Cited Source URLs vs. Uncited Pages: A Distribution Mapping Playbook

Start with the AI Citation Source Map CSV template. It contains the observation, owned-page, routing, receipt, and follow-up fields used in this playbook. The example rows use reserved `example.com` domains and are marked `synthetic`; delete them before collection.

Create one observation row per query, answer surface, run, and cited URL. If one answer exposes three source URLs, store three rows under the same run_id. If the answer exposes no source URL, store one row with cited_url=none; an empty cell is ambiguous and should not stand in for an observed absence.

Version the map instead of overwriting it. Document the map_version, timezone, geography, provider, model or interface, search mode, and exact query text. A later refresh receives a new version while the original remains unchanged.

Run each frozen query on each selected answer surface. Record only URLs the interface exposes as sources; do not infer a citation from an unlinked brand mention or from text that resembles an owned page. Inspect the citation or Sources view in ChatGPT Search with OpenAI's search guide ↗. Inspect Perplexity source links with the Perplexity Help Center ↗.

Save the answer itself as a receipt: screenshot, export, or immutable internal capture. The receipt must show the query, visible sources, interface, and collection time. If you rerun a query, issue a new run_id and preserve both observations rather than replacing the earlier row.

Complete the frozen panel without editing query wording midway. If a provider fails, record the provider failure and retry policy separately; do not convert a failed run into cited_url=none.

Classify the query-level state only after the provider sweep is complete. Apply the labels in order: no relevant owned page means Absent-No-Page; an unavailable, non-indexed, unclear, or structurally incomplete page means Absent-Structural; a relevant and review-ready page that was not cited means Absent-Fixable; an exposed source URL on your domain means Present.

The label describes what you observed and what work is admissible next. It does not explain why an answer engine selected one source instead of another. Keep hypotheses such as authority, format, freshness, or distribution in a separate notes field until an action produces a new observation.

If several owned pages could answer the same query, select one canonical candidate before assigning a lane. Competing internal candidates create an ambiguous target and should be resolved as a structural issue.

The rows below are synthetic and use reserved example domains. They demonstrate how the same collection protocol produces different work queues. The cited source type and owned-page state determine the route; the mere absence of your domain does not.

Notice that only the first row enters distribution immediately. The second needs structural repair, the third needs a page, and the fourth requires no acquisition action. This prevents a team from pitching publishers for a URL that is not yet eligible to represent the query.

For each Absent-Fixable row, route from the type of source that is currently exposed and the target surface you can legitimately influence. Select one lane, one target URL, one owner, and one required receipt. A list of vague ideas is not an action plan.

Use approval-first execution for any action that leaves your systems. The map may prepare a draft or identify a target, but a person with the appropriate account and context should approve editorial requests, community participation, or partner publication.

If the answer exposed no source URL, do not invent a distribution target. Route the row to direct-answer and technical review, then collect a new observation after the documented change.

A completed action proves delivery, not impact. A published asset, sent pitch, approved community response, or live partner update should receive its own receipt. Keep that receipt separate from the later answer-engine observation.

Choose the review date before executing the action. Thirty days is a practical first operating window for this template, not a universal latency promise or a causal window. At review, rerun the frozen query on the same surface and interface mode, assign a new run_id, and preserve the previous version.

If the cited URL changes, write `observed_shift` in `followup_observation`. Do not write `caused_lift` unless the action belongs to a preregistered controlled design with an eligible readout. For a broader measurement protocol, use the quarterly AI citation review workbook.

Do not force a comparison when the observation contract changed. Mark the break, preserve both versions, and restart the affected panel. A smaller comparable map is more useful than a large table that mixes different questions, surfaces, or evidence rules.

Archive completed queries instead of deleting them. Retire a query from the active queue when the team has no current action to execute, while keeping its observations available for later durability checks. Add new queries through the same admission step used at the start.

If the map repeatedly produces only Absent-Structural and Absent-No-Page rows, pause distribution work. The evidence is routing the team back to technical repair and content creation; continuing editorial requests would only create activity without an eligible asset.