Use AI for Company Research, Then Check Three Sources

What Does a Model Get Wrong About a Company?

It gets specific, checkable facts wrong: a headcount that shrank in a layoff the model's training predates, a CEO who left, a product that was discontinued or renamed, a funding round that closed a year after the model's knowledge stopped updating.

None of this comes with a warning attached. A model answers a factual question about a company the same fluent way whether the fact is current or eighteen months stale, because fluency and accuracy are different properties and only one of them is guaranteed. This isn't a hypothetical failure mode confined to obscure companies, and it isn't a reason to avoid the tool either. There is measured evidence that a tool which helps on tasks inside its strength can produce confidently wrong answers on tasks that only look similar. In a 2023 field experiment with 758 consultants at one firm, those given GPT-4 on a task deliberately chosen to sit outside the model's range were about 19 percentage points less likely on average to reach the correct answer than the control group, 84.5 percent of whom got it right ¹. That was one task and one 2023 model, and it is not evidence that models make people worse in general. What it names is the shape of the risk, and company facts have that shape. They look like the kind of thing a model should know effortlessly, and sometimes it simply doesn't, with no visible seam between the parts it has right and the parts it doesn't.

The stakes here are asymmetric in a way worth naming plainly. Not knowing a fact costs you nothing in the room; the interviewer just tells you. Stating a wrong one confidently, in front of someone who knows it's wrong, costs more than silence would have, because it reads as a claim you didn't check rather than a gap you're honest about.

Check Three Sources Before You Say It Out Loud

Checking doesn't mean re-researching the whole company from scratch. Three places catch nearly everything a stale model answer gets wrong, and each one takes a few minutes, not an evening, because you're confirming a handful of specific claims rather than starting over:

• The posting and the company's own careers or about page. This is the one source guaranteed to reflect how the company describes itself right now, including team structure and current priorities.
• The most recent earnings call, funding announcement, or investor update. For a public or well-funded company, this is the fastest way to confirm headcount, leadership, and financial direction as of this quarter rather than as of whenever the model's training ended.
• A recent product changelog, release note, or press item. This catches the product-line and feature errors a model is most prone to, since these change constantly and quietly.

Reserve this checking for anything you actually intend to say in the room. You don't need to fact-check background reading you'll never repeat aloud, only the specific claims you're about to put your name behind in front of three people who work there.

A useful habit is separating the two categories as you research: context you're absorbing to sound generally informed, and specific claims you plan to state as fact. Only the second category needs the three-source check, which keeps the whole process to a few minutes rather than turning company research into a second job the night before an interview. Startups without earnings calls have a rough equivalent worth checking instead: a recent funding announcement, a founder's public post, or coverage from the last few months.

Ask About the One Thing You Actually Noticed

The target of a night of research isn't fifteen facts you can recite. It's one specific thing you noticed and can genuinely ask about, which is a much smaller, more durable outcome, and a much shorter list to hold onto walking into the room.

An interviewer can usually tell the difference between a list of recited facts and a real observation after a follow-up question or two, and the recited version is the one that reads as less prepared, not more, because it invites the obvious next question: why does that matter to you. Using a model to generate questions rather than facts works well for finding that one thing. Ask it what's changed about the company in the last two quarters based on what you paste in from the sources above, or what an outsider might be curious about given a specific recent announcement. At least one company already tells candidates this directly. Anthropic's candidate guidance, dated 10 July 2025 and one the company says it will revise, encourages applicants to use its assistant to research the company and prepare questions, while asking them to complete take-homes and live rounds without it unless told otherwise ². That is one employer's published rule rather than an industry norm, and it may change, but it treats prep and performance as different acts with different rules.

What you're building toward is the moment follow-up questions expose whether someone actually understands the answer they just gave: a genuine observation survives being pushed on, because there's a real reason behind it. A recited fact usually doesn't, because there wasn't one to begin with. The same logic applies to what good AI use in an interview actually looks like: the pattern that holds up is demanding a source for a claim and opening it, not accepting the first fluent answer a tool hands back.

This is also the difference that survives a nervous night before an interview when your prep time was short. Fifteen facts memorized under pressure are the first thing to blur together when someone asks a real question about them. One observation you actually noticed, checked, and can explain why it caught your attention, is much harder to lose your grip on, precisely because it isn't a list you're holding in short-term memory. It's something you actually thought about.

The shift is from research as a volume exercise, where more facts means better prepared, to research as a small verified set of things worth actually discussing. It takes less time, not more, once you stop trying to hold a loose list in your head at once.

See how it works >