Release Summary 26.20 | Oct 06, 2026
The following key features and improvements, along with bug fixes, have been released in ADA Global Digital Experience Personalization (DXP) in the release version 26.20.
Ensemble AI
Automated Quality Checks for Ensembles and Virtual Try-On Images
Virtual try-on Validation: Virtual try-on validator (LLM as judge) now reviews every virtual try-on image against a set of quality checks and flags it as valid or not valid. The review confirms that all garments are applied accurately, including color, pattern, print, texture, and logo placement. It also confirms that a single, naturally proportioned model is fully dressed, with no duplicate or overlapping garments in the same slot, such as two bottoms. Jackets worn over tops are treated as a valid combination. Images are also checked for background match, correct orientation with no mirrored details, and the expected aspect ratio and framing.
When an image does not meet the criteria, a new virtual try-on is generated automatically, with up to two retries. If it still does not meet the criteria, it is excluded from the outfit list and tracked separately to help identify the cause. Merchandisers also have an option to review through the portal and flag them and exclude any outfits with virtual try-ons.
Ensemble Validation: Added a validator (LLM as judge) that ensures each part will have the products that are as per the slot definition in prompt. This will ensure the outfits will not have same categories in different slots.
Virtual Try-On Images Now Available Immediately in Ensemble AI Responses
The recommendations job that adds virtual try-on images to Ensemble AI responses now runs in sync with the virtual try-on job. As soon as the try-ons are generated, the recommendations job updates their URLs, so they appear in the API response right away. The sequence is now: Vector job, LLM job, Ensemble AI job, recommendations job for ensembles, and then the virtual try-on job followed by a recommendations job for both try-ons and ensembles.
Every ensemble with a virtual try-on now returns it in the API response, so shoppers see complete outfit visuals from the start.
Jira: ENG-33258
Exclude Brands Filter in Ensemble AI Style Definition
Merchandisers can now exclude specific brands from outfit generation in the Ensemble AI style definition. The new Exclude brands option sits alongside Include brands and removes products of the selected brands from the seed products, for both free form and structured styles. For free form styles, an Apply for all parts option extends the exclusion to every part of the ensemble, so no product from an excluded brand appears in any generated outfit or virtual try-on. Brands are selected from the client's catalog, with a scrollable list of up to five brands at a time, search, and a paste option to add many brands at once without duplicates. A clear all option removes every excluded brand in one step.
Existing outfits that contain excluded brands are regenerated or filtered on the next refresh cycle. This gives merchandisers direct control over brand presence in outfits and virtual try-ons without manual cleanup.
Jira: ENG-33206
Social Proof
Message Template Reporting in Social Proof
Message level reporting in Social Proof now shows message templates instead of individual message variations. For example, "10 people bought today" and "25 people bought today" are grouped under one template, "@usercount people bought today". This gives merchandisers a clear view of how each message performs, rather than a long list of messages that differ only by count.
The template view applies to any message that includes a user count or event count, and to messages that embed attributes such as categories, brands, or product attributes. Messages with no counts continue to appear as they are. Merchandisers can compare message effectiveness at the template level and make better decisions about which messages to keep.
Jira: ENG-33231
Find
% Zero Results KPI in the Find Report
The Find report now includes a new KPI, % Zero Results. It shows the share of search views that returned no results, calculated as the number of searches with zero results divided by total searches (views). The KPI appears in the Find report table alongside the other Find metrics and includes a line chart, so you can see how the rate changes over time.
Merchandisers can spot rising zero result rates early and act on them, for example by adding synonyms or improving catalog coverage, so shoppers find what they are looking for.
More Accurate Zero Results Search Terms Report
The Search Terms with Zero Results report now counts a search term only when the first page of results is empty. A request for page 1, or a request with no page specified, that returns no results is recorded. Requests for later pages, such as page 20 or 35, are no longer counted, even when they return no results. A term such as "cotton t-shirt" that has matching products therefore stays out of the report, while searches that return no results on page 1 are still reported.
Merchandisers now see only the search terms that truly have no matching products. This makes the report a more reliable guide for closing catalog gaps, adding synonyms, and improving search coverage.
