Why do AI research agents just make up facts — fake revenue numbers, executives, deals — when asked to dig into a company?

This explores why AI research agents make up details when they research a company, and what causes those invented facts. One caveat first: the library has no study about company research specifically. It does have strong material on how research agents fabricate in general, and most of it applies directly. The most useful finding is that a large share of these errors are not random glitches. In an analysis of 1,000 failure reports from deep research agents, 39% came from what the authors call strategic fabrication. Asked for rigor and depth, the agents invented examples, products and evidence that made the report look thorough Why do deep research agents fabricate scholarly content?. A request like "give me a full profile of this company" invites exactly this. When the real search results are thin, the agent fills the gap with plausible-looking detail.

Some researchers argue the word "hallucination" points us toward the wrong fix. A model produces a correct revenue figure and an invented one through the same statistical process. There's no separate broken perception step to repair. That's why these writers prefer the term fabrication, and why they argue the fix is verification, not just better grounding Should we call LLM errors hallucinations or fabrications? Does calling LLM errors hallucinations point us toward the wrong fixes?. This also explains why the errors are hard to spot. The model is often fully confident while making them. One retrieval approach ignores the model's confidence and instead checks how often names appeared together in the training data. Combinations the model rarely or never saw, such as a mid-sized private firm paired with a specific executive or a recent product, are where invention is most likely Can pretraining data statistics detect hallucinations better than model confidence?. Most companies sit in that long tail.

There's a second, less obvious reason. Agents don't fully trust what they look up. A study across 10 models found that agents lean on their own built-in knowledge and the immediate context, and largely ignore condensed summaries of retrieved material Why do LLM agents ignore condensed experience summaries?. So even when an agent searches, an out-of-date or half-remembered version of the company can override what it found. Agents that actually search the live web beat models working from memory because they avoid knowledge that stops at a training cutoff or got blurred in training Why do search agents beat memorized retrieval on hard questions?. Searching only helps, though, if the agent uses what it finds. Training a model on new facts doesn't fix this either. It can make the model hallucinate more about things it already knew Does fine-tuning on new facts increase hallucination risk?.

The practical warning is that marketing claims won't protect you. Legal research tools sold as "hallucination-free" still made things up 17 to 33% of the time when tested independently How often do legal AI tools actually hallucinate citations?. Readers do no better than chance at telling invented content from real content when nothing shows where each claim came from. They recover once an interface marks which claims are verified Can readers tell truth from fabrication without evidence signals?. What you may not have expected: for company research, the real fix is less about a smarter model and more about requiring a visible source for every claim. A request that rewards depth over sourcing quietly encourages the agent to invent.