Skip to content
AskVault Docs
Try Free →

The content_recommender skill

Written by , Founder, AskVault · Reviewed by Aashiq
Last updated: · 4 min read

What the skill does

After answering a customer's question, the bot can suggest related content the visitor might want to read next. Recommendations surface in about 200 ms after the main answer completes, drawing from up to 5,000 indexable documents per workspace. Three patterns:

  • "Want to learn more?" Bot suggests a deeper blog post on the topic just discussed.
  • "Other customers like you read..." Suggests case studies that match the visitor's apparent industry or role.
  • "Related guides" Surfaces tutorial content for visitors in onboarding flows.

Lift varies by industry, but average session-engagement (pages-per-session, time-on-site) goes up 15 to 25%. On marketing sites the lift can be larger; on pure support sites smaller.

The skill is on Growth and above. Growth+

How recommendations are picked

The skill embeds the conversation's recent context (last 2 to 4 customer messages) into a vector. It then runs vector search against your indexed content, filtered to "recommendable" sources (you mark these during setup).

The top 2 to 3 matches surface as cards or links at the end of the bot's reply:

Here's how to enable two-factor auth: open Settings, click Security, click Enable 2FA, follow the prompts.

You might also want:

Tune the number of recommendations and where they appear under AI Agents > Skills > content_recommender > Display.

What sources are recommendable

By default, the skill recommends from:

  • Blog posts (if indexed from WordPress, Ghost, or via URL crawling).
  • Case studies (if you've tagged documents with case_study audience).
  • Help articles (if ingested from Zendesk, Confluence, or similar).
  • Tutorials and guides in your indexed content.

Configure which sources are recommendable under Skills > content_recommender > Sources. Set per-source weight (lift the prominence of case studies, downweight old blog posts).

For documents you don't want recommended (legal pages, internal docs), exclude them with the not_recommendable audience tag.

Setup

Three steps to working recommendations.

  1. Index recommendable content. Either crawl your blog/help site via URL crawling, connect Ghost, WordPress, or upload as files.
  2. Enable the skill. AI Agents > Skills > content_recommender > Enable.
  3. Configure display settings. Number of recommendations (default 3), placement (after every answer, only on long answers, or only when explicitly asked).

Test by asking the bot a question whose answer matches one of your topics. You should see 2 to 3 related links surface alongside the response.

Trigger conditions

Three trigger patterns, configurable per workspace:

  1. After every bot answer (most aggressive). Always shows recommendations. Use when content discovery is a primary goal.
  2. After longer answers (default). Skip recommendations on quick yes/no responses; surface them after detailed explanations.
  3. On explicit request only (least aggressive). Visitor asks "do you have more on this?" or similar.

Most teams start with the default and tune up or down based on engagement data.

Personalization with identity verification

For identity-verified visitors, the skill personalizes by:

  • Reading their plan tier from your CRM (via HubSpot or Salesforce) and matching content tagged for that tier.
  • Reading their industry or company size and matching case studies that fit.
  • Reading their recent product activity (via the metadata field in chat requests) and matching tutorial content for features they haven't yet adopted.

Personalization makes recommendations significantly more relevant. Anonymous visitors get topic-based matches without personalization. Growth+

Two display modes:

  • Inline links. Recommendations appear as a short "You might also want" footer in the bot's reply. Less intrusive, lower CTR.
  • Visual cards. Recommendations render as clickable cards with title, excerpt, and image. Higher CTR but takes more vertical space.

Visual cards work in the website widget channel. Other channels (WhatsApp, SMS, voice) fall back to inline links because rich media isn't always supported.

Limits

  • Plan availability. Growth and above.
  • Number of recommendations per reply. 1 to 5. Default 3.
  • Source pool size. Up to 5,000 recommendable documents per workspace.

Common pitfalls

Recommendations feel off-topic. Source pool includes documents that aren't tagged correctly. Audit under Skills > content_recommender > Source Audit and exclude noise.

Recommendations are too uniform. Skill keeps suggesting the same 3 popular posts. Diversify by enabling Skills > content_recommender > Diversification which spreads recommendations across topics.

Visitors ignore the recommendations. Display mode is too subtle. Switch to visual cards (if widget channel).

Recommendations dilute the main answer. Too aggressive trigger setting. Switch to "after longer answers" or "on explicit request only".

FAQ

Yes. Tag them with the not_recommendable audience tag under Knowledge Hub > [document] > Audience. The recommender skips them.

By default, recommendations point to URLs you've indexed (your own site). For external recommendations (e.g., link to industry research from third parties), add them as Q&A pairs or snippets with source_url pointing to the external site.

How fresh are the recommendations?

As fresh as your indexed content. If you publish a new blog post and run a re-crawl, the new post becomes recommendable within minutes.

Can I A/B test recommendation strategies?

Yes. Create two workspaces with different recommendation configurations, split traffic between them at the widget level. Compare engagement metrics in Analytics.

Will recommendations affect SEO if I'm running them on a marketing site?

The recommendations are bot responses, not page content. They don't affect SEO. Visitors who click through to your blog posts increase internal traffic, which is a small positive signal.

Was this page helpful?