Ad systems need accurate, intermediate ground-truth signals because spending can move from profitable to unprofitable very quickly.
2
Ads and organic product systems should share measurement and signals instead of being built as separate systems.
3
Streaming only helps when the whole product is designed around the latency it needs, rather than when a batch job is simply replaced with a streaming job.
Summary
Andrew Yates describes the engineering and organizational problems behind ad platforms at companies such as Facebook and Pinterest. Ads differ from ordinary recommendations because the system spends someone else's money in real time, and a small change in efficiency can determine whether a marketplace is profitable. He argues that ad systems should share data and measurement with the organic product, while preserving strict definitions for impressions and clicks. He also explains why streaming systems matter for pacing and calibration, yet often fail when teams adopt them without a product strategy or clear latency target. Promoted.ai was built for large marketplaces that need an integrated measurement and machine learning layer. It collects marketplace signals, accepts scores from existing models, and models on top of them. Yates is direct about the limits of infrastructure alone. Versioning and feature stores can help, but leadership must resolve ownership, migration, and experimentation conflicts.
Ads became attractive because the largest technology companies depend on them
Andrew Yates says he entered ad tech after recognizing that Google, Facebook, Amazon, Microsoft, and Apple all have major advertising businesses. He had worked at Facebook Ads and Pinterest before starting promoted.ai. He also saw a gap in the startup market. Facebook and Google had been so dominant that few people were building ambitious ad technology companies. A conversation at Pinterest reinforced that view, since there were no ad tech products that met Pinterest's standards for use or acquisition. Yates treats advertising as the business that pays the bills for many technology companies, even though engineers often avoid identifying with ad tech.
An honest ad system needs one organizational goal and strict measurement
Yates says promoted.ai is built around the idea of an honest ad system. The goal is to make profits for all sides while keeping the system fair, rather than changing attribution or measurement to pull revenue into the current quarter. Click prediction is already established technology. The harder problem is getting the whole organization to optimize for one goal instead of keeping an extra lever available whenever the core business needs a short-term revenue boost. Clear definitions for impressions and clicks matter because consistent measurement produces better training data and more accurate models.
Advertising has a profitability threshold that makes efficiency unusually valuable
Yates describes how machine learning changed mobile advertising around 2010. Facebook and Google used their own data for click and conversion prediction, and Facebook's mobile app install ads became an early breakout example. Ads behave differently from a recommendation system where a small quality improvement usually produces a small gain. If an ad campaign remains unprofitable, it should not be purchased. Once it becomes sufficiently efficient to be profitable, the marketplace can buy a very large amount of it. That threshold makes modest improvements in ad efficiency financially significant.
Ad and organic systems lose value when they are built separately
Large companies often create a separate ads team that owns event collection, measurement, machine learning, and delivery. An orchestration layer later combines ad results with organic results. Yates argues that the two systems should share signals and measurement. An organic listing may appear immediately below an ad, yet the teams may not coordinate whether the user would have seen that listing anyway. The same issue affects pricing. An advertiser should pay less, or nothing, to win a position that the organic system would already have awarded. Combining the systems can also make use of the full app's data when training ad models.
Real-time systems matter because ads spend money while the system is running
Yates explains that batch systems can work for some jobs, and Facebook used daily batch training when he worked there. Ads add a different risk because they spend other people's money in real time. A model that is wrong for two or three days may continue spending an advertiser's budget in a bad way. Facebook therefore used real-time calibration and pacing systems alongside batch training. Streaming becomes more important when the product needs to adjust budgets, predictions, or delivery during the day. The organic side can benefit from streaming too, but the financial consequences make the ad case more urgent.
Signal ownership and model versioning are organizational problems
The teams consuming a signal need to know who owns its definition and what happens when it changes. Yates says this requires versioning, migration periods, and enough budget to run old and new versions in parallel. Feature stores and machine learning platforms can automate parts of the process, but they do not resolve disagreements between teams. He thinks strong technical leadership must decide when to accept mixed or ambiguous experiment results and continue toward a coherent architecture. Forcing every shared service through a perfectly clean A/B test can slow the organization until teams bypass the process. Technology can support signal management, but it cannot settle ownership and tradeoffs by itself.
Promoted.ai adds a measurement layer around existing marketplace models
Promoted.ai targets large marketplaces and focuses first on proving that its models produce more revenue. The company built its own streaming data and measurement infrastructure, collects marketplace events, and joins signals in real time. Customers can send user scores, item scores, embeddings, relevance scores, or other model outputs. Promoted.ai does not need to host those models. It can ingest their results, model on top of them, and report how much each signal contributed. Yates says the system can also show where existing models missed useful information. This lets promoted.ai work with an in-house data science team instead of replacing it.
Streaming needs a product latency target and a management commitment
Yates names two common streaming failures: a lack of commitment and a streaming system that is so slow that it behaves like batch. Replacing a Hive job with a Flink job does not automatically create a real-time product. Teams need to decide how quickly the product must react, then carry the signal through the full pipeline within that limit. If an event waits three hours before reaching the recommender, streaming may have little value. A real-time product requires strategy, infrastructure, and leadership that are designed around immediate feedback. It cannot be added as a small upgrade and judged only by a standard A/B test.
"I don't think this can be purely solved from a technical and infrastructure standpoint in terms of feature stores and Signal Management and versioning."Andrew Yates27:05
Who should watch
You are building ads or monetization for a marketplace and need to understand why measurement quality can affect the company's profitability.
Your organization is separating ad ranking from search and organic ranking, and you want concrete reasons to share data and measurement.
You are considering streaming ML infrastructure and need to set a real product latency target before replacing batch jobs.
