The cold start problem nobody warned me about

Two years in, I can report that everything written about the marketplace cold start problem is true and none of it prepared me for the actual experience of it.

The literature describes a chicken-and-egg standoff: no buyers because no sellers, no sellers because no buyers, and your job is to break the loop on one side. Fine. Accurate. What it doesn't tell you is that once you start breaking the loop by hand, you lose the ability to tell whether it's still a loop.

The measurement problem#

We have listings. We have brands. Deals happen. Money moves. From outside — and increasingly from inside — Exposely looks like a working marketplace.

But I'm in the middle of a large fraction of those deals. Not visibly. I'm emailing a creator to say a brand is looking for someone in their lane, I'm nudging a brand toward three profiles they'd never have surfaced on their own, I'm unsticking a negotiation that stalled. All of it in the background, none of it in the product.

So when I look at the numbers I cannot answer the only question that matters: how much of this happens if I stop?

I genuinely don't know. I've tried to estimate it and the estimate moves depending on my mood that week. That's not a metrics problem I can solve with better instrumentation, because the intervention isn't logged anywhere — it's in my sent folder.

What I should have built first#

The thing I'd tell a version of me starting this in 2013: instrument the manual work from day one. Every time you touch a deal, record it. Not for the customer, for you. Build the ability to say "this month, 40% of GMV was assisted, down from 55%" before you need it, because by the time you need it you'll have two years of history you can't reconstruct.

I have no such history. I have an inbox and a feeling.

The other thing nobody says#

Doing it by hand doesn't just obscure your metrics. It obscures your product.

Every time I match two parties myself, I'm patching over a specific failure in the software — bad search, missing filters, a discovery surface that doesn't surface the right people. The patch feels like customer service and it's actually a bug report I'm swallowing instead of filing. The product doesn't improve, because the pain that would have driven the improvement got absorbed by me being helpful.

That's the trap. It isn't that manual work doesn't scale. Everyone knows manual work doesn't scale, that's the least interesting version of this. It's that manual work is an anesthetic. It removes the signal that would tell you what to build.

What I'm changing#

Three things, and I'm writing them down so I can be held to them.

Log every assist. A spreadsheet is fine. I need the ratio more than I need it to be elegant.

Every assist gets converted into either a product change or a decision not to make one. If I had to introduce a brand to a creator they should have found in search, that's a search bug and it goes on a list. If I decide search will never solve it, that's a real answer too, but it has to be said out loud.

Pick one segment and stop assisting it entirely. Not the whole market — that's a suicide test. One category, no help from me for a quarter, and I find out what the software actually does when I'm not standing behind it.

I expect the answer to be worse than I'd like. That's the point of running it. Two years of not knowing has been comfortable in a way that I've started to distrust.