From Intuition to Precision: How Coinbase Built a General-Purpose Targeting Engine
TL;DR: The future of growth is precision. For every communication we send, from educational content to new asset launches, there is an ideal target audience.
We built Smart Targeting to move beyond using manual segmentation to build target audiences towards automated, intelligent discovery. This platform dynamically identifies high-propensity users to target by learning from small, initial exploration tests and then expanding to look-alike audiences who mirror the traits of those early responders. Applying Smart Targeting to campaigns enables us to rapidly iterate and deliver content to the users that are most likely to engage, without building bespoke ML models.
The Challenge: The "Long Tail" of Relevance
At Coinbase, we rely on machine learning to power targeting for our most critical, always-on notifications: price alerts, price milestones, etc. For these core inventory items, investing in bespoke, highly tuned models makes sense.
However, a significant portion of our user communication lives in the "long tail." These are ad-hoc or temporary campaigns: a new asset launch, a seasonal promotion, or an educational series on a specific Web3 topic.
For these campaigns, we have historically faced a resource dilemma:
Bespoke models are too expensive: We cannot spin up a dedicated engineering team to build and train a new model for a campaign that might only run for a week. Furthermore, new campaigns suffer from the "cold start" problem – we simply don't have the historical training data to build a model before launch.
Finding the right audience is inefficient: Without a model, campaign managers rely on heuristics to determine the target audience (e.g., "Select users active in the last 7 days who hold >$100"). Finding the right heuristic requires running multiple small-scale experiments to tune the logic, requiring time and effort while rarely achieving optimal targeting.
We needed a middle ground: a system that could offer the precision of machine learning without the high overhead of bespoke development.
The Solution: Smart Targeting
We built Smart Targeting to solve the "cold start" problem for the long tail, serving as a general-purpose platform that allows us to launch a campaign without a pre-existing model and rapidly converge on the best audience.
The Engine: Universal User Representations
We convert holistic user activity into dense vector embeddings, where users with similar behaviors naturally cluster together in the data space.
The foundation of the system is our General-Purpose User Embeddings. Rather than engineering specific features for every possible campaign type (which doesn't scale), we built a single, robust representation of every user based on their holistic history.
We ingest a user’s interactions with the app – navigation, page views, and activity patterns – and compress this high-dimensional data into a dense vector. This embedding acts as a behavioral fingerprint. It captures nuanced similarities between users (e.g., users who explore similar assets or engage with similar content) without requiring us to explicitly define those relationships.
Solving Cold Start: Explore, Learn, Expand
To solve the cold start problem, an automated feedback loop learns from a small initial test group and expands targeting to behaviorally similar "look-alike" users.
The true power of Smart Targeting is how it operationalizes these embeddings. Since we rarely have training data at the start of a campaign, the system uses an automated feedback loop:
Exploration: The campaign begins by targeting a small, randomized set of eligible users.
Signal Capture: As users interact (or don't interact) with the notification, the system captures this ground-truth data as labels.
Expansion: The model identifies the embeddings of the engaged users and finds similar users in the vector space – the "look-alikes."
Scaling: The campaign scales up, targeting these high-propensity users who are behaviorally similar to the early adopters.
While the default optimization target is engagement (clicks), the platform is flexible. For complex campaigns, managers can provide custom SQL queries to define specific success metrics (e.g., "users who viewed the asset page" or "users who completed a trade"), allowing the model to optimize for more than just engagement.
Architecture and Integration
Smart Targeting is built as a decoupled service that orchestrates the end-to-end workflow, connecting UI configuration with heavy compute and downstream delivery platforms.
We designed Smart Targeting as a decoupled service that integrates seamlessly into our existing notifications stack.
The workflow operates in distinct stages:
Configuration: A campaign is configured in our internal campaign management tool. The manager selects the "Smart Targeting" strategy and defines the budget constraints.
Orchestration: Our service handles the job lifecycle. It coordinates the fetching of labeled training data, computation of user scores, managing the dependencies between data availability and model execution.
Scoring: The compute layer calculates distances between user embeddings and the campaign's target profile, generating a scored list of eligible users.
Consumption: Downstream services that target and deliver content to users query the Smart Targeting service to check the status of a user. Before sending a message or rendering a banner, the delivery system asks: "Is User X targeted for Campaign Y?"
This separation of concerns allows us to update the underlying embedding models or scoring logic without requiring changes to the campaign management UI or the delivery infrastructure.
The Impact
By shifting from manual heuristics to automated discovery for targeting audiences, we have seen a marked increase in campaign efficiency. We are effectively trading volume for relevance – sending fewer messages, but sending them to the right people.
Significantly Higher Engagement: We have observed increases ranging from 41 to 92% in Click-Through Rate (CTR) when running various A/B tests comparing heuristic-based targeting vs. smart targeting across various channels (push and email). Since moving all our lifecycle email campaigns to Smart Targeting, we have seen an overall 30% increase in CTR and 30% reduction in user unsubscribe rates across the entire program.
Reduced Noise: Because the model identifies the highest-propensity users, we can achieve our campaign goals while messaging a smaller fraction of the user base, reducing the risk of user fatigue.
Operational Velocity: Teams no longer spend weeks tuning segmentation logic. They can launch a campaign with a broad hypothesis and let the system discover the ideal audience automatically.
Challenges & Opportunities
Automating ML intelligence for self-serve “no code” tools and non-technical users comes with unique challenges from bespoke domain-specific models:
Audience Sizing: Automatically determining the right volume reduction (or minimum score threshold) targets for every campaign from a cold start, balancing positive impact (e.g., engagement or clicks) against negative impact (e.g., unsubscribes).
Label Selection: There is no “one size fits all” label to identify engaged and similar users. For some campaigns, the optimal label may be as simple as a click or view, but for others, compound labels or further downfunnel labels may yield better results.
Application Relevance: Campaigns with continually refreshed engagement data such as evergreen CRM lifecycle emails fit this model well, while one-off campaigns (e.g., product announcements, go-to-market) may benefit less since they require a “learning period”. In these cases, a model which captures the semantic meaning of the campaign’s content (e.g., modeled by language embeddings) may be more effective in absence of data specific to the campaign itself. It’s also hard to cold start in cases where downfunnel labels take substantial time to compute (long term ROI, feedback in days or weeks instead of hours).
For users applying Smart Targeting to their campaign for the first time, the results can seem like magic. However, even with a powerful general purpose solution, we have to ensure that the use case is aligned to the strengths of the tool and we don’t become overly confident too quickly.
Our design principles are to deliver an experience for internal users which hides the complexity of this system, build trust through observable results, while exposing advanced controls and configurations when needed.
Conclusion
Smart Targeting represents a shift in how we approach growth engineering at Coinbase. We built a platform that enables every campaign to be powered by machine learning to move away from relying on static rules and bespoke models for every targeting use case. Smart Targeting allows us to respect our users' attention by ensuring that when we target users with specific content, that decision is backed by data, not just intuition.
