Promoted supports many technical features, analysis tools, and security components. Each feature is described in more detail in its own section, and an overview is given here.

Product

Custom models. A solution built for your own marketplace, using the features and optimizations that are right for you.
Promoted listings. In search and discovery, promote the best listings at the top and optimize their order through an advanced machine learning model.
GSV optimization. Optimize sales and ad ranking directly for Gross Service Value, which represents the total sales in dollars. Similar concept as post-purchase value optimization, or revenue/profit maximization.
Native ads. Promoted optimizes native ads for revenue and computes prices using an auction.
Content embeddings: The flexibility to load content, user, or context embeddings of any type into our CMS — Promoted's model will learn.
Impression tracking libraries: MIT-licensed user engagement logging libraries for standardized visibility logging, consistent with IAB advertising standards. Useful for all types of visibility logging useful in search and discovery.
Attribution pipeline: Attribute purchase credit to clicks, impressions, and other signals.
Real-time feedback: In-session recommendation changes for users to optimizes search and ads based on the user's real-time activity.
System replacement: At times when Promoted is able to reproduce other personalization, recommendation, and ranking systems already in use — or perform just as good alone as in conjunction with them — you may optionally choose to sunset existing systems to save infrastructure costs, lower end-user latencies, and reduce maintenance costs.
User retention: Promoted maximizes retention of users by avoiding strategies like disruptive advertising, which may result in long term revenue decreases, even if short-term revenue is increased.
Unified system: Promoted encourages an end-to-end system that returns personalized and optimized results through a full loop of collecting and feeding activity and outcome data to the discovery engine.
Relevance ranking: Promoted ensures items ads and search results are relevant to the query and user.
Semantic search: Increase search relevancy by analyzing the context and intent behind the search query, not only its lexical characteristics.
Semantic relevance : Estimate how relevant an item will be to a human or AI relevance rater, as opposed to how engaging or profitable an item will be in a context.
Geosearch: Ability to optimize search via geolocation as a search parameter.

Blender

Blender: A customized solution for deciding priority of content, using different rules to allocate items at certain positions. Blender's input is an unordered set of listings and outputs a ranked list.
Blender and hyper-parameters: Allocation is controlled by a dynamic configuration language that itself is controlled by customizable hyper-parameters that can be continuously optimized. For example, the utility function or insertion or blending rules are defined and optimized using the blender system in coordination with the 3rd stage ranker.
Utility function optimizer: A blender application for optimizing the definition of "quality" to achieve long term product goals like long-term revenue and user retention.
Diversity rules: Blender rules that encourage variety by penalizing or rewarding repeated allocation of similar items. After an item is allocated at a position, diversity rules influence future item selection by increasing or decreasing the score of other items based on the similarity in their properties.
Trimmers: Remove items in Blender that exceed a score threshold, such as having too low of a quality score or too high of a spam likelihood.
Long-term value optimizer: The correction of Blender rules themselves, such as optimizing the quality score definition and learning the optimal mix of diversity and promotion load to maximize future user retention and revenue. Promoted optimizes directly for long-term business outcomes like sales and revenue, not proxy metrics like clicks or views. This prevents falling into a trap of having positive initial results but negative revenue impact over time.
Challenger slots: An customized implementation of Blender giving you control to force new items to be shown in a reserved slot to improve new seller activation and success rates.
Cancelation and response rate controls: An application of Blender to downrank or "soft shadowban" items with poor seller behavior to encourage better seller behavior in the future.

Reporting and Statistics

Introspection: Understand what is promoted, and why, through a comprehensive report detailing the system.
Blender score access: Promoted returns all pieces of our scores for use in your own reranking function on top of Promoted, if needed. For example, this may include p(CTR), the probability of a click given an ad impression, and p(CVR), the probability of any conversion (typically, a sale) given an ad click or impression.
Instant Monitoring: Promoted gets notified if anything in the system breaks or needs work, so it can be fixed quickly. For example, if the data received is missing parameters that are usually sent.
Shared slack channel: Promoted creates a Slack channel shared among your team and Promoted, where daily model statistics and feature reports are automatically sent. You can directly respond to these messages to raise them to Promoted's or your own team's attention.
Daily model statistics: Numeric model quality statistics computed from offline evaluation about the latest model.
Daily feature reports: Links to spreadsheets describing what features were seen and used in which models, how important these features were, and if these feature importances are changing over time.
Descriptive feature importance: A simplified, unified report that best answers "what features drive model performance overall" in an intelligible way. Excludes item engagement features to help show the importance of content features that are otherwise made redundant by item engagement features. This worksheet is useful when trying to verify the value of recently added features.
Debug feature importance: A more technical feature report that shows feature importance for all features for all models, which is useful for engineering and debugging. This includes measurements of the Click Through Rate and Conversion Rate models' feature importance, per-day, for all models trained. This worksheet is best used to understand changes in the model's use of feature data.
Feature coverage report: The feature coverage statistics reports show features that frequently appear for your customers in Promoted's models and stats about their frequency and typical value have when they appear. This report verifies whether a seemingly unimportant feature is actually unimportant, or appears so due to integration issues.
Model performance comparison: Each day, new models are generated and compared against production models, in the Area Under the Curve and Normalized Cross Entropy metrics. If the new models perform better in both, the models will be published to production. Typically the same models stay in production for a few days.
Profile logs: These contain sensitive information, and are mutable. An example of a Profile record is our User log record.
Transaction logs: These are designed not to contain sensitive information, and are immutable and easy to separate from the actual user.
Input rank echoing: Promoted returns the original ranks provided by the customer, which is useful for testing.

