Define the event you intend to count
A citation benchmark starts with a counting rule, not a target percentage. Decide whether you are studying answers that cite your site, individual links to your pages, or appearances of your brand in cited passages. Those are different events and can move in different directions.
For a simple answer-level measure, count a qualifying answer once when it contains at least one visible citation resolving to an approved domain. Divide qualifying answers by all eligible collected answers. Multiple links in one answer still contribute only one qualifying answer under this definition.
Write the denominator before collection
Eligibility needs an explicit policy. An answer without citations may be a valid observation, while a failed request may not be. An assistant that did not invoke search might belong in the main denominator or in a separate group, depending on the question you are studying.
Consider a purely arithmetic example: if twelve eligible answers include three qualifying answers, the answer-level citation rate is three divided by twelve. These invented counts demonstrate the calculation only; they are not a Kept AI benchmark or a claim about provider performance.
Choose comparisons that answer the same question
Group prompts by intent, language, provider, and collection conditions before comparing results. A troubleshooting question often needs a different source than a product comparison. Combining both can make a change in the prompt mix look like a change in citation performance.
Prefer a stable panel of relevant prompts for repeated reporting. Keep an exploratory panel separate so new questions can be investigated without silently rewriting the baseline. If the provider or collection interface changes, annotate the break rather than pretending the series is fully comparable.
Make the counting procedure reproducible
A reviewer should be able to reconstruct a classification from the saved answer and its visible links. Preserve the original URL as well as any normalized domain. Redirects, subdomains, and syndicated copies can change attribution, so record how each is handled.
Before a reporting cycle, apply this checklist to a small pilot batch. Resolve disagreements in the written rulebook before processing the rest. A rule revised after seeing attractive results needs a new version and, where feasible, a recalculation of earlier observations.
- Save the prompt, provider, collection time, answer, and visible citation URLs.
- Label collection failures separately from eligible answers without citations.
- Have another reviewer classify ambiguous links using the same rule.
Report uncertainty beside the rate
Publish the underlying counts and collection window with any percentage. Repeated answers to the same prompt are related observations, not necessarily independent draws. A narrow interval calculated as though every run were independent can convey more confidence than the study design supports.
Inspect results by prompt before celebrating an overall increase. One frequently repeated question can dominate a pooled average. Show both the pooled result and the distribution across prompts when they answer useful, distinct questions for your audience.
Keep the benchmark within its scope
A visible citation does not prove that a model trained on the page, that a reader clicked it, or that a sale followed. Likewise, an absent citation does not prove the content was ignored. The benchmark describes observable citation behavior under recorded conditions.
Use the result to prioritize a closer source review or a controlled content improvement. Avoid turning a category-specific sample into a universal industry target. The most useful benchmark is one a colleague can audit and whose limitations remain visible when the chart is shared.