Public Methodology
A ghost citation occurs when a customer-owned URL is cited in an answer but the customer brand is not named in the prose. The diagnosis logic shipped in WS4 records that event as a first-class failure mode because citation without attribution can still send the recommendation to a competitor.
1. Resolve owned domains for the workspace.
2. Inspect replica citations persisted in evidence_cited.
3. Confirm at least one owned URL was cited.
4. Confirm the brand entity was not mentioned in the extracted answer entities.
5. Persist a row in ghost_citation_events with cited URL, query text, competitors mentioned, and answer excerpt.
Code path: packages/monitor/src/geo_monitor/services/diagnosis.py.
Seer Interactive's March 24, 2026 article LLM Ghost Citations: Why Your Content Is Working and Your Brand Isn't reports analysis across 541,213 LLM responses, 20 brands, and 6 AI platforms. A widely cited summary of that dataset reports citation rates of 53.1% when the brand is named in the answer versus 10.6% when it is absent. We treat that as industry benchmark data, not as a replacement for each workspace's own observed event stream.
When ghost-citation rate is elevated, RAID shifts toward grammatical_subject_anchoring. The intent is simple: attach the claim to the brand in the sentence itself so the model cannot quote the evidence without also carrying brand attribution forward.