The Citation Problem in AI Search
Assistants increasingly answer without sending traffic. Publishers and readers are both affected, in different ways.
A search engine that answers a question directly still, usually, shows you where the answer came from. An assistant that answers the same question may cite nothing, may cite selectively, or may cite a page that does not contain the claim it is attached to. The third case is the one that should worry publishers and readers equally, and it is the least discussed.
Start with what is reasonably well established. Assistants that perform live retrieval do produce links, and those links are often accurate. But several independent evaluations of AI search products have found meaningful rates of confident attribution error: answers presented without qualification, sourced to pages that do not support them, or attributed to syndicated copies rather than originators. Rates vary substantially by product and by question type, and any specific figure should be treated as a snapshot of one system at one moment rather than a property of the category.
Two distinct problems sit underneath that finding, and conflating them produces bad analysis.
The first is economic. When an assistant answers fully, the reader has no reason to click. For publishers whose business depends on visits — advertising, subscription funnels, affiliate revenue — the value created by their reporting is captured at the interface and not returned. This is a continuation of a decade-long trend rather than a break from it; featured snippets and knowledge panels did the same thing at smaller scale. What is new is the proportion of queries that can be satisfied without a visit, and the fact that the answer is synthesised from several sources, which dilutes attribution even when it is present.
The second is epistemic, and it is more serious. A citation in an assistant's answer is a claim about provenance that the reader is not in a position to check without doing the work the assistant was supposed to save them. When that claim is wrong, the error inherits the authority of the format. Footnotes signal diligence; a footnote to a page that does not say the thing is worse than no footnote, because it converts an unverified assertion into an apparently verified one.
For publishers, the strategic responses available are narrower than the advice industry suggests. Blocking crawlers protects content from ingestion but also from citation, which removes the only remaining benefit. Licensing deals are available to a small number of large publishers and not to anyone else. Structuring content so that claims are attributable — clear datelines, explicit sourcing, stable URLs, factual statements that survive extraction without their surrounding context — improves the odds of being cited correctly, and is also just good editorial practice. The most durable position is being the primary source of something: original data, original testing, original reporting. Synthesis of other people's work is exactly what assistants do cheaply.
For readers, the practical adjustment is to treat a cited answer as a starting point rather than a verified one, particularly where the stakes justify a check. The citation tells you the system found a page. It does not tell you the page said this.
The limitations of this piece should be explicit. The published audit figures come from evaluations with differing methodologies, question sets and dates, and the products have changed since; we have deliberately avoided quoting a single headline percentage because the number would be more precise than the evidence supports. Traffic-impact data is largely self-reported by publishers with an interest in the framing, and clean counterfactuals are scarce. And the systems are improving on exactly this dimension, which means any assessment of attribution quality has a short shelf life.
What is unlikely to change is the structural point. An interface that answers rather than directs breaks the link between the value a piece of reporting creates and the mechanism by which its creator is paid. That is a question about business models and, eventually, about policy. It is not one that better citation accuracy alone will resolve.
Every claim above is sourced to a document, a named person, or a record we hold. Where we could not verify a claim, we say so. Read our standards and corrections policy →
