Overview
The Contextual Recommendation and Matching Engine from EX.CO improves content relevance and user engagement through video matching and boosts content circulation through AI RSS. How it works and the benefits it brings are explained in this guide.
How It Works
-
Contextual Recommendation:
- The engine maintains a score for each video based on its contextual relevance to the page it's served on.
- Videos are selected for the playlist based on the feed history (such as the last 7, 14, or 30 days) for the article the player is loaded on.
-
Contextual Matching Engine:
- Integration: The engine is fully integrated into the EX.CO platform, requiring only the enabling of domain scraping based on your list of publisher's partner domains.
- Application to Playlists/Players: When applied to a playlist or player, it actively scans each new article, ensuring that content is always fresh and relevant.
- Contextual Relevancy Scoring: Videos are scored for relevancy to the content of the current page, ensuring the most relevant videos are displayed.
- Dynamic Content Selection: The engine can select videos based on various recency settings, allowing customization of content presentation.
Even Without Video, There's a Solution:
If you don't have video content, there's no need to worry. You can integrate your website's XML feed with the EX.CO AI-RSS Solution, which is designed to create video experiences for content circulation.
By combining this with the contextual engine, you can craft an engaging on-page video experience. This setup will recommend content that is relevant to your readers, enhancing every story on your website.
Related guide:
Results
Implementing this engine has shown significant improvements:
- Revenue Uplift: Up to 17% to 49% increase in revenue.
- Dwell Time Uplift: A 49% increase in user dwell time on pages.
Benefits
- Enhanced User Engagement: By presenting contextually relevant videos, users are more likely to engage with the content.
- Increased Revenue Opportunities: More relevant content can lead to higher ad engagement and revenue.
- Automated Content Tailoring: The automated system ensures that the content displayed is always suited to the specific context of each page.
Contextual Recommendation Q&A
1. What does the MRSS feed need to contain for each video asset?
Every video item in your MRSS feed is ingested into your EX.CO Media Asset Management (MAM) catalog, becoming a managed media asset on our platform. Any metadata you include in the feed carries through to that catalog entry, and because the recommendation engine reads directly from the catalog, every attribute you provide directly improves match quality.
Required: media:content, with attributes url, duration, fileSize, width, height, lang/language.
Optional, each improving recommendation quality:
- title — video title
- pubDate — publish date, used for freshness/recency
- link — click-through URL
- media:thumbnail — preview image
- media:name — fallback title if title isn't present
- media:group — container for multiple media:content renditions
- media:category — content categorization
- media:keywords — additional topical tagging
- guid — unique identifier per item, used to correlate recommended results back to your own asset IDs (recommended for this integration specifically)
Automatic AI enrichment (optional capability): when EX.CO's contextual AI analysis is turned on, every video is automatically enriched with additional metadata, no extra publisher effort required:
- A full transcript of the video's audio, generated automatically as part of the same analysis process and available for your own use
- AI-generated summary and semantic description of the video's content
- Topics and named entities detected (people, organizations, places, events)
- Content categories mapped to IAB standard taxonomy + custom publisher categories
- Automatically extracted keywords
- Auto-generated chapter markers with timestamps
- Sentiment and emotional tone
- Brand-safety and kid-safety flags
- Detected language (and regional accent, where relevant)
- Audience affinity signals
2. What metadata is used to determine recommendation relevance?
EX.CO's recommendation engine determines relevance through semantic AI matching: both the article and each candidate video are converted into vector embeddings capturing their actual meaning and content, not just keywords, then compared using similarity scoring. The videos with the highest similarity to the article are the most relevant matches. Richer video metadata, especially when AI-enriched with automatic topic, entity, and category detection, translates directly into stronger matches.
3. Is there a waiting period before a video becomes eligible for recommendation?
No fixed waiting period or embargo. Once a video is added to the feed, AI analysis runs automatically. Completion time depends on video length and current processing queue depth, but EX.CO targets individual asset analysis completing within a few seconds, so videos become eligible shortly after being added.
4. How long does it take to index an article and generate recommendation results?
There are two distinct processes here. First, the page is crawled and analyzed to understand its content. This happens once per article.
Second, recommendations are generated by matching that article's analysis against the video catalog. This matching step happens essentially immediately once the article has been analyzed. The first visit to a new article triggers the one-time analysis; every recommendation lookup after that returns fast, ranked results.
5. What happens when there are no strong recommendation matches? Is there a relevance threshold?
Recommendation is based on ranking, not a fixed relevance cutoff: the engine always scores and ranks the available videos by relevance to the article, and the highest-ranked match is served first. A recommendation is always available, ordered by how closely it matches the article's content.
6. How does the player retrieve recommendations?
Only the fully managed player implementation is available today: the EX.CO player calls the recommendation engine directly and renders the results automatically, with no additional integration work required on the publisher's side.
7. What recommendation experiences are currently supported?
The supported recommendation today is our content-matching scenario, in which we match articles to relevant videos. Both the videos and the articles are passed through a vectorization model to obtain their embeddings, and those vectors are then compared using similarity metrics, where high similarity scores indicate that the articles are contextually related to the video or videos. Based on those similarity scores, the videos are ranked and compiled as a playlist, presented in order of relevance to the article. On our roadmap, recommendation will become session-based and will incorporate viewer engagement, content recency, and popularity.
8. What reporting and analytics are available for recommended content?
Recommendation performance isn't yet broken out in reporting today. This is part of the active roadmap, building A/B testing for recommendations so performance and conversion impact can be directly measured against a control experience.
9. How are recommendations seeded initially?
Recommendations are seeded entirely through content-based matching from day one, vectorizing the article and video content and ranking by similarity. There's no reliance on accumulated user engagement or behavioral data at this stage, so no learning or warm-up period is needed before recommendations start working.