What it does

Agents rate an article thumbs up or thumbs down from the knowledge search control, optionally with a comment. Each rating is a row in the feedback table, and the number shown on the article is a rollup average of the normalised ratings — not a like count.

Key facts

• The agent-facing control is binary: thumbs up or thumbs down. The comment box appears on thumbs down.
• The rating displayed on the article is knowledgearticle.rating, a rollup field defined as Aggregate(Average(feedback.normalizedrating)). It sits between 0 and 1, with decimals.
• Normalised rating is calculated as (Rating - Minimum Rating) / (Maximum Rating - Minimum Rating). The min and max are stored on each feedback row, so different sources can use different scales and still aggregate.
• The number next to the thumbs-up icon is the average rating, not the count of positive ratings. Everyone misreads this at least once.
• Being a rollup, the value is not real-time. It updates on the platform's rollup schedule, so a fresh rating does not move the number immediately.
• Feedback rows carry a Source field distinguishing internal feedback from portal feedback, so agent and customer ratings can be told apart.
• Feedback can also be created manually from the Feedback subgrid on the article, with Title, Regarding, Source, Comments, Rating, Minimum Rating, Maximum Rating and an optional contact.
• Created By is set automatically on creation; Closed By is set automatically when the feedback record is deactivated.
• Service representatives can create, view and edit only their own feedback. Customer service managers and knowledge managers can see everyone's.
• Feedback must be enabled on the table before any of this works — KnowledgeArticle has to be enabled for feedback/ratings in the table's settings.
• The Enable feedback toggle in Copilot Service admin center > Knowledge > General settings controls whether agents see the rating control at all.
• View counts are tracked separately in KnowledgeArticleViews and are also a rollup, with the same delay characteristics.

When to use / skip

Turn it on. It costs nothing and the alternative is having no signal at all about which articles are actually useful.

What you should be sceptical about is what the client plans to do with it. Feedback volume from agents is low — single-digit percentages of article opens is normal, and it skews negative because people rate when something is wrong. Any target built on average rating will be noisy and easily gamed. A KPI like "all articles above 0.8" is not a KPI, it is an invitation to stop rating.

Where it genuinely helps is triage. A knowledge manager looking at a view of articles with recent thumbs-down and reading the comments gets a real backlog of things to fix. That is the use case to design for: comments as work items, not ratings as a score.

Customer-facing ratings from a portal are a different proposition — higher volume, less context, and worth separating from agent ratings using the Source field before anyone builds a report.

Configuration decisions

• Whether feedback is on for agents, for portal customers, or both — and whether the two are reported together or separately.
• Whether thumbs-down comments are mandatory in practice. The product does not force a comment, so if you want one you need a flow or a customisation.
• Who owns the feedback backlog. Feedback with nobody triaging it is worse than none, because it teaches agents that rating does nothing.
• Whether feedback rows are deactivated when actioned (populating Closed By) or left open. Pick one and build the views around it.
• Whether managers need a dedicated view or dashboard of recent negative feedback, and which security role sees it.
• Whether any custom feedback source beyond the built-in internal/portal values is needed — for example feedback arriving from a chatbot or a survey.

Gotchas

• The rollup delay means a client testing feedback in UAT will rate an article, see no change, and log a defect. Explain the rollup behaviour before UAT, not during it.
• The 0-to-1 average is routinely mistaken for a count of likes. Rename the field label on the form if the client will look at it daily.
• Agents can only see their own feedback. A team lead testing with an agent role will conclude feedback is missing.
• Turning off Enable feedback hides the control but leaves existing feedback rows and the rollup value in place, which looks odd on the form.
• Feedback is tied to the specific article version record. Create a new major version and you start from a clean slate — historic ratings do not follow the content forward.
• Normalised rating depends on min and max being set correctly on each row. Records created by an integration with min and max left blank will distort the average.
• Deleting feedback to "clean up" the average is possible and irreversible, and someone will suggest it.

Consultant notes

• Demo the knowledge manager's view of thumbs-down comments, not the agent's thumbs. The value is in the backlog, not the click.
• Tell the client to expect low participation and to treat rating volume as a health metric in itself, rather than chasing the average.
• Agree the triage owner and cadence before go-live. Put a name against it in the RACI.
• If a portal is in scope, insist on separating internal and portal feedback in reporting from day one. Retro-fitting that split is painful.
• Check before go-live that KnowledgeArticle is actually enabled for feedback in the target environment — it is a table-level setting that does not always travel with the solution as expected.

---

Worth revisiting if the client starts publishing to customers, or if Microsoft moves the rating control beyond thumbs up and down.