Jira: ENG-33120
Recommend
Improved Image Similarity with Color Filtering
Image Similarity recommendations now use a newer image model that delivers more accurate matches, especially for fashion and apparel products and for photos featuring a model. The model keeps all existing seed, sort, filter, and option choices, and it adds a color filter. In the Models section of Configurable Strategies, you can select Match Items with a Similar Color to return products in the same color as the seed product, or Match Items with a Different Color to return products in a different color. You can select only one option, and when neither is selected, no color filtering is applied.
Colors are matched against a standard set of 30 colors, so variations in how clients name colors, such as "Dunkelgrün" and "Grün", are treated as the same green. The filter also works when a product comes in multiple colors, returning the product in the color you ask for. Shoppers see closer, more relevant similar products, and merchandisers have better control over color consistency.
ENG-33057, ENG-33058, ENG-32565
Edit List Option for Fixed Product Lists in Configurable Strategies
When Fixed Product List is selected as the personalization seed, an Edit list button in Advanced Settings now opens a side panel for managing your products. You can search the list, select several products at once and remove them with the action bar at the bottom, and reorder products. Products are also added to the list from this panel. Product counts appear at the bottom of the panel and in Advanced Settings. Managing a long product list is now faster, saving merchandisers time.
Rename Existing Compound Strategies
You can now rename an existing Compound Strategy from its edit page without recreating it. A pencil icon next to the strategy name makes the name editable, and you can save or cancel the change. The new name appears right away in the Compound Strategy list and wherever the name is shown, such as Strategy Rules, while the configuration and rules stay unchanged. The same name requirements that apply at creation, including uniqueness and length, apply when renaming.
A left arrow next to the strategy name now replaces the Back button, matching Guided Selling and Dynamic Experiences, and the "Edit Compound Strategy" prefix is removed from the page title. Merchandisers can keep strategy names accurate and organized with ease.
Strategy Name and Message in Recs Test Drive for Email
The recs test drive for email product recommendations now shows the strategy name above each set of recommendations. A strategy message row appears between the strategy name and the recommendation images, so you can confirm that the email content looks right. The strategy message appears when the layout definition includes it.
Test drive requests are excluded from reporting, so previews do not add to view counts or click-through data. The email service recognizes the dev=true parameter on click links and passes it through to the click URL, which lets the test drive retrieve the strategy name without affecting reports. The HTML Code tab keeps its original requests so the code you copy is ready to use.
Warning When Adding a Content Manually with Regional Attributes
When you add content manually through the content catalog, if the content has regional attributes, a warning message appears if the content is set to be recommendable globally. This helps you confirm the setting before the content becomes available across all regions.
Jira: ENG-32745
Other Feature Enhancements
The following feature enhancements and upgrades have been made in the release version 26.20.
|
Jira # |
Module/Title |
Summary |
General Availability |
|---|---|---|---|
|
Discover: Performance Reporting Data for New Discover |
We have added performance reporting data for New Discover, covering discover visits, views, clicks, add to cart actions, conversions, sales, and orders, along with comparable measures for visits without a discover interaction. It also includes key KPIs such as conversion rate, revenue per visit, average order value, exit rate, findability, and weighted click-through rate, giving merchandisers the data needed to review discover performance and optimize configurations. |
06-Oct-26 |
|
|
Recommend: Locale-Based Recipe Recommendations |
Recipe recommendations now use the locale passed in the request. When none is passed, they use the site's default locale, built from the Dashboard Locale Language and Country settings, such as en-US. Recipe language, ingredient names, and cultural tagging match the site's configured locale instead of being inferred from the shopper's region. The new meal type and main ingredient display name fields from the recipe catalog are also recognized. Recipes for regional variants of a locale, such as en-IE, are adapted from the primary locale's recipes, such as en-GB, instead of being created from scratch. They are translated and lightly adjusted where needed, then added to the Universal Recipe Catalog. Recipes for en-IE and en-SA were created and reviewed with Product Management. Shoppers see recipes in the right language on product detail pages, and reusing existing recipes and images lowers the cost of supporting new locales. |
06-Oct-26 |
|
|
Recommend: Viewed Product Shown for Its Matching Recipe Ingredient |
