01 / Platforms · Publisher Pistachio

Rebuilding targeting around how publishers sell

Publishers had to describe their inventory placement by placement, which meant maintaining thousands of configurations to express something simple. The cheaper option was to leave it alone and let them absorb the work. I argued for rebuilding the targeting model instead.

The situation

Configurations in Prebid Server Premium, Microsoft’s server-side header bidding platform for publishers, determine what inventory a publisher sends to which demand partners. Historically that was expressed at the placement level, so a publisher describing their full offering had to create and maintain thousands of individual configurations. For larger publishers it reached into the tens of thousands.

Customers were direct with me about how prohibitive this overhead was. They had given the same feedback before, and the design had not changed. There were other friction points, but this was the piece that had to be solved. The platform worked, which made rebuilding it easy to keep deferring.

Constraints I was handed

• The existing model carried real publisher revenue, so nothing could break for anyone already using it.
• The only alternative was making no investment and leaving first-party and third-party publishers to absorb the friction.
• Ownership of the core targeting logic was unsettled between engineering teams; engineering leadership had to resolve it before work could be assigned.

The call I made

The expected option was to leave it. Prebid Server Premium worked, rudimentary bulk tools existed, and the pressure was coming from customers rather than an internal metric. I argued for rebuilding the model so publishers could express intent at the level that matched how they sell—from run of site to an individual placement, with geographic, device, segment, and key-value targeting alongside it. That meant asking multiple engineering teams to re-architect a live system over more than a year on the strength of customer complaints.

Time to ship

Rejected · Leave the model aloneNothing to ship; the friction continues

Chosen · Rebuild the targeting modelOver a year from scoping to general availability

What it fixes

RejectedNothing structurally; the bulk tools reduce the symptom slightly

ChosenLets publishers express intent at the level that matches how they sell, and unlocks better inventory-to-demand matching

Cost of being wrong

RejectedPublishers keep absorbing overhead we already knew was prohibitive

ChosenA long investment, plus migrating every existing publisher off a live model without disrupting revenue

I made the case with three forms of evidence: customer feedback established that the problem was real and specific; competitor comparison showed that our model was the outlier; and internal configuration, revenue, and usage data established the scale. Each was arguable on its own. Together they were hard to set aside.

What shipped

Publishers now express targeting at the granularity that matches how they sell, with geographic, device, segment, and key-value dimensions alongside it. Active configurations per publisher fell by about a third from before the work began to two months after general availability, while revenue grew. We rolled out first to API publishers to prove architecture, reliability, and performance; then to a deliberately vocal group using the interface; then to general availability. Both models ran in parallel until remaining customers were migrated by script and the old architecture retired.

The configuration count is the visible proxy. The substance is that publishers can now say what they mean once, instead of restating it placement by placement.

33%

fewer active configurations per publisher two months after general availability, while revenue grew through the transition.

What I would redo

At general availability, I wanted both models to coexist for a while so publishers could move at their own pace. The team working directly with customers quickly reported that having both options visible caused confusion and churn: what looked like optionality to me looked like ambiguity to a publisher. We had staged the technical rollout carefully, but not the customer transition; the gap was communication and education, not software. I reversed the decision, built migration and communications plans with engineering and services, and moved everyone.

“Treat the customer experience of a transition as part of the design.”