Why General-Purpose LLMs Fall Short at Company Discovery

The companies that could change your next deal won’t show up in ChatGPT or Claude. FounderNest was built to find them first. TL;DR General-purpose large language models like ChatGPT and Claude are optimized to return the most probable answer. Company discovery for M&A, corporate development, and competitive intelligence requires the opposite: the improbable-but-true companies that aren’t in the consensus. This is a structural mismatch, not a prompting problem, and it doesn’t close as the models improve. Nonetheless, effective company discovery requires a purpose-built engine sitting on top of proprietary, structured data: a queryable graph of companies, patents, papers & articles, and clinical trials, with filtering, ongoing monitoring, and decision-ready output. Click here if you want to know more about how 100+ Fortune 500 teams use FounderNest. Company discovery is a recall problem, not a chat problem There is a quiet assumption spreading through strategy teams: that a general-purpose chatbot is now good enough to find companies in a market. Open ChatGPT or Claude, describe the space and get a list. It does feel productive For most knowledge work, that instinct is correct. A general-purpose model is extraordinary at summarizing a document, drafting an email, or explaining a concept, tasks where you want the single most likely, most reasonable answer. Company discovery is a different job. The entire value of sourcing (in mergers and acquisitions, in corporate development, in competitive intelligence) lives in the companies you couldn’t have named yourself. The early-stage player two funding rounds before it becomes obvious. The competitor that doesn’t show up in the trade press. The acquisition target nobody else has mapped yet. The output from LLMs was not so different from what we already knew internally; it was like revalidation of what we already knew. Business development director, European specialty ingredients company In information-retrieval terms, this is a recall problem: the goal is comprehensive coverage of a space, including its long tail. A general-purpose LLM is optimized for precision on the most probable answer, exactly the wrong objective when the names that matter are, by definition, the unlikely ones. A better-written prompt does not fix this. It is a mismatch between what the tool is built to optimize and what the task requires General-purpose LLMs are optimized for the most probable answer not the most complete one. The real signal lives in the long tail: early-stage, under-covered, specialized companies that don’t rank high on the public web. Why does ChatGPT return only the obvious companies? Ask a general-purpose model to list the companies in almost any space and the first names back are the ones a domain expert could have written from memory. Ask for humanoid robotics and you get Tesla, Boston Dynamics, Figure, Agility, Xiaomi. All correct. All consensus. The top five biggest companies are always the same: the super large corporates. Finding them doesn’t tell me anything. Only in FounderNest do I find the next five competitors, and they usually fit very well to my needs. Business development director, European specialty ingredients company This happens because the model generates from the highest-probability region of its training distribution: the names mentioned most often across the public web. Those are, almost by definition, the well-known players. The companies that change a strategic decision are rarely the most-mentioned ones; they are the under-covered, recently founded, or narrowly specialized firms that haven’t yet accumulated a large public footprint. When the obvious answer is the complete answer, you didn’t need a tool. When it isn’t (which is most of the time in serious sourcing), a consensus-seeking model is structurally unable to reach the names that matter. Why does asking for “more companies” return repeats? A common workaround is to ask the model for more: give me more, give me more. In practice, the returns diminish fast, and often invert. Each additional request asks the model to keep generating from the same pool of high-probability names. So the more you push, the more it repeats names it has already given you, and the fewer genuinely new companies you discover. In one test of a niche therapeutic area, five follow-up requests produced 17 companies, of which six were duplicates, leaving only 11 unique. The curve bends the wrong way: effort goes up, new information goes down. A purpose-built discovery engine should do the opposite: the deeper you dig, the further into the long tail it reaches, surfacing more signal, not the same names again. Ask ChatGPT for 5 batches of companies and you’ll get 17 results — but only 11 are unique. Each extra request just recycles the same pool of familiar names. More effort, same blind spots. Click here if you want to know more about how 100+ Fortune 500 teams use FounderNest. Why do general-purpose LLMs miss relevant competitors? The highest-stakes failure is the one that looks like success. Ask a general-purpose model for companies similar to a specific startup and it returns a clean, confident, well-organized list that quietly omits several of the most relevant competitors. In casual use, an incomplete list is a minor annoyance. In M&A and competitive strategy, it is a wrong decision waiting to happen. If a competitive set is missing the three companies that matter most, every downstream conclusion rests on a false map: the moat is misjudged, the deal is mispriced, the wrong initiative is greenlit. And the gap is invisible, because the list looked complete. There is no error message for “the company that would have changed your mind isn’t here.” But six months later, your VP hears about that missing company, asks you about it, and your entire world falls apart. We are evaluated and measured on number of deals that we can source and bring in. This is where FounderNest comes into play and helps us find those companies before our competitors. Business development director, European specialty ingredients company Confidence is not coverage. A fluent, well-formatted answer with a silent gap in it is more dangerous than no answer at all, because it invites trust it hasn’t The post Why General-Purpose LLMs Fall Short at Company Discovery appeared first on FounderNest .