Start with the decision a view must support
A shared visibility dashboard earns its place when someone can use it to make a concrete decision. A content lead may need to choose a page to improve; a product reviewer may need to identify misleading answers. These needs share evidence but require different summaries and different levels of detail.
Write a one-sentence purpose for each view before selecting charts. For example: identify which deployment questions need factual review this week. A view with that purpose should surface questionable answers and their sources, rather than allowing a broad mention trend to displace the work the reviewer actually needs to do.
Build outside Kept AI with authorized inputs
This workflow describes an external shareable dashboard built from authorized MCP reads or exported samples. It does not assume a native Kept AI custom dashboard publisher. Choose a reporting environment your team already governs, and decide whether recipients need a static snapshot or an independently maintained external report.
An external connected assistant can help organize information retrieved through Kept AI’s read-only MCP tools. Keep the source project, collection time, prompt identity, run identity when available, and evidence links with each row. Reading observations does not automatically authorize republishing their full text to everyone with a link.
Define the denominator before the headline
Create a data dictionary for every displayed measure. If a card counts answers mentioning a brand, specify the eligible runs, handling of failed responses, matching rules, and collection period. A missing answer should not silently become a negative observation. Show coverage alongside the summary so readers can see incomplete collection.
Keep prompt groups stable across comparison periods or annotate the change prominently. Adding many new discovery questions can alter an aggregate even when previous questions behave identically. Where membership changed, provide a matched subset comparison and treat the expanded collection as a separate view.
- Name the audience, decision, prompt group, and reporting window.
- Check row provenance, measure definitions, and missing observations against the authorized sample.
- Review sensitive content and interpretation, then publish through the external reporting workflow.
Give each audience a useful route to evidence
A planning view can show prompt groups, unresolved gaps, and proposed next actions. A review view can show the exact answer excerpt, source page, and factual issue. A leadership view can summarize what was checked and what decision is pending, with a clear link back to the underlying evidence.
For a deployment-question review, put the incorrect or incomplete claim beside the approved documentation passage and a review status. This makes the dashboard actionable without inventing a composite score. Keep the original answer accessible so an excerpt cannot hide a caveat that changes its meaning.
Review access and meaning before sharing
Before publishing the external dashboard, check both permissions and interpretation. Remove private prompt text where necessary, verify that source links are appropriate for the audience, and label the collection window. Have a reviewer unfamiliar with the analysis explain one card back to you; confusion is a reason to revise it.
Add annotations for page releases, changed prompt wording, unavailable observations, and corrected classifications. Describe them as context, not causal explanations. If a change in visibility follows a content release, the dashboard should show the sequence without claiming that timing alone proves an effect.
Maintain a report people can trust
Assign a refresh owner and a visible last-updated label. State whether updates are manual, externally scheduled, or snapshots; do not imply that the read-only MCP connection publishes or refreshes a dashboard by itself. Archive the data and definitions behind important decisions so later readers can reproduce the summary.
A dashboard reflects its selected questions and observation conditions, not the entire AI search landscape. Small samples, answer variation, inaccessible sources, and changed group membership limit comparisons. Retire cards that no longer inform decisions, and preserve explicit unknowns instead of replacing them with reassuring colors.