AI visibility and analytics

Read stored evidence first

The overview and citations reads are provider-bounded and truncated. Treat them as bounded samples, not a census, and use a refresh only when new collection is needed for the decision.

Interpret the measurement

For one prompt, visibility is citations or qualifying mentions divided by eligible completed runs for that prompt, engine set, locale, and window. Preserve the completed-run count, citation count, and percentage together. Zero citations across zero completed runs means the measurement is unavailable, not zero percent.

Compare equal-length windows and do not silently merge engines, locales, or time windows. A recommendation is an inference until a later comparable measurement shows a change.

Refresh only when you need new evidence

A refresh collects citation data for the workspace and returns an accepted asynchronous result with the workspace id, queued query count, credit ledger id, provider receipt, receipt, and graph effects. Read the terminal evidence after the queued work finishes before reporting a result.

When it goes wrong

A missing connection, stale window, queued scan, empty denominator, or unavailable provider data is a limitation to report. Correct the scope, wait for terminal evidence when work is queued, or state what connection or evidence is missing. Do not claim that a recommendation improved visibility until a comparable later measurement shows it.

A recommendation becomes a measured change only after a later comparable run. Preserve equal-length windows and matching engine and locale scope, then state whether the later evidence actually changed.

If a scan is queued, wait for terminal evidence before reporting a result. An accepted queue receipt proves that collection was scheduled; it does not prove that the requested citations or rankings were found.

An empty or truncated response is a limitation to name. It may represent a bounded sample, provider coverage gap, incomplete collection, zero eligible runs, or a different scope from the question being asked.

A successful read has a result with evidence and its measurement bounds. A successful refresh has terminal evidence that can be read and compared, not only an asynchronous acceptance receipt.

When numbers disagree, re-read the result and compare prompt, engine set, locale, window, completed runs, citation count, provider coverage, and completeness before reporting an improvement or decline.

Read before refreshing

The dashboard sidebar opens AI Visibility at /dashboard/aivisibility. Use the screen to inspect stored evidence before asking for new collection or turning an observation into a recommendation.

Visibility is a knowledge-layer pack. It owns analytics, attribution, SEO, share of voice, competitor evidence, brand presence, AI answer-engine citations, meetings, and feedback about how business content is seen.

Visibility does not own social post creation, provider connection setup, or a new people or company list. Use Social for posts, Integrations for credentials, and Tables for a working list.

The overview read is free and read-only. Its bounded result includes visibility, citation coverage, rank, share of voice, engines, sentiment, brand, and evidence, so it is useful for describing what was measured rather than making a promise.

The citations read is also free and read-only. It returns cited domains, citation events, pages, competitor differences, source gaps, and evidence, which makes the underlying observation inspectable.

A measurement needs its window, engine set, locale, completed-run count, and denominator. Do not compare a narrow provider window with a wider multi-engine window as if they are the same measurement.

For a prompt, visibility uses citations or qualifying mentions divided by eligible completed runs for the selected prompt, engines, locale, and window. A zero denominator means the result is unavailable, not a zero-percent performance claim.

Citation measurement and rank tracking answer different questions. Keep them separate together with source, date, provider coverage, and completeness, rather than flattening partial evidence into one health score.

A refresh is an asynchronous collection pass. Its accepted result carries the workspace id, queued query count, credit ledger id, provider receipt, receipt, and graph effects, but the useful evidence arrives only after the queued work reaches terminal state.

The refresh capability has a dynamic credit cost in the catalog and the current page documents 0.31 credits per call. Treat that number as a quoted credit effect for this capability, not as a price for every visibility operation.

Cost and comparison

Paid live work uses a quote when the runtime needs a price. Keep the returned estimate and rerun the same parameters with an unchanged max_credits ceiling, because changing the request invalidates the quoted work.

A mutation must state its mode. smoke_test inspects static schema and policy without writes or audit records, dry_run previews against workspace state without business writes, provider spend, or outward action, and live executes.

Ask Architect to read the current overview and citations before it proposes a refresh. Architect uses the visibility skill for measurement practice and canonical discovery controls for the actual capabilities.

When evidence changes a durable claim about the business, record the conclusion and source in Context. Context is the workspace knowledge store; visibility evidence remains measurement material with its coverage and limitations.

A provider can require the workspace’s own connection or key. If provider evidence is unavailable, report the missing connection or unavailable data rather than inventing a citation, rank, or competitor result.

Check the evidence in the AI Visibility screen after a terminal result. A reading is useful only when its bounds travel with the number, and a paid refresh is useful only when it addresses a specific evidence gap.