Security

Pseudonymization: User identifiers like User IDs are substituted with a replacement key before activity data is written in a durable format. The substitution mapping cannot be reconstructed and is stored separately from the subject data.
User Deletion Requests: Promoted supports user deletion in the User content management system (CMS) API. In addition to deleting the primary record, we support row-level deletion to delete any derived records associated with the user.
Disabling Personalization: Personalization is when Promoted uses user-specific data as ML features to optimize listings in search, feed, and ads. This can be disabled per Delivery request via the Delivery API and per user via the User CMS. When personalization is disabled, no user-specific or session-specific information is used in the ML model to optimize delivery. When this happens, the optimization is similar to when a new user starts a new session — the results are still optimized, but they are not personalized because no information about the user or session is used.
ML Opt-out: Promoted can exclude traffic from use in training examples, features, and aggregate metrics in machine learning through the ignoreUsage flag. When requested, engagement directly attributable to an insertion, like impressions and clicks, will be excluded.
Data Processing Addendum: We offer a standard Data Processing Addendum (DPA) as part of our sales contract that documents our commitments to use your data responsibly, and never sell it or use it for cross-site or interest-based advertising purposes.
GDPR compliant: Promoted's GDPR local agent ID is ZF9XAE3. Promoted acts as a “processor” as defined by the GDPR as to the personal information provided by our customers.
CCPA compliant: Abides by the California Consumer Privacy Act to protect your data and give you full control over your own data. Promoted acts as a “service provider” as defined by the CCPA.
Data minimization: A guiding principle of our products — we have designed our services to collect the least amount of personal data necessary. We never sell the information we collect or use it for cross-site or interest-based advertising purposes.
Siloed data: Promoted silos user data per customer, and does not combine user data or other confidential data across customers.
Encryption: Promoted encrypts log records and database traffic in transit and at rest.
Trust website: You can request to view a complete description of our information and organization security policies at trust.promoted.ai.
Virtual private cloud: All of Promoted's servers are in a virtual private cloud, which means Promoted reserves its own space on the cloud, increasing privacy and reliability.
Private subnet: Promoted uses a private range of IP addresses (the subnet) that have access controls in place and are not publicly accessible.
Custom data retention policy: Promoted's solutions include default retention periods, such as a standard retention of one year with unneeded model training data, but clients can override these defaults to meet their retention and privacy needs.
Invalid Traffic Filtering: Promoted excludes the following: metrics explicitly identified as invalid, like internal employees or automated testing; external user traffic explicitly labeled as a bot; and traffic that reports as valid but is actually detected to be a bot or otherwise invalid through using industry-standard practices.

Data Processing

Future joins: Promoted can associate purchases — even those days in the future — with clicks and impressions to properly attribute credit to specific insertions.
Protobuf support: Promoted sends binary protobufs over the network, taking up 80% less data than JSON.
Event batching: Promoted sends data in batches when possible, saving energy while still delivering real-time metrics. This means users won't notice any performance or battery impact with services that use Promoted.
Discount promotions: Promoted provides optimized discount promotions ("offers") merchandizing, and offers for your marketplace, with live metrics showing how the promotion drives sales.
Expected value estimation: In calculating promotion metrics, Promoted estimates the lift and conversion probability with and without the promotion, then subtracts the cost of the discount. This is used to optimize promotions to show only to the most profitable users.

