Why Keyword Search Fails for Deal Sourcing

Every major deal sourcing database works roughly the same way. You type in some words or set some filters, and the system returns companies whose profiles contain those words or match those filters. It seems reasonable. It's also why you keep missing good companies.

The fundamental problem with keyword search

Keyword search finds what you literally typed. It doesn't find what you meant.

Consider a PE firm looking for "industrial automation software companies." A keyword search returns companies that use those exact words in their description. But the company you're actually looking for might describe itself as "workflow digitization for discrete manufacturers." Same thing. Completely different words. Keyword search misses it.

This problem compounds as your thesis gets more specific. The more nuanced your criteria — the type of customer, the specific workflow being addressed, the go-to-market motion — the more the best-fit companies slip through a keyword filter. They use their own language to describe themselves, not yours.

The other problem is that keyword search is symmetric. It treats "industrial automation software" as a set of words that either appear or don't. It has no concept of which companies are more or less relevant. You get a list of results sorted by some combination of recency and keyword frequency, not by how well they actually fit what you're looking for.

What filters make worse

Most database tools respond to this problem by adding more filters. Revenue range. Employee count. Geography. Funding stage. Investor names. The idea is that if you add enough constraints, you'll narrow the list to companies that fit.

The problem is that filters encode your thesis in a rigid structure that the data often can't support. A filter for "manufacturing software" requires every company in the database to be classified correctly under that category — which they aren't. Tagging is inconsistent. Companies that operate across categories get pigeonholed. Companies that use non-standard language to describe themselves get missed entirely.

Filters are also slow. Building the right filter set for a nuanced thesis takes time, and every search is effectively starting from scratch. There's no memory of what you've searched before, no learning from what turned out to be relevant.

What semantic search does differently

Semantic search converts your query into a mathematical representation — a vector — that encodes meaning rather than words. It then compares that vector against pre-computed vectors for every company in the database and returns the ones that are closest in meaning.

The company that describes itself as "workflow digitization for discrete manufacturers" has a vector that's close to the vector for "industrial automation software" even though the words don't overlap. Semantic search finds it. Keyword search doesn't.

This matters most at the edges of your thesis — the companies that are clearly relevant once you read their description but would never appear in a keyword or filter search. These are often the best opportunities because they're the ones competitors are missing too.

Tools like Radar are built on this foundation. Instead of a filter panel, you get a chat interface where you describe what you're looking for the way you'd describe it to a colleague. Radar converts that description into a semantic query and searches across millions of companies for the closest matches. You can add specific filters on top — geography, size, stage — but the core search is semantic, not keyword-based.

The compounding effect on personalization

Keyword search has another limitation: it's generic. The same search returns the same results for every firm. There's no concept of which results are relevant for you specifically, given your portfolio, your track record, and your thesis.

Semantic search opens the door to personalization because it operates in a continuous vector space where distance has meaning. A firm that invests in industrial IoT companies can have their results weighted toward companies that are closer to their existing portfolio in that vector space — not because of a filter, but because of learned proximity to what they've found relevant before.

Radar does this by analyzing a firm's existing portfolio and investment criteria and incorporating them into a hybrid score: part semantic similarity to the query, part fit to the firm's thesis. Two firms searching for the same thing see different orderings because their context is different.

What this means in practice

The companies that appear in a keyword search are the same ones appearing in every competitor's keyword search. They're the obvious targets. Everyone is calling them.

The companies that only appear in a semantic search are the ones whose language doesn't match the standard terminology. They're harder to find. That's also why they're often better opportunities — less picked-over, more receptive, earlier in the process.

If your sourcing process is built around keyword databases and filters, you're working from the same information as everyone else. The information advantage is zero. Semantic search is one of the few ways to change that.

Radar is built on semantic search from the ground up. Try it on your own thesis and see what keyword search has been missing.