hausresearch published a focused audit last week, and the conclusion is blunt: roughly one-third of Perplexity's citations contain no matching figures in the source text.
The report spent some time on the Hacker News front page with 32 points and 6 comments. For a small independent audit with transparent methodology and a concrete sample, that's a decent amount of discussion.
The more revealing angle isn't the error rate itself, but the type of error — it points to an underappreciated structural weakness in AI search.
The audit: 33% of citations don't check out
hausresearch took a measured approach: they selected a batch of questions requiring specific figures (like "What was Company X's Q3 revenue?"), had Perplexity answer them, then manually verified each citation to confirm the source actually contained that number.
The sample is modest (the report describes hundreds of question-answer pairs), but the screening criteria were strict — only queries demanding precise numbers were tested, because that's exactly the scenario where users rely most heavily on citations.
Results:
- About 33% of citations had no matching figure in the source document
- The error pattern is consistent: the model produces a plausible-looking number, attaches a plausible-looking source link, but clicking through reveals either a different number or no mention of that figure at all
The report doesn't directly benchmark ChatGPT Search, Gemini, or Grok on citation accuracy (that kind of head-to-head requires significant manual labor, which is tough for an independent audit to absorb), but based on Hacker News discussion threads and scattered hands-on tests, this isn't a Perplexity-only issue — it just got quantified first.
Hallucinated citations are a different beast from regular hallucinations
Standard LLM hallucinations are familiar territory — invented people, fabricated papers, events that never happened. Users know to cross-check.
Hallucinated citations are something else:
- They come with a URL that loads
- They have a title that looks like a real source
- They might include page numbers and paragraph references, formatted to look academic
- But the source text simply doesn't contain the number or quote the model claims
This kind of hallucination is harder to catch. The core reason users turn to AI search is "citations included, verifiable" — seeing a citation triggers a mental shortcut that the first layer of verification has passed. When the citation itself is fake, that verification mechanism collapses.
The examples in the hausresearch report (the full Q&A pairs are public) are telling: Perplexity returned what looked like an authoritative earnings report citation, but the actual source contained a different number. This isn't a case of "off by a decimal point" — it's the entire factual chain breaking at the source.
Why this is worse than ChatGPT making up stories
When users use ChatGPT to write a story, they don't treat it as a factual source in the first place.
Perplexity users come for the citations.
Product positioning sets user expectations: Perplexity's pitch is "sources included, links included, traceable." The UI places citations right after each statement, and the design clearly emphasizes "this is a grounded answer." Users choose it over a standard LLM specifically to skip the manual verification step.
Turns out: 33% of the time, the skipped step is the most important one.
This isn't a model capability failure. It's a product trust failure.
When a standard LLM hallucinates, users blame the model. When an AI search product hallucinates its citations, users blame the product. Once trust shifts from "can the model answer correctly" to "can the product verify itself," what's at stake expands beyond technology into product promises.
Where RAG's citation mechanism actually breaks down
From an engineering standpoint, a typical RAG citation pipeline looks like this:
- Retrieval: Pull candidate document chunks from a vector store or search engine based on the query
- Rerank: Use a more precise model to reorder the candidates and surface the most relevant ones
- Citation alignment: The model generates an answer grounded in the selected chunks and tags each sentence with its corresponding citation
A failure at any stage can produce a hallucinated citation:
- Retrieval failure: The system pulls documents that are topically related but don't match the specific detail — the model then generates from the wrong source material
- Rerank failure: The reranker promotes a less relevant chunk, and the model cites the wrong source when generating
- Alignment failure: This is the sneakiest one — the model's generated sentence is correct (it may genuinely be grounded in a source passage), but the citation it attaches points to a different passage. This typically happens in long answers with mixed sources, when the model fumbles the "which citation does this sentence belong to" mapping
The hausresearch findings likely point to alignment-stage failures as the main culprit, with retrieval-stage issues playing a supporting role. If it's alignment, the model's self-tracing accuracy — its capacity to remember "where did that sentence come from" — sits around 67%.
This isn't something Perplexity can fix in isolation. Any AI search product following the RAG-plus-generation approach can hit the same pothole. ChatGPT Search from OpenAI and AI Overviews from Google almost certainly have similar issues — they just haven't been subjected to the same rigorous independent audit.
How users can avoid getting burned
Short answer: don't trust them blindly.
A few specific tips:
- Numbers, dates, names, quotes — for these precise pieces of information, always click through and verify the source. AI search is reasonably accurate on broad summaries and thematic synthesis, but hallucination rates spike on exact numbers
- High-stakes scenarios (report writing, investment analysis, academic citations) — don't rely on AI search as your only source; at minimum, keep a manual verification step in the loop
- Cross-source check: Ask the same factual question across 2–3 AI search products and compare their citations. If all three cite the same number, it's probably right; if only one provides a specific figure, be cautious
- Reverse verification: Don't just confirm the link exists — confirm the link supports the claim. A working URL doesn't mean the content matches
Where AI search works fine:
- Quick orientation in an unfamiliar field (topical level)
- Hunting for candidate source leads (then filter manually)
- Summarizing long documents (provided you already trust the document itself)
Where to skip it:
- Financial, medical, or legal information requiring precise numbers
- Academic writing citations (your professor will check)
- Any scenario where "I can't afford to be wrong" applies
Editorial take
This audit deserves serious attention because it turns a widely discussed but never quantified phenomenon — citation hallucination in AI search — into a concrete number (33%).
My read: citation alignment will become the next competitive battleground in AI search. The current race is about "can it answer"; the next round will be about "do the citations hold up." That shift is especially lethal for products like Perplexity that built their moat on citations — if users stop trusting the citations, the moat crumbles.
A counter-consensus observation: this won't drive users away from AI search, but it will change how they use it. From "I believe what you say" to "I use you as a source of leads." The design implications might matter more than the technical improvements — future AI search products may more explicitly label "which parts are model summaries" versus "which parts are direct source quotes," rather than today's ambiguous blend.
One uncertainty: hausresearch's sample isn't large, so strictly speaking the confidence interval on that 33% figure is probably wide. But even if the real error rate is only 20%, it's unacceptable for something that markets itself as "citation-first."
Citation alignment has historically been treated as a minor RAG engineering detail. After this audit, companies should start treating it as a first-class concern — either by improving alignment accuracy or by redesigning the citation mechanism entirely (for example, forcing the model to emit "where this sentence came from" as a separate token during generation, which makes post-hoc verification tractable).