Testing

Shadow tier: The Shadow Tier is the real Promoted hosted system, but instead of waiting on the Promoted Delivery response and using the Promoted response to change the user experience as in full, deployed production, the Promoted response is “ignored” for optional asynchronous logging. The Shadow Tier can be used in our standard integration on unified ads + organic or ads only, and its designed to estimate value, costs, and infrastructure loads in a safe yet true production environment without risking any user experience impact. The Shadow Tier is the fastest, safest, most well-supported, and most realistic way to test and evaluate hosted Promoted systems.
End-to-end latency and infrastructure load estimation: The shadow tier produces an accurate estimate of the latency and infrastructure load of the Promoted system, without any integration risk that the data used in estimation analysis will not be available or feasible in live production.
Simulated billable events: Pacing systems consist of a feedback loop that reacts to live production context, such as modeling budget spent over the projected budget as a function of real-time aggregates of billable ad clicks. While shadow tier tests cannot use real user events to demonstrate the true feedback loop, Promoted can simulate billable ad clicks using a Bernoulli process on the expected clicks of the ads that Promoted recommend that should have been delivered to demonstrate that the live pacer system works as expected. From this, Promoted can estimate an improvement in ROAS for advertisers in the bid pacing system in the shadow tier.
Sum of expected clicks, conversions, and sales: Promoted estimates semantically meaningful, calibrated event probabilities for clicks and conversions with position and placement awareness. By summing the expected clicks, conversions, and sales given the Promoted-suggested allocation in the Shadow Tier traffic, the customer can estimate bid-weighted clicks, conversions, and sales that would be generated from ads and compare this estimate to the actual production system metrics to estimate lift.
A/B Testing: Use Promoted's optimizations only on a portion of your search or ads to compare performance with and without Promoted. This is a powerful tool to measure Promoted's impact quantitatively, and Promoted collaborates with each client to run experiments.
Holdback testing: An A/B test after launch to verify that the expected impact is still being realized. Sometimes called a "backtest."
Xray debugging: A function call tracing system for debugging. It aims to expose and understand the internal workings of a program, and involves examining the code and its execution at a granular level, much like how an X-ray reveals the internal structure of an object.
Offline Data Analysis: Promoted may be able to estimate click and conversion model performance on artificial data offline, without necessitating changes to the code. This limited data analysis cannot be used to evaluate the streaming data infrastructure, optimized serving systems, and real-time ad delivery controls like auto-bidders and pacers which would be necessary to realize the business impact in user A/B testing and live production.
Proxy endpoint: Instead of sending user event logging data directly to Promoted from client devices like a web browser and mobile apps, user events are sent first to your own servers and then forwarded to Promoted's servers. This is faster, avoids ad blockers, and it keeps the client fully in control of all data Promoted receives, if any.

Machine Learning

Item and User Feature Stores: Delivery fetches a page of "largely stable" ML features per Item or User ID which have been pre-transformed and processed for efficiency from unstructured JSON. The underlying Content Management System (CMS) preserves past versions of features with the assumption that these features are not updated continuously (less than several hundred times per day, or more commonly, at most once per day). Both dense and sparse features may be generated from a Delivery API Request.
Real-time Feature Store: ML features that represent live traffic of users and items may continuously update. Promoted saves these features with a latency of a few seconds or much less.
Content Embeddings: Some types of media (text, images, videos, audio) are not appropriate for direct use in ML and need to be transformed into compressed numerical representations. This service computes embeddings associated to the item for use in machine learning.
String and Text Processing: Promoted supports search through simple text matching, with algorithms such as longest common sequence. These text matching scores become features for ML.
Request Distributions: Promoted computes the the relative rank of certain features, such as the average click-through-rate, price, and distance, in comparison to all other ranked items. For example, this simple and cheap calculation helps view the cheapest or closest item available in this page compared to other items.
Special Transformers: Promoted automatically transforms features like distance, time-to-now, radians, and more for domain-specific applications, and additional transformations are continuously added with new customers.
Multi-task Learning for Clicks and Conversions: In search, feed, and ads recommendation, we need to predict the probabilities of multiple type of events, i.e. Clicks, Like, Save, Purchases, etc. We learn all events in a single model by building an MLP task layer for each task on the Top MLP layer. The task layer is only trained with the data for that specific task. The shared layer is trained with the data from all tasks. The multi-task architecture is particularly beneficial for tasks with very little data, like Purchase. Further, the total cost of ownership of training, managing, and creating training data is reduced. This contrasts the common approach of one model for each task where each model has its own “supporting cast”, including the data pipelines, training jobs, predictors, and computational resources.

API and SDK

Low latency delivery API: Promoted's delivery has minimal delay (typically <100ms p99), and Promoted can decrease the latency further if needed. One strategy is to deploy to your specific AWS region to keep latencies low; the major driver of latencies is networking across different data centers on the open internet.
Front-end Client Logging Evaluation: While Promoted offers free, open-source client-side event logging SDKs for iOS, Android, and Web, you may instead send your own client events, generated by your existing logging systems, to the server-side Promoted Metrics SDK. Promoted can evaluate how the Promoted client SDK can improve data quality to improve reporting correctness and sales optimization performance using the customer's own event definitions.
Delivery API: Matches content to the user and delivers ranked results.
Content API: For sending and logging data about content and users to Promoted. Any data sent via the Content API can also be sent directly via the Delivery SDK per delivery request.
Metrics API: Logs records to promoted.
Delivery API Request: Any feature can be passed on a Delivery API Request or Request Insertion as unstructured JSON. Delivery transforms this information for use in machine learning in real-time in a similar way as Item and User Feature Stores are transformed. Such features could be information from other ML systems (e.g., retrieval score), which is one way of implementing model stacking. Delivery also accepts common properties like "query" and other common attributes. Both dense and sparse features may be generated from a Delivery API Request.
IAB (Interactive Advertising Bureau) and MRC (Media Rating Council) standardization: Promoted abides by rules and standards set by the Interactive Advertising Bureau and Media Rating Council for events like impressions, clicks, and video viewing, for all client metrics in reporting, optimization, analytics, and ad billing.