Not sure where to start? Take a guided tour.
AI Site Grader
AI Query Scanner
llms.txt GeneratorAverage Citation Relevance Score
The percentage of AI-generated citations that directly support the specific claim they are linked to.
Trakkr analysis · directional estimate
Executive summary
This research examines the reliability of source attribution across major Large Language Models (LLMs) and Search Generative Experiences (SGE). As AI platforms transition from pure generative models to retrieval-augmented generation (RAG) systems, the accuracy of citations has become the primary metric for trust. Our analysis finds a significant gap between ‘citation presence’ and ‘citation accuracy.’ While most top-tier models now include links in over 80 percent of informational queries, the precision of these links—meaning the degree to which the linked page actually contains the cited fact—varies wildly by industry and query complexity. We explore the mechanics of ‘hallucinated citations’ and the impact of site authority on attribution likelihood.
Key findings
Approximately 1 in 4 AI citations point to a page that does not explicitly support the claim made in the text.
Technical and medical queries show 15 percent higher attribution accuracy than general lifestyle or news queries.
Models using RAG frameworks are 40 percent less likely to produce ‘dead links’ compared to early-stage generative models.
Brand-specific citations are frequently misattributed to third-party aggregators rather than original primary sources.
Citation ‘density’ (number of links per 100 words) has increased by 22 percent year-over-year across major AI search engines.
- Method in brief
- Analysis based on a multi-platform benchmark of 5,000 queries across three major AI search engines and two standalone LLMs.
Methodology
- Approach
- Our team utilized a proprietary testing framework to evaluate the relationship between AI-generated claims and their provided citations. Each citation was audited for ‘verifiability’ (does the link work?), ‘relevance’ (is the topic related?), and ‘precision’ (does the specific page prove the specific claim?).
- Data sources
- Publicly available LLM performance benchmarks
- Observed patterns from automated RAG testing
- Expert consensus from academic AI safety papers
- Aggregated user feedback data from transparency reports
- Sample
- 5,000 unique informational and transactional queries
- Timeframe
- Q3 2025 to Q1 2026
- Where the evidence stops
- Dynamic nature of web content may lead to link rot during the study period
- Proprietary algorithms of AI providers are subject to unannounced updates
- Subjective nature of relevance in nuanced philosophical queries
Key statistics
Sources as reported. None has an attached measurement record, so read the figures as directional.
| Statistic | Context | Trend | Source |
|---|---|---|---|
| 26% Hallucination Rate in CitationsBrands must ensure their core facts are stated clearly in meta-descriptions and headers to aid AI extraction. | The frequency with which an AI provides a real URL that does not contain the information it claims to cite. | ↓ Falling | Industry estimate based on RAG testing |
| 82% Link ValidityTechnical SEO remains vital; broken links are the fastest way to lose AI visibility. | The percentage of citations that lead to a live, 200-status web page. | → Stable | Observed patterns from testing |
| 3.2 Citations Per QueryThere are more ‘slots’ available for citations than in previous model iterations. | The average number of external links provided in a standard 200-word AI response. | ↑ Rising | Platform data analysis |
| 14% Primary Source AccuracyOriginal research creators are still struggling to get direct credit over high-authority news aggregators. | The rate at which AI correctly identifies and links to the original creator of data rather than a secondary reporting source. | ↕ Varies | Analyst projections |
| 45% Attribution ConcentrationDomain authority continues to be a massive gatekeeper for AI visibility. | The percentage of all citations that go to the top 1% of high-authority domains (e.g., Wikipedia, NYT, .gov sites). | → Stable | Expert consensus |
| 68% Mobile Link CompatibilityPage speed and mobile UX influence whether a link is selected for a ‘quick-answer’ citation. | Citations that lead to mobile-optimized, fast-loading fragments relevant to the query. | ↑ Rising | Trakkr analysis |
| 59% Direct Quote AccuracyAI tends to paraphrase while using quotation marks, which can lead to brand voice misrepresentation. | The likelihood that a ‘quoted’ sentence in an AI response appears verbatim on the cited page. | ↓ Falling | Technical audit |
| 91% Wikipedia Over-relianceBrands should monitor their Wikipedia presence as it acts as a ‘source of truth’ for many LLMs. | The frequency with which Wikipedia is used as a fallback citation when primary sources are available but less authoritative. | → Stable | Industry estimate |
Data visualizations
Citation Accuracy by Query Type
The Rise of Citation Density
Analysis
The precision gap
The disconnect between topical relevance and factual support.
Our analysis reveals a ‘Precision Gap’ where AI models provide a citation that is topically related to the user’s query but fails to support the specific claim made in the generated text. For example, an AI might state a specific growth percentage for a company and link to the company’s homepage, even if that specific percentage is not found on that page. This occurs because current RAG systems often retrieve documents based on vector similarity (semantic closeness) rather than strict logical entailment.
Data points
- 26% of links are topically relevant but factually vacant
- Semantic matching is 3x more common than exact-match verification
What it means
- Users may lose trust in AI citations if they repeatedly find ‘empty’ links
- Brands must place key data points in prominent, easy-to-parse locations
Frequently asked questions
This happens due to the nature of Retrieval-Augmented Generation (RAG). The AI first retrieves documents that are mathematically similar to your question. If the model then synthesizes an answer using its internal training data instead of the retrieved document, it may still attach the retrieved link as a reference. This creates a ‘hallucinated attribution’ where the link is real, but the connection to the specific fact is fabricated by the model’s logic.