---
title: "Deposits, Not Impressions, Running an Affiliate Desk Like an Engineer"
description: "A case study in affiliate and creator partnerships: the \"verified streamer\" proposition, audience verification as a precondition for every deal, deposit-level attribution and reporting, and the tooling the desk built for itself when nothing off the shelf would do."
url: https://www.afonsomartins.com/affiliate-desk
author: Afonso Martins
published: 2026-07-12
updated: 2026-09-18
---

# Deposits, Not Impressions

A case study in running a creator-marketing desk like an engineer: how I managed an influencer media-buying agency, made audience verification a precondition for every deal, measured everything in verified deposits, and built the tooling the desk needed.

**Status:** The agency (SkinBet Agency) operated from 2024 to 2026, until I moved to building [21Core](https://www.afonsomartins.com/how-i-built-21core) full time. Client names and commercial terms are withheld.

## The proposition: the "verified streamer"

Performance creator marketing has a trust problem in both directions. Brands get burned by creators with manufactured audiences; honest creators get lumped in with the fraud and squeezed on terms.

The agency's answer was a single promise: **every creator we put in front of a brand has passed a fraud check, and every deal is reported in verified deposits, not reach.** Impressions are an input someone else can fake; a deposit is an outcome on the brand's own books.

That one constraint shaped the entire desk.

## How the desk ran

**Sourcing.** I come from the player side of this scene, CS2 native, fluent in where its creators actually live (Twitch, Kick, YouTube, Telegram, Discord). Sourcing wasn't a list purchase; it was knowing which channels players actually trust, and why.

**Vetting.** Every candidate channel went through the fraud pipeline I built for exactly this purpose: a 0 to 100 score over follower-ratio, engagement and growth-velocity signals, screening tens of thousands of profiles per session. Clear fails never reached a brand. The full story of that system is its own case study: [Scoring 75,000 profiles a session](https://www.afonsomartins.com/fraud-scoring).

**Budgets.** I managed the monthly spend across partners, with a single partner running at €3k a month.

**Deal structure.** CPA, revenue share, or hybrid, chosen per creator based on what the vetting data supported. A creator whose audience quality was proven could carry performance terms; unproven channels didn't get funded on hope.

**Measurement.** Deals were tracked to deposit-level outcomes with promo-code and link attribution. Reports to brands showed the arithmetic (spend on one side, verified player value on the other) so a campaign decision never rested on a screenshot of a view counter.

## What running the desk taught me

- **The only honest metric is the one the counterparty can't manufacture.** Everything upstream of the brand's own ledger is negotiable fiction.
- **Verification changes the negotiation.** When you can show a creator their audience is real (or isn't), pricing conversations get shorter and more honest on both sides.
- **Fraud is a moving target.** The vetting rules that worked in month one were partially gamed by month six; re-validation had to be routine, not incident response.

## The tooling behind the desk

The desk's tooling outgrew the desk:

- **StreamRadar Pro**, the vetting pipeline, built out as internal tooling for the desk rather than sold on.
- **The decision layer around it:** creator scores, ROI-ranked deal estimates and GEO-confidence checks. Weak evidence was marked as weak instead of being turned into a confident recommendation.

## What this demonstrates

- **The full affiliate cycle, owned personally:** sourcing, vetting, negotiation, campaign management, attribution, reporting.
- **A measurable trust edge:** verification-before-spend as the product, not a slogan.
- **Builder range:** when the desk needed tooling that did not exist, I built it and used it in the same commercial process I managed.

---

*Related: [Scoring 75,000 profiles a session: the fraud engine behind the desk](https://www.afonsomartins.com/fraud-scoring) · [How I built 21Core around prediction markets](https://www.afonsomartins.com/how-i-built-21core)*