Recipe recommendations on product detail pages now show the product the shopper is viewing as the recommended product for its matching ingredient. This applies to every returned recipe that includes that ingredient. Other ingredients continue to receive personalized product recommendations as before, and recipes that do not include the matching ingredient are unchanged. Shoppers no longer see a competing product for something they have already selected, which keeps recipe recommendations consistent with the page they are on. |
06-Oct-26 |
|
|
Enterprise Dashboard: Simplified Query Understanding Configuration |
The Query Understanding section is now simpler to configure. You select one or more attribute names, such as brand or category, and choose whether to boost, filter, or exclude them, with an option for multi-filter match. Rules now apply to the attribute as a whole rather than to specific values. For example, if a search keyword contains brand names, all matched brands are boosted or filtered, with no need to set a rule for a brand such as Nike. The DC drop-down is removed from Advanced parameters, and deleting a configuration from the list also resets its detailed view. Merchandisers can control matching behavior with fewer steps and less setup. |
06-Oct-26 |
|
|
Recommend: Site Setting for the Solr Catalog Model |
A new site setting controls whether a site can use the Solr Catalog model in Configurable Strategies. The interface checks this setting to show the model only for sites that are enabled for it. This keeps the model available to the right sites and gives administrators control over who can use it. |
06-Oct-26 |
|
|
Find: Faster Global Rank Processing with a Recommendable Products Filter |
The global rank publisher job can now be set to calculate rankings for recommendable products only. Sites that are enabled for this setting skip products that are not recommended, which shortens job run time and keeps the datasets smaller. This makes ranking updates faster and more efficient for sites with large catalogs. |
06-Oct-26 |
|
|
Social Proof: Independent Settings for Each Social Proof Messaging Mode |
Social Proof now keeps the settings for Interval based, Recency based, and Personalized messages separate. Each mode has its own default options and threshold ranges, so changes in one mode no longer carry over to another. Entries you make in each mode are held for the session, so you can switch between modes and return to your selections before saving. Changing the event type keeps your ranges, intervals, and messaging as they are. When you save, only the selected messaging mode is saved. Merchandisers can configure each mode with confidence, without resetting defaults or re-entering values. |
06-Oct-26 |
|
|
Social Proof: Region Added as a Context Feature for Social Proof |
Region is now available as a context feature for Social Proof, alongside the existing context features. Social Proof optimization can use the shopper's region when choosing which messages to show. This helps Social Proof messages fit the shopper's location, making them more relevant and effective. |
06-Oct-26 |
Bug and Support Fixes
The following issues have been fixed in the release version 26.20
|
Jira # |
Module/Title |
Summary |
General Availability |
|---|---|---|---|
|
Discover: Boost and Bury Now Takes Priority Over Personalization in Discover |
We have fixed an issue in Discover where personalization could outweigh boost and bury settings on list pages. Products you boost now appear above personalized products, and buried products are placed as configured. |
06-Oct-26 |
|
|
Enterprise Dashboard: Fixes to the DSW Status Page |
We have fixed several issues on the DSW status page. The latest status now loads once when the page opens instead of refreshing every 5 seconds. Run history in a strategy's details shows the job ID when a status is not available, replacing the "unknown" entries. Edit schedule now pre-selects the query currently in use. Run details history also loads reliably whenever job history data is returned. |
06-Oct-26 |
|
|
Find: Search Results Fixed for Certain Phrases in Find |
We have fixed an issue that affected Find for a specific client site, where searches for certain phrases returned no results even though matching products were available. Those searches now return the expected products. |
06-Oct-26 |
|
|
Recommend: Recommendations Now Return for Users with User Attributes |
We have fixed an issue where recommendations were not returned for some placements when the request included both a user ID and user attributes. These placements now return recommendations as expected. |
06-Oct-26 |
|
|
Science: Faster Style Seed Product Processing for Large Catalogs |
We have fixed an issue where retrieving seed products for a style on a site with a very large catalog could run for days and use significant database resources. Seed products now load efficiently, so processing completes without delay for that site. |
06-Oct-26 |
|
|
Recommend: Production Startup Issues Resolved |
We have fixed startup issues that affected two services during a production deployment. The recommendation server now starts correctly when loading advanced merchandising rules, and the model master service now starts with its required components in place. |
06-Oct-26 |
|
|
Discover: Pagination Corrected in New Discover |
We have fixed an issue in New Discover where a request could return more products than the number requested. Each page now returns the requested number of products, so pagination works as expected. |
06-Oct-26 |