Definition
A gateway records only the traffic that is routed through it. Every tool that calls providers directly is dark traffic: its spend, prompts, and failures leave nothing in the organization's shared record, even when fragments survive in provider consoles or the tool's own logs. Coverage, the share of traffic routed through, is the ceiling on which questions the record can answer.
The first three pages assumed the specimen request went through the gateway. This last one asks the quieter question: what about the requests that don’t? A gateway is not a camera on every wall. It is a doorway with a guest book, and it knows exactly nothing about the side doors. This page is about how side doors appear, what they cost, and the one thing that closes them.
How dark traffic accumulates without anyone deciding
The most important thing about dark traffic (model traffic that goes straight from a tool to a provider, never crossing the gateway) is how it comes to exist: by drift rather than decision. Scrub the six months below. The first tools arrive after the gateway exists and get pointed at it. Then a script gets written with a provider key in an environment variable, because that was the fastest way to make it work. Then an extension shows up with its own endpoint baked in. No meeting approved any of this.
month 1 · The platform team sets up the gateway and points the coding agent at it. Coverage: everything.
By month six, half the organization’s tools are dark and coverage has drifted to 50%, one perfectly reasonable shortcut at a time. This is the normal trajectory, and it is why coverage is a thing to be maintained rather than assumed.
What a gateway can and cannot see: routing each tool through or around it
Now take control of it. Below are the four tools with their routes drawn. Solid dots cross the gateway and feed the record; hollow dots go direct, and notice that they still reach the provider and still get served. Hollow means unrecorded while the request still works, which is exactly what makes dark traffic so easy to live with until the day it isn’t. Hover a tool to hear how it ended up on its route; tap it to re-route it, and watch the questions grid react.
0 of 4 questions still answerable. The hollow dots are dark traffic: served fine, recorded nowhere the organization can query. Micro-challenge: find the single tool that turns them all back to yes.
The questions grid is the honest scoreboard. Capture writes the guest book as traffic crosses the doorway, and coverage (the share of traffic that crosses) is the ceiling on every answer computed from it: spend-by-team is really spend-by-covered-team, and “did customer data leave?” is really “did it leave through the front door?” Take the micro-challenge in the caption; the tool you find is the nightly script, and the question it darkens is precisely the kind that matters at the worst possible time.
Why a dashboard cannot fix a coverage gap
When a question comes back “unknown,” the instinctive fix is better reporting: a dashboard, a chart, a weekly email. Here is that instinct, as a button. The 3am question is on the table. Press “add a dashboard” as many times as you like.
The question on the table: "What did the 3am job ask the model to do?" unknown
no rows between 03:00 and 03:20. The script's traffic never crossed the gateway, so its story was never written here.
each one is a prettier view over the same rows
Two buttons. One adds a view over the record; one adds rows to it. Find out which one answers the question.
The dashboards render beautifully over zero rows. A dashboard is a view over the record; it cannot conjure rows that were never written. The other button works because it is a routing change: the moment the script’s model traffic crosses the doorway, its requests and replies get written, and the question collapses into a query. That is the closing thought for the whole series: the gateway’s powers (the record, the routing, the policy, the answers) all flow from one unglamorous fact about topology. The traffic goes through it. Everything else is what you build on that.
For the reference treatment of the metrics and records, see AI gateway observability; for what the captured record contains, see AI session capture.