AI cites an old URL: a canonical and redirect audit
Trace an unexpected AI citation through redirects and canonical signals, fix evidenced URL problems and verify visibility separately.
If an AI answer cites an old or unexpected URL, preserve the exact link and check where it leads before changing content. The visible citation, redirect destination, declared canonical and search engine’s selected canonical can be different URLs. A canonical tag is not an instruction that forces every answer engine to cite your preferred page.
Start with the citation, without a site-wide rewrite
Our suggested audit begins with one observed answer. Save the question, engine and surface, date, language, cited link and the claim it supports. Open that link and record the final destination. If the old address redirects correctly to the current evidence, the action differs from a link that ends on a missing page or an unrelated homepage.
Keep two conclusions separate: whether the reader reaches the right evidence, and whether a technical signal needs repair. A changed URL is not automatically a lost citation, and a working redirect does not prove that the answer used the latest facts.
Compare four addresses
- Observed citation: the link displayed in the answer, including any parameters.
- Final destination: the URL reached after redirects, with the actual status and visible content.
- Declared canonical: the preferred URL stated in the page HTML or response header.
- Google-selected canonical: the URL reported by Search Console URL Inspection, where you have access. This observation applies only to Google.
Google documents redirects and rel="canonical" as strong canonicalization signals, and sitemap inclusion as a weaker one. They should agree with your intended destination. Google may still select a different canonical; its troubleshooting guide recommends checking the selected URL and technical inconsistencies.
An example: a renamed evidence page
Consider a hypothetical guide moved from /research/old-guide/ to /research/current-guide/. The old address still appears in an AI answer. If it permanently redirects to the relevant current guide, record a working reader journey. Then check that internal links and the sitemap point to the current guide and that its canonical matches.
If the old URL returns 200 with an obsolete copy, decide whether it is still a distinct resource. When retiring a genuine duplicate, use an appropriate permanent redirect. If both pages serve different purposes, preserve their differences and make their scope clear. Do not redirect unrelated research to a homepage just to remove an old address.
A practical decision log
- Wrong destination or broken page: repair the evidenced routing problem and retest the full path.
- Conflicting canonical, links or sitemap: align the signals for the intended equivalent content.
- Correct destination but outdated claim: update the source evidence and record the factual change separately.
- Correct source and no evidenced defect: retain the observation; avoid a speculative site-wide change.
For each change, record the before-and-after URLs, the observed status, the responsible person and a review date. This is our recommended handover format. No provider requires it, and it carries no guarantee of a new citation.
Verify technical completion and visibility separately
After deployment, test the public URL rather than relying on the repository. Check the article, redirect path and metadata. For Google’s AI features, its documentation says normal Search eligibility applies, including indexing and snippet eligibility; eligibility does not ensure inclusion.
Repeat a comparable set of questions after the change and save both the answer and link destination. Report a fixed redirect as a technical result. Only report a citation change when observed, and do not attribute it to the fix without evidence that supports that conclusion. These Google references do not establish how ChatGPT or other providers resolve canonicals.
Documentation reviewed on 19 October 2026. This is a documentary guide and suggested audit workflow. It does not include live tests of every answer engine.