Contextual signal that powers Criteo's brand-safe ad targeting
Problem
Building products to improve personalized ad experiences means turning high volumes of unstructured image data into signal. Criteo wanted to scale how it produces and improves that signal across multiple project teams — to identify consumer products with complex backgrounds and classify whether a web page is safe (age-appropriate).
Solution
Labelbox Annotate became the primary platform to produce the signal and connect Criteo's teams, cutting the back-and-forth needed to turn unstructured image data into training signal.
Result
Criteo's Publisher Content team immediately saw a 40% gain in signal delivery speed and comparable increases in signal quality.
Criteo's models classify web context and ensure brand safety. Labelbox's platform produced the expert-reviewed visual and contextual signal, with a 40% gain in delivery speed.
Note: This story is a recap based on a panel discussion at Labelbox Accelerate featuring Hong Noh, Senior Product Manager at Criteo.
The challenge
Criteo is one of the world's leading ad platforms, building contextual advertising that delivers personalized, brand-safe experiences. It uses AI and ML to make product recommendations, ensure brand safety on a page, and classify the context of websites across the open web so marketers reach relevant, reputation-safe sites. Its Publisher Content Analysis team hit a wall many ML teams face: it couldn't reliably improve signal quality and efficiency or surface the edge cases where models underperformed. The team managed unstructured data in Excel and defined label quality over endless internal email.
The approach
Criteo used Labelbox as a data engine to produce the signal its models needed, on core initiatives like background removal for better product identification and classifying whether a web page is safe (age-appropriate). The platform connected product and data science teams and cut the daily back-and-forth. For background removal, where a few pixels swing the result, expert human review in one dedicated platform moved model performance. And the platform's stability let globally distributed contributors produce signal securely, anytime.
The outcome
Criteo's Publisher Content Analysis team immediately saw a 40% gain in annotation delivery speed and comparable gains in signal quality. Next, the team is extending its ML products to fake news detection, contextual signal detection on video content, and tracking user attention (eye tracking).
Where this goes
Contextual understanding is a signal problem. Expert-reviewed signal is what lets an ad model tell a safe page from a harmful one — at the speed the open web demands.