Count a retrieved URL and a cited URL as different events. A provider can consult more sources than it explicitly cites, and a brand can be mentioned without its own site appearing in either set.
Keep three linked records
The retrieval record lists sources consulted when that information is available. The citation record stores explicit source references attached to the answer. The brand record stores mentions or recommendations in the answer text.
OpenAI's web search documentation distinguishes its consulted-source list from inline citations. Do not apply that exact payload contract to another provider without checking that provider's documentation.
A worked counting example
Illustrative example: a response exposes six consulted URLs, cites two of them in the answer and mentions three brands. One of the brands has no owned-domain URL among the consulted sources.
You have six retrieved URLs, two cited URLs and three mentioned brands under the stated counting rules. You do not have six citations, and you cannot conclude that every mentioned brand's own website was retrieved.
Group counts at the intended level: unique URL, page identity, domain or observation. Label each chart so readers do not compare units that mean different things.
Preserve the relationships
- Keep the prompt and provider surface with every record.
- Link source entries to the originating execution.
- Distinguish structured citations from plain URLs typed in prose.
- Preserve the answer text and any available citation spans.
- Mark a missing retrieval list as unavailable, not as zero sources.
Some integrations save only a flattened source array. If you cannot establish whether it represents retrieved or cited material, use a neutral label and document the limitation. Do not retrofit a more precise claim from the field name alone.
Use the right evidence for the decision
A large retrieval set can help explain research breadth but does not prove citation quality. A small cited set can still support several claims. Open a consequential cited page before making a support judgment.
For content planning, investigate why an available source fits the question. Do not treat an uncited retrieved URL as a confirmed endorsement or use retrieval frequency as a substitute for market demand.
The AI citation guide explains the answer-level concept. Next, inspect one observation in the citation tracking workflow, label the available source evidence accurately and keep unknown source types out of a citation total.
Fact-checked: 17 October 2026. Examples are illustrative; no customer results are reported.