Insights

Lost in Citation: Why Research Agents Claim More Than They Know

By Will Hardman · · 4 min read

More on the theme of trusting research agents.

Last week I discussed content credibility: how can you make sure your agents are only sourcing reliable documents (a particular problem when they’re searching the web)?

A closely related topic is citation support: once they have a document, do they actually pull information from it correctly?

Most research agents provide citations, noting inline within their responses where they have relied on different sources. Very few of us ever bother to check up on the cited sources but if we did, we’d discover a problem: not infrequently, the source will not fully support the claim the agent wants to make.

Here’s a measured example: on a clinical QA benchmark (Zhang et al., 2026), 15% of the claims made by Claude Opus 5 on the basis of retrieved evidence were not fully supported by said evidence. That’s actually only the tip of the iceberg, however, because a further 48% of the claims relied on evidence that came from elsewhere in the sources (i.e. not from the regions that Opus said it had cited).

This matters because if a human checks the cited fragment of the source, they’ll conclude that the claim is fabricated - when it is merely mis-attributed within the source. If you’ve ever run an evaluation exercise for a research agent, you’ll know this is a problem: reviewers will penalise the model for fabrication when they should not.

The key is to realise why this happens: given rich context, models will tend to write the most plausible answer that conforms with all the retrieved information - modulated only by any observations made during the reasoning process. Combining blended sources with a propensity for sycophancy and confirmation bias can often steer the model’s claims beyond that which the sources actually license.

There’s actually a bunch of other citation-related failure modes in research agents - for example redrafting answers often breaks citations or some claims aren’t cited when they should be - but to keep this article short, we’ll focus on citation support, which is arguably the most impactful of the lot.

Aside: the astute reader might notice I’ve skipped past the most widely known citation-related problem: fabrication. This is where the model simply makes up the source. I’ve skipped this because it’s easily solved: simply have the research agent place all of its sources into a temporary file-store then check during a review step that every citation exists in that store.

My Solution

So, how do you ensure that your agent makes only well-supported citations? Here’s my recipe:

  1. When sources are retrieved, perform simple parsing, chunking, tagging and cleaning then place them into a repository. Do not allow any interpretation of the source to occur at this stage.
  2. Before your agent starts to compile a response, have it develop a skeleton of the sections and list the questions that need answering within each section. The key insight here is that binding questions to the response structure before the full response is written reduces the room for bias when hunting for the evidence.
  3. Place a “blind extractor” tool between the agent and the sources. This tool should accept a question from the writer and should then compile a list of fully-supported facts - one fact per sentence and one citation per fact - for the response. The facts are returned to the writer. The point here is that the extractor is a “disinterested intermediary”. It is not hunting for evidence that conforms with a prior belief. Two points to make this work well: You just need to ensure that the questions it accepts are semi-open (”What do the sources say about X?”); questions implying a prior (”Find me evidence that X causes Y”) should not be allowed. You should allow “Not addressed by the sources” to be the default answer returned by the extractor.
  4. The writer then compiles its response using only the licensed content and it must provide the citation for each fact it uses.
  5. Finally, don’t forget the revision step - a simple check that each claim is fully supported by the extracted fact. The presence of the extracted facts makes this a simple and reliable process step. (Contrast with the other way of doing this: to hand an LLM judge the citation and the source and asking “Does the source fully support the citation?”. This is a far less accurate process.)
  6. Return the report, citations and the facts that were tied to each citation.

So there you go - a citation support subsystem that should not only substantially boost the quality of research agent results but should also - thanks to the attached facts - make it far easier for a human to review.

Ready to move from pilot to production?

Book a free 30-minute consultation. You'll leave with a straightforward answer on where agentic AI can work for your organisation, and the first steps to get there.

Book a free 30-min consultation →