Use a four-part evidence model
Direct page and crawl facts
Status codes, metadata, extracted claims, sources, dates, headings, canonical URLs and page inventory.
Timestamped provider results
Exact query, model, response time, returned citations, target position and competing domains.
Disclosed comparative estimates
Scores derived from observable signal groups and published weighting, never presented as vendor rankings.
Private system facts
Undisclosed training inclusion, internal weights, unseen user personalisation and guaranteed future results.
Define citation rate before reporting it
A citation rate needs a fixed denominator. For example: target cited in 8 of 40 provider-query observations during a stated seven-day test window. Report the providers, locations, model versions where available, query set, and repeat count. A percentage without this context invites false comparison.
Design a balanced query set
Cover branded questions, category discovery, comparisons, evidence-seeking questions, definitions, implementation questions, and high-value long-tail decisions. Avoid changing the prompt after every unfavourable result. Version the set and rerun it on a schedule so changes are interpretable.
Record source competition
Count the domains cited across the same observations and classify why they may be preferred: primary data, official status, freshness, directness, format, or topic depth. “Competitor X appeared in 18 observations” is more actionable than “we have low AI authority”.
Keep readiness separate from visibility
A page can be citation-ready but not yet observed, and it can be cited despite weak structure. Readiness measures the content and access conditions a team controls. Visibility records what a provider returned. Present both, but do not silently merge them.
Connect observations to business outcomes
Use tagged landing pages, referral reporting where available, branded search lift, assisted conversions, sales-call mentions, and content engagement. Attribution will be incomplete, so label direct, assisted, and qualitative evidence separately.
A defensible report line
“IndexHalo observed the target domain in 6 of 30 repeatable provider-query checks between 1 and 7 August. The result is an observed citation rate for this test set, not evidence of model training or a prediction of future citations.”
Add governance to the dashboard
- Show the evidence class beside every metric.
- Preserve timestamps and raw returned citation URLs.
- Version query sets and scoring weights.
- Do not replace missing data with a confident-looking zero.
- Keep sample or demonstration data visibly labelled.
- Document provider outages, rate limits, and unconfigured checks.