LinkedIn can see who you know. It cannot see who you actually talk to.

You have two thousand connections and about eleven relationships. Both look identical from the outside.

Scroll your list and try to sort it honestly. The university friend you have not spoken to since a wedding. The recruiter who added you in 2019. The person you worked beside every day for three years and still message most weeks. On the screen they are the same size, the same blue, the same first-degree label. Only you know which of them would pick up the phone.

That is fine until you need to borrow somebody else’s knowledge of the same thing. You ask a colleague for an introduction to a buyer. The platform says they are connected. The colleague says “I think we met at a conference once?” and the introduction dies in a draft.

It is worth being current about LinkedIn before saying what it cannot do, because it is shipping quickly. It launched AI people search in November 2025. It sells AI hiring software, and AI selling software inside Sales Navigator, and its agentic hiring line is reported at roughly $450M annualised. Open to Work is no longer a bare switch either: it is a structured panel carrying wanted job titles, location types, locations, start date and employment types, with three-way visibility including a recruiters-only mode, and in India it also collects notice period and expected salary. Anyone who tells you LinkedIn does not capture what people want has not looked recently.

We know, because an earlier draft of our own thinking said exactly that, and it did not survive a check.

The gap that does survive is narrower and more useful. TeamLink will tell you that a colleague is connected to your target. It cannot tell you how well they know them, when they last spoke, or whether an introduction from that person would land. That is not a missing feature waiting on a roadmap. It has never seen a single message, so there is nothing for it to know that from. It reads the connection graph, not the correspondence.

A connection is a claim about the past. A live thread is evidence about now.

The distinction matters because the two graphs were made in different ways. The graph everyone competes on is the one people curated for an audience: who they chose to accept, what they chose to list, which title they were using that year. The useful graph is the one that fell out of two years of actually replying to each other. Nobody performed it, which is exactly why it is hard to fake.

In Deals On, path strength is a real term in the ranking. It is computed from correspondence rather than from a connection count, and it is computed from headers only: who wrote to whom, and when. Never message bodies. The agent reads who you correspond with, and it does not read what you said.

That deserves more than one line, so it has its own post. The short version: the Gmail permission we ask for is gmail.metadata, which is narrower than read-only access and structurally cannot return a message body.

What changes, once path strength is in the ranking, is the order of your day. Somebody two warm hops away outranks a stranger who looks better on paper. The stranger has the better title. The two-hop contact has a route in, and a route is the part you cannot manufacture on a Tuesday afternoon.

There is a second effect, quieter but worth having. Knowing the path in advance changes how the ask gets written. A note travelling through a former colleague reads differently from one travelling through an investor, and the difference is not decoration: it is which fact leads, and how much context has to be restated. When the route is known early, the first draft can be written for the route.

Keep your two thousand connections. As an address book it is genuinely useful, and nothing here suggests deleting it.

Just do not mistake it for the map. The eleven relationships were always the asset. The difference is whether anything can see them.