Define the monitored sample first
A citation measurement should begin with a versioned portfolio of buyer questions that represent actual commercial, editorial or research decisions. Record the exact wording, language and intended audience. Choose the provider set before collecting the baseline and preserve every provider, model, response time, returned URL and target position.
The resulting percentage describes that monitored portfolio at those times. It does not describe every user, geography, personalisation state, model version or answer format. Calling it “total AI market share” would expand the conclusion beyond the evidence.
Use “observed citation coverage in the fixed monitored sample” or “share of returned citation appearances,” then publish the denominator beside the value.
Keep distinct metrics separate
A composite visibility index can summarise these signals for navigation, but its formula must remain visible. One practical calculation weights citation coverage at 50%, position-weighted presence at 30%, and returned-source share at 20%. The underlying values—not the composite—should drive diagnosis.
Reject unlike-for-like trends
A longitudinal series is comparable only when the buyer-question version and provider set remain fixed. A run with a different question portfolio answers a different measurement question. A run with no verified provider evidence contains no citation outcome. Preserve those runs in history, but exclude them from movement calculations and explain why.
Provider errors should not become “not cited.” Track attempted and verified checks separately. A failed request cannot support a conclusion about the target’s visibility.
Analyse provider movement individually
An aggregate gain can hide a provider loss. For each provider, report current checks, citations, coverage, change, average position and source retention. Source retention compares the domains returned in consecutive runs. A low value signals that the provider’s source pool is volatile even when target coverage stays flat.
Cross-provider source overlap reveals convergence. When several providers repeatedly select the same primary datasets, government pages or expert references, those sources describe an evidence pattern worth understanding. The action is not to copy them. It is to identify the provenance, specificity, format or authority that makes them useful.
Turn query movement into recovery work
Build a question-by-provider matrix. For every combination, retain cited, not cited or unavailable state, target position, target URL and competing domains. Compare the newest run with the previous comparable run to find gains, losses and persistent gaps.
- Recover. A provider stopped citing the target. Inspect the sources that replaced it and the owning page’s access, freshness, evidence and canonical state.
- Establish. No provider cites the target. Strengthen or create the mapped owner around the evidence patterns shared by selected sources.
- Grow. Some providers cite the target. Close the remaining evidence gap without fragmenting the winning page.
- Protect. The target holds broad coverage. Preserve the cited URL, evidence, redirects and internal ownership through future releases.
Maintain a returned-source ledger
Count each domain’s current appearances, change, query reach, provider reach and persistence across runs. Label new, gaining, declining and lost sources. Classify obvious government, academic, reference, documentation and community domains with transparent rules, while leaving ambiguous sources in a general publisher or commercial group.
A Herfindahl–Hirschman Index can describe concentration inside the current returned-source set. It does not measure the entire web. High concentration means a small number of domains dominate this monitored sample, which may increase the value of understanding those evidence patterns.
Make every movement operational
A professional citation report should name the affected buyer question, provider, target page, before and after state, replacement sources, action owner and repeat test. Recovery should use the unchanged question and provider. Protection should preserve the exact cited URL and its supporting evidence. Changed prompts should begin a new baseline rather than rewriting history.
Keep the evidence boundary visible
Timestamped observations prove what those stored responses returned. They do not prove training inclusion, secret ranking logic, permanent visibility or causation. A disciplined observatory is valuable precisely because it refuses to claim more than the evidence can support while still telling a team what to do next.