Content intelligence

Most Intelligence screens work from an index of what your published entries and image assets mean. To build it, choose Intelligence → Embeddings in the sidebar, then Index library. Content queues every changed source and says how many it queued. Background indexing works through the queue over time; Run worker processes a batch straight away, and Retry jobs puts failed jobs back in the queue.

The Library & jobs tab lists the Indexed library, each source with its type, number of vectors and when it was updated, and the Background jobs, each with its status, attempts and details. Images get one vector each, and long entries are split into several chunks. Failed jobs keep their reason, and jobs held back by a limit keep their work and try again later. The figures across the top show the number of Content vectors and Audience vectors, and how many indexing Requests today have been used of the daily budget.

Indexing and changing its settings need the Manage content types, fields and blocks permission, which the Owner and Admin roles have.

The Intelligence screens draw on three sources:

• The index of published entries and image assets, built under Embeddings. Semantic Search, AI Tag Suggestions, Duplicate Detection, Content Gaps and the Knowledge Graph use it, so they're empty until the library is indexed.
• Your structure: taxonomy terms and references from Content architecture.
• Recorded activity: the events your site sends, covered in Audiences, and the decisions your content slots make.

Opening the Intelligence screens that use the index needs a role that can see the Intelligence, Recommendations and Audiences sections, which Owner, Admin and Editor can. Each of these screens has Refresh, Export to download its rows as JSON, and a box to filter the results. Similarity scores show how alike two things are; they aren't probabilities or predictions.

The Configuration tab of Intelligence → Embeddings controls what gets indexed. Turn these on or off:

• Background indexing: off pauses new indexing and keeps existing vectors.
• Images: index image assets.
• Published content: index published entries as searchable chunks.
• Audience interests: build interests only for profiles that granted personalization consent.

Then set the limits: Daily ingestion budget, Content vector limit (100 to 50,000), Topic vectors per user (0 to 5) and Interest half-life (days) (1 to 90). Choose Save changes; changes apply to later indexing runs. If the screen says indexing can't run until a key is set up on the server, existing vectors still work for similarity.

Intelligence → Semantic Search finds images and published entries by what they mean, even when they don't share your words. Describe what you want, like "a bright product photo on a wooden table", and choose whether to search All sources, Images or Content. Results are ranked by similarity, with editorial adjustments applied, and show "No matching results" when nothing is close. Only indexed sources can be found.

Intelligence → AI Tag Suggestions suggests taxonomy terms an entry is missing, based on similar indexed entries that have them. Each suggestion shows the entry, the suggested term and its taxonomy, the similarity and the similar entries it's based on. Pick how close the evidence must be, from Similarity ≥ 65% (the default) to 95%.

The screen reviews 100 sources at a time: Next sources and Previous sources move through the rest. It uses stored vectors only, so it doesn't use any indexing budget. To accept a suggestion, assign the term under Architecture → Taxonomies.

Intelligence → Duplicate Detection pairs up published entries that are at least 80% similar, each with its similarity. Raise the bar with the threshold menu, from At least 80% similar to 95%; it starts at 90%. The screen reviews 100 sources at a time, with Next sources and Previous sources. Check intent before you merge or remove anything.

Intelligence → Content Gaps lists taxonomy topics with too few indexed published entries. Pick the target, from Target: 1 per topic to 10 (3 to start). Each topic shows its taxonomy, how many indexed entries it has, how many it's short of the target, and the entries that cover it. It measures how much you've published on a topic, not how much readers want it. When every topic meets the target, the screen says so.

Intelligence → Knowledge Graph shows how published entries connect: through references, taxonomy terms they're tagged with, and similarity of at least 75%. The diagram draws up to 40 entities; dashed lines are inferred similarity. Click a node to filter the relations to it.

The Relations table lists each connection, its type and its evidence. Filter by relation: All relations, References, Tagged, Taxonomy or Similar. The screen reviews 100 sources at a time, with Next sources and Previous sources.

Intelligence → Decision Explanations lists the reasons Content gave when your app asked a content slot what to show: which variant won and why. Pick the period, Last 7 days, Last 30 days (the default) or Last 90 days, and the kind, starting on All decisions. Decisions for visitors who denied personalization consent aren't recorded. Personalization explains how slots decide.

Four screens report on how content and audiences perform, from the events your site sends and daily totals:

• Content Performance: a traffic trend of views, engaged sessions and conversions, the channel mix, performance by type, underperforming entries and top content.
• Audience Performance: engagement by segment, the journey mix, reach against quality, and a segment leaderboard.
• Funnels & Journeys: the acquisition funnel, drop-off between steps, entries against conversions, and top paths.
• Topic Interest: interest momentum, maturity by stage, and a topic leaderboard.

Until the project has recorded data for a report, the screen shows sample figures marked Sample data. Don't read anything into them.