What Is Citation Share in AI Answers and How to Measure It

Write the eligibility rule before the first run. Mark an observation eligible only when the planned query was completed on the planned answer surface, a synthesized answer was returned, the run belongs to the frozen panel, and the stored evidence can distinguish answer citations from other links or organic results. This rule keeps a provider failure from being counted as an uncited answer and keeps an unknown source list from being treated as verified evidence.

Use an explicit status vocabulary in the raw sheet: eligible and cited, eligible and not cited, no synthesized answer observed, provider error, run not completed, and unknown provenance. Only the first two states enter the answer-level denominator. Preserve every other row with its exclusion reason. Excluded does not mean failed, and missing does not mean zero.

Record the effective answer surface rather than only the requested setting. If the collection system cannot establish which surface produced the stored links, classify provenance as unknown and keep the row out of proof metrics. Never infer provenance later from the presence of a URL because that conditions eligibility on the outcome being measured.

Create the panel from questions tied to real research and evaluation work. For every prompt, record the buyer intent, topic cluster, target geography, selected collection surface, visible configuration label, account state, and planned collection cadence. Keep branded navigational prompts separate from category and comparison prompts because they answer different commercial questions.

Freeze the prompt text and settings before collecting the baseline. If the panel changes, create a new version instead of rewriting prior rows. A versioned panel makes a trend interpretable: the reader can see whether a difference came from new observations or from a changed question set. Save the panel version on every run.

Choose a panel your team can complete consistently. There is no universal prompt count that makes every niche representative. Document what the panel covers, what it omits, and why the selected scope is useful for the decision. Report the completed observations beside the planned observations so partial collection remains visible.

For every eligible answer, save the raw response evidence permitted by your collection policy, the displayed source URLs, the collection timestamp, and the effective answer surface. Preserve the original URL before normalization. The raw value is the receipt; the normalized value is the analysis key.

Normalize URLs with one documented function. Lowercase the hostname, remove the fragment, resolve only redirects you are authorized to request, and apply a written rule for tracking parameters. Do not merge different paths merely because they share a domain. For exact-asset reporting, compare the normalized cited URL with the frozen canonical URL assigned to that asset.

Deduplicate at a declared grain. For answer-level rate, multiple links to the same target domain inside one answer still produce one cited answer. For source citation share, decide whether the event grain is unique normalized URL per answer or every displayed source occurrence, then keep that rule unchanged. Store both URL and registrable domain so later analysis can answer page-level and domain-level questions without reparsing the raw response.

Calculate each metric from the frozen eligible set. Answer-level citation rate equals eligible cited answers divided by eligible answered runs. Source citation share equals target-domain citation events divided by all citation events at the declared event grain. Prompt coverage equals eligible prompts with a target citation divided by eligible prompts with at least one completed answered run.

For a hypothetical panel, three of six eligible answered runs contain an exact target-domain citation, so the answer-level rate is 3 / 6 = 50.0%. For a hypothetical event table, three of ten normalized citation events belong to the target domain, so source citation share is 3 / 10 = 30.0%. For a hypothetical prompt panel, three of four eligible prompts produced a target citation, so prompt coverage is 3 / 4 = 75.0%. Treat every figure here as a synthetic input, not a measured result or market benchmark.

Report results by engine or surface before producing any combined number. A pooled result gives more influence to strata with more eligible observations. If a combined metric is necessary, publish the weighting rule and the eligible count in every stratum. Never replace an empty denominator with zero percent; show the metric as unavailable and explain why.

A later observation can differ after a page was published, updated, or distributed, but timing alone does not establish that the action caused the difference. Label an uncontrolled comparison as observational. Preserve the action date, exposure evidence, and metric window so the result can inform the next test without being promoted as verified lift.

For causal language, use a protocol that freezes assignments and outcomes before exposure, protects control prompts from publication, measures at the assigned cluster or prompt grain, and revalidates the readout from stored evidence. Keep randomized single-brand case studies separate from cross-brand confirmatory evidence. The evidence class belongs beside the verdict.

When the rate changes, inspect the components before rewriting content. Check whether the eligible denominator changed, whether one surface failed more often, whether the prompt panel changed, whether provenance coverage improved, and which exact URLs gained or lost citation events. Route the next action to the stage where the evidence changed.

Citation metrics describe source selection inside the measured answer panel. They do not by themselves establish traffic, leads, or revenue. Build a separate asset-path funnel beginning with the exact cited or distributed URL. Record visits carrying that path, then bind a signup only when the attribution contract permits it, and count first-paid only from a confirmed billing event.

Keep each funnel row independent: published asset, exact citation, asset-path visit, signup bound, and confirmed first-paid. Show a numerator and denominator at every stage. Keep the acquisition-source label separate; use asset-path attributed unless a trusted acquisition field establishes a narrower source.

This separation makes failure diagnosis possible. Citation without visits points to a different question than visits without signup. Signup without first-paid belongs to activation, trust, pricing, or another commercial stage. A blended score hides those distinctions and cannot tell the operator what to change.