digital agency he built platforms with Nike, Google, Meta, and Airbnb. He founded the UX Design program at Central Saint Martins College, University of the Arts London, and is a Fellow of the UK's Higher Education Academy. Based in Hamburg, he works closely with GS1 on Digital Link adoption; info.link is headquartered in Hamburg and Berlin and counts GS1 Germany among its investors.

Follow Max on LinkedIn.

Interested?

From compliant digital labels to AI-verified product answers, we help leading brands ensure their products are visible and accurately represented everywhere consumers look. Book your free consultation and demo.

digital label preview
digital label preview
digital label preview
How accurate are AI search answers, and where they fail | info.link