Phantom URL
- Phantom URLs are fabricated source addresses — the AI invents a plausible path on a real domain.
- The user experiences your broken page, not the AI's error: the brand inherits the dead end.
- Measure from both sides: resolve AI-cited URLs, and watch server logs for 404s with AI referrers.
- Recurring phantom paths can be converted into working entry points with a permanent redirect.
Consensus definition
Citation fabrication is a documented failure mode of generative search engines. In the first systematic audit of citation quality across four generative search engines, only 51.5% of generated sentences were fully supported by their citations, and a quarter of citations did not support the sentence they were attached to.1 Phantom URLs are the sharpest end of that spectrum: the cited address itself is invented. The model assembles a URL that matches the pattern of a brand's site architecture — right domain, plausible path — without verifying that the page exists. The result is a citation that performs authority without providing any.
rhinegold operator refinement
Rhinegold treats phantom URLs as a brand-risk signal, not a curiosity. The user who clicks a phantom URL does not blame the AI — they experience your broken page, at the exact moment the AI positioned you as the answer. Operationally this makes phantom URLs measurable on two sides: in citation-level AI monitoring (which cited addresses return 404?) and in your own server logs (which 404s arrive with AI-platform referrers?). A recurring phantom path is also actionable: if an AI repeatedly invents the same plausible address, a redirect to the real page converts a fabricated citation into a working entry point.
Operational use
Track a phantom rate alongside Citation Rate: of all URLs the AI attributes to your domain, the share that does not resolve. Review server logs for 404s with AI referrers as the ground-truth complement. High-frequency phantom paths are candidates for permanent redirects to the nearest real page.
Measurement boundary
Phantom-URL detection requires citation-level monitoring — answer-level metrics like Mention Rate cannot see it. Detection is also provider-dependent: systems differ in how much they ground answers in live retrieval versus generate from memory, so phantom rates are not comparable across platforms without noting the grounding mechanism behind each.
Distinct from
Against Hallucination, the broader category: hallucination covers any fabricated claim about a brand, while a phantom URL is specifically a fabricated source address. Against a stale citation — a real page that has since been removed or moved. The stale citation once existed and can be redirected from knowledge; the phantom never existed and must be discovered.
Obstacles & resolutions
Observed pattern in practice
Common mistakes
- Treating every AI-cited URL as a real page — citation lists must be resolved, not just counted, or the phantom share silently inflates apparent authority.
- Blaming the 404s on site migration before checking referrers — AI-fabricated paths and genuinely broken links need different fixes.
- Redirecting every phantom path reflexively — a redirect is for recurring, high-traffic phantoms; blanket wildcard redirects mask real architecture problems.
Where consensus is missing
There is no standard phantom-rate metric and no cross-provider study of how often fabricated URLs carry real brand domains. The term itself circulates informally (phantom citation, hallucinated URL, fabricated source); measurement conventions — resolve-time, redirect handling, treatment of soft-404s — are unstandardized.
Sources & deeper reading
- rhinegold citation-resolution methodology — phantom-rate measurement and AI-referrer 404 analysis, shared with engaged clients and partners.
