How to build a contextual targeting product
From a taxonomy to a signal buyers trust, without user data
1 min read
TL;DR
- Pick one buyer and one job before you pick a taxonomy.
- Classify pages with a model, but price on outcomes, not on labels.
- Ship a small set of segments that beat cookies on one metric.
- Sell it as a test with a clear success bar, then expand.
The problem
Buyers lost the user signal. Cookies are going, device IDs are capped, and AI answers keep people off the page. What is left is the page itself: what it says, and what the reader wants while reading it.
A contextual product turns that page into a signal a buyer can bid on. The publisher gets a higher CPM. The buyer gets reach without user data. You get a product that works in every privacy regime.
The VP calls
The call
Who is the first buyer?
Agencies · Brands direct · DSP partners
Agencies. They test fast and talk to each other.
The call
Build or license the taxonomy?
IAB standard · Custom · Hybrid
Start with IAB so buyers can map it. Add custom segments only where you win.
The call
How do we price it?
CPM uplift · Flat fee · Revenue share
CPM uplift on the first deals. It ties our price to their result.
The build
Freeze the taxonomy
Owner: PM + Data science2 weeks
Pick the segments buyers already ask for. Write one sentence per segment that a human labeler can follow.
Build the eval set
Owner: Data science3 weeks
Label 2,000 pages by hand. This is the bar every model change has to beat.
Classify at serve time
Owner: Engineering6 weeks
Cache labels per URL, refresh on change, and return a segment in under 20 ms.
Pass the signal to the auction
Owner: Engineering + Partnerships4 weeks
Send segments in the bid request so every DSP can read them. Start with deal IDs.
Text version
- Page: URL + text
- Classifier: model, cached
- Segment: IAB + custom
- Bid request: deal ID
- Winning ad: CPM uplift
Mistakes to avoid
- Shipping 400 segments when buyers use 20.
- Measuring label accuracy and calling it success. Buyers care about outcomes.
- Selling before you have one case study.
Go to market
Package it as a test: four weeks, one brand, one metric. Price the test low and the renewal on uplift. Write the case study before the test ends, then fill in the numbers.
Checklist
0/5Next episode
How to build an AI agent that takes actions for users
Scope, evals and guardrails for an agent people let touch real money