Decisioning Frequently Asked Questions decisioning-faq

An offer is capped as soon as any single condition is met. When multiple capping rules exist, the offer stops being displayed once any rule reaches its threshold.

Example:
If you define two capping rules for an offer:

• 5 times per profile per week
• 100 times total across all users

The offer will stop being displayed to a user once they’ve seen it 5 times in a week, even if the total cap of 100 hasn’t been reached yet. Similarly, once 100 total impressions are reached, the offer stops being displayed to all users.

Learn more about capping rules.

When configuring AI models, both datasets and audiences serve distinct purposes.

• Datasets: Capture conversion events (clicks, orders, revenue) that serve as optimization targets for the model.
• Audiences: Function as predictor variables that enable the model to personalize recommendations based on customer segment membership.

Audiences do not restrict or expand the model’s scope. Instead, they provide contextual attributes that improve the model’s ability to make personalized predictions across different customer segments.

Both components are required for effective personalized optimization models model performance.

The auto-optimization model serves traffic to the next best available offer based on traffic data from the last 14 days, whether the personalized optimization model uses traffic data from the last 30 days.

When several offers are removed simultaneously and the remaining offers have minimal traffic data within the 14-day or 30-day window, the model may exhibit suboptimal behavior, including random distribution patterns or bias toward offers with higher conversion rates based on limited impression data.

Best practice: When modifying offer collections significantly, verify that remaining offers have sufficient historical performance data to maintain model effectiveness.

AI models identify and begin testing newly available offers on their next training cycle:

• Auto-optimization identifies and begins testing new offers on its next training cycle. Auto-optimization training occurs 3 to 4 times daily, approximately every 6 hours.
• Personalized optimization identifies and begins testing new offers as they are added to the offer strategy. They will be included in random exploration traffic. Then, those offers will be personalized in the model’s next training cycle, occurring on a weekly basis.

Once identified, both models will start serving the new offers to some visitors immediately in order to test their performance and gather data about their effectiveness.

Learn more about auto-optimization and personalized optimization models.

Both auto-optimization and personalized optimization models employ an “explore-exploit” strategy that eliminates the need for dedicated control groups.

• Initial phase: Models begin with 100% exploration, testing different offers to establish baseline performance data.
• Adaptive optimization: As behavioral events accumulate and prediction accuracy improves, models automatically balance exploration and exploitation.
• Ongoing learning: The system progressively allocates more traffic to high-performing offers while continuing to test alternatives.

This ensures continuous learning and optimization across all traffic without requiring separate control groups.

Adobe recommends the following minimum thresholds to ensure effective model performance:

• 1,000 impressions per offer/item per week
• 100 conversion events per offer/item per week

By default, the system will not attempt to build personalized models for offers/items with fewer than 1,000 impressions or 50 conversion events.

Learn more about data collection requirements.

AI models generate greater personalization benefits when offers appeal to distinct customer segments. When offers are highly similar, two outcomes are typical:

• Equivalent performance: Offers perform identically and receive approximately equal traffic distribution.
• Dominant offer: Minor differences cause one offer to outperform others across all segments, capturing the majority of traffic.

Best practice: Design offers with meaningful differentiation that align with distinct customer segment preferences to maximize AI model effectiveness.

Traffic anomalies are incorporated into the model proportionally within the 30-day rolling window, which provides model stability during temporary traffic fluctuations. Short-term spikes or drops do not significantly disrupt model predictions or performance.

A temporary traffic spike (for example, twice the daily traffic) has minimal effect on overall model performance because the anomalous traffic represents a small fraction of the 30-day dataset.