Citation tracking guide
How to track AI citations
Build a source history that can be audited. Preserve the original response, normalize citations without erasing their observed form, and investigate changes before assigning a cause.
Quick answer
Quick answer
Define what the answer surface exposes as a citation, save its prompt and response context, extract and normalize URLs, classify source ownership, compare like-for-like observations, investigate gains or losses and log any action without claiming causation. Keep mention metrics and citation metrics separate.
1. Define a citation event
Document the surfaces and fields that qualify: provider citation metadata, a visible source card or an attributable link associated with the response. State how redirects, unsupported protocols and inaccessible sources are handled. Apply the rule before analyzing brands.
2. Save the prompt, response context and timestamp
Store the prompt key and version, engine, model when available, retrieval mode, locale, response text, raw citation payload, timestamp and completion status. Keep raw observations immutable so improved extraction can be rerun later.
3. Extract and normalize cited URLs
Preserve the observed URL, then create a separate normalized URL and registrable domain. Remove fragments and approved tracking parameters, standardize host casing and record redirect targets. Do not discard page paths when page-level persistence matters.
4. Classify source ownership and type
Match domains to reviewed lists for owned properties and named competitors. Classify independent sources by useful types—editorial, review, directory, research, community or other—without treating a listing as positive coverage. Unknowns remain unknown until reviewed.
5. Compare citations across runs and engines
Compare the same prompt version and observation mode. Report unique domains, cited pages, owned-domain citation rate and source persistence. Segment by engine because a source appearing in one system does not imply it is available or selected in another.
6. Investigate gains, losses and replacements
Review access status, page changes, the cited passage, competing sources, provider release notes and repeated runs. A replacement may be more accurate, more recent or simply sampled differently. Use a review queue for material or ambiguous changes.
7. Record actions without assuming causation
Link each proposed action to the observation that motivated it, then record approval, execution and later outcomes. Say that a change preceded an observation unless the study design supports a stronger conclusion. This protects the monitoring history from becoming a collection of post-hoc success stories.
Primary sources and further reading
- Overview of OpenAI crawlers — OpenAI
- Perplexity crawlers — Perplexity
- General structured data guidelines — Google Search Central