Deep research cites you but gets you wrong

A deep research agent is an AI system that runs many search and fetch steps over minutes rather than seconds, then writes a long cited report, and the evaluation work published this year shows something uncomfortable about that surface: the citations look excellent and the claims attached to them frequently are not supported by the page being cited.

This matters because deep research reports are the highest-stakes AI output most buyers will ever read about your category. Somebody asks for a vendor comparison, waits five minutes, gets twelve pages with sixty footnotes, and treats the result as researched. If your page is footnote 34 next to a sentence your page does not actually support, you have been cited and misrepresented at the same time.

#The citation looks fine. The claim does not.

The clearest measurement comes from Cited but Not Verified, which evaluated 14 models on 130 research queries drawn from DeepResearch Bench and BrowseComp and scored every citation on three separate dimensions: whether the URL resolves, whether the source is topically relevant to the claim, and whether the specific factual assertion is supported by the source.

The first two dimensions are close to solved. In that evaluation, link validity sat above 94% and topical relevance above 80% across the frontier models. The third dimension is not solved at all. Fact-check scores for frontier models ranged from 38.9% to 76.8%, with Claude Opus 4.5 at the top of the range and GPT-5 Mini at the bottom. Open-source models were worse on every axis, and fewer than half of them completed a cited report at all, with task success between 17% and 40% against 83% to 100% for frontier models.

Read those three numbers together. The agent finds the right page, links to it correctly, and then writes a sentence the page does not back up. That is not a retrieval problem. That is a synthesis problem, and it is the one you cannot fix by ranking better.

#The longer it researches, the worse it cites

The finding with the most operational weight is about depth. As search depth scaled from 2 tool calls to 150, fact-check accuracy fell roughly 42% on average. The per-model spread is wide: GPT-5.4 dropped from 78.6% to 16.7%, while Claude Opus 4.6 held far better, moving from 80.0% to 57.9%.

That runs against the intuition everyone brings to these products. More sources should mean a better grounded report. What the data shows is that more sources means more claims to bind, longer context to hold them in, and more opportunity for a number to drift away from the sentence that qualified it. The deep research mode is the mode where your page is most likely to be quoted incorrectly, and it is the mode marketed as the rigorous one.

#The mistake happens after your page is read

If the error were happening at retrieval, on-page work would fix it. It is not. Who is the Agent to Blame?, published in August 2026, instrumented multi-agent deep research pipelines and attributed each citation error to the specific agent that produced it.

In the AI-Q system, 84.7% of final-report errors originated at the orchestrator, the component that synthesizes agent outputs into the finished report. The gap between stages is stark: search snippets carried a 3.8% mistake rate while the researcher stage carried 70.8%. In MS-Agent the searcher's synthesized output sat at 0.9% against 16.8% at the orchestrator.

The signal is consistent. Retrieval brings back roughly correct material. The damage is done when that material is compressed into prose. The same paper found that simple interventions raised citation recall by 5 percentage points and precision by 3 to 7 points without hurting output quality, which tells you the failure is a fixable engineering gap on the vendor side rather than a limit of the technology. It also tells you it is not currently fixed.

#Broken links are a smaller, separate problem

The URL layer has its own failure mode, and it is worth keeping distinct. Detecting and Correcting Reference Hallucinations checked 53,090 URLs from 10 models on DRBench plus 168,021 URLs on ExpertQA, and separated non-resolving links into two categories: stale, meaning the page once existed and has a Wayback Machine snapshot, and hallucinated, meaning no snapshot exists and the URL probably never did.

Across tested models, 3% to 13% of citation URLs were hallucinated and 5% to 18% were non-resolving. Engines differ by a lot: Gemini 2.5 Pro deep research hallucinated 13.3% of the 11,309 URLs tested, while OpenAI deep research came in at 3.5% and Claude models between 3.0% and 3.2%. The paper also notes that deep research agents emit substantially more citations per query than search-augmented chat while hallucinating URLs at higher rates, so raw citation counts from these surfaces are inflated relative to chat counts before you check anything.

The stale category is the part you own. A URL you retired, redirected badly, or moved without a redirect becomes a dead citation in somebody's report, and no reader can tell your link rot from the model's invention.

#What to change in how you measure

Three practical shifts follow.

Separate the surfaces. Deep research citations and chat citations come from different pipelines with different error profiles, so pooling them produces a number that describes neither. This is the same denominator discipline that makes AI citation studies disagree with each other when they look like they are measuring the same thing.

Score the claim, not the link. A citation audit that checks whether your URL appears is measuring the dimension the engines have already largely solved. The dimension in play is whether the sentence next to your link is something your page says. For each deep research citation you find, record the claim as written, then mark it supported, distorted, or unsupported. That is a manual read, and it is the only read that reflects the actual risk.

Do not confuse this with absorption. We have written before about citations that appear without shaping the answer and about counting verified mentions rather than raw ones. Those measure whether you were used and whether you were real. This measures whether what was said about you is true. A citation can pass both earlier tests and still fail this one.

#What you can actually control

Very little of the synthesis layer is yours. Two things are.

Keep the qualifier in the same sentence as the number. If a figure depends on a period, a sample, a unit, or a scope, put all of it in one sentence rather than establishing context in a prior paragraph. Compression drops what is not adjacent. A statistic whose conditions live two sentences away is a statistic that will be restated without them.

Keep your URLs alive. Redirect anything you retire, and do not restructure paths without one. The stale-citation category is entirely self-inflicted, and it is the only one of these failure modes that a competent afternoon of work removes.

What you should not do is treat deep research visibility as a scaling problem to be won with more pages. The depth finding cuts the other way: the more thoroughly an agent researches your category, the less reliably it represents any individual source in it. Being cited more on that surface is not the same as being described correctly, and until the orchestration layer improves, the two will keep coming apart.