When AI Cites a Page You Already Moved
A site redesign renames your catalogue URLs and the citations pointing at them quietly stop resolving. What breaks, what survives, and what to check first.
A redesign ships. The catalogue moves from one path to another, the datasheets get re-uploaded with cleaner filenames, and the old product pages are retired because nobody was looking at them anyway. Weeks later an assistant still names your company and links to a page that no longer exists, or worse, stops naming you at all for a query it used to answer with your product. Both symptoms come from the same place: the address changed, and the things that pointed at the old address did not all follow.
Two separate ways a citation dies
It helps to keep these apart, because the remedies differ.
The link is stale but the content survived. The assistant reproduces a URL it absorbed during training, or that it found referenced elsewhere, and that URL now returns a not-found page. The buyer clicks, lands on an error, and forms an opinion about your company that has nothing to do with your products. Your content still exists somewhere on the site, under a different address, but the trail to it is broken.
The content itself became unreachable. Retrieval tried to fetch the page at answer time, got an error or a redirect chain that gave up, and moved on to something else. The assistant then answers from whatever it could reach: a distributor, a marketplace listing, an old press release, a competitor's comparison page. You are not cited because, at the moment it mattered, there was nothing of yours to cite.
The failure modes that actually cause this
Redirects that point everywhere except the equivalent page
The most common migration shortcut is to send every retired URL to the homepage. It avoids error pages, it looks tidy in a report, and it destroys the specific thing that made the old address worth citing. A system fetching a page about a stainless steel pump housing and receiving a corporate homepage has been handed nothing usable. Human visitors react the same way, they just complain more audibly.
Redirect chains and refresh tricks
A permanent redirect, applied once, from the old address to its direct equivalent, is understood by essentially everything that reads the web. A chain of three or four hops is understood by less. A page that returns success and then bounces the browser somewhere else with a script or a meta refresh is understood by less still: an automated fetch may see only the empty intermediate page and take that as your content.
Pages that return success while saying nothing
The quiet one. A retired product page that returns a success status and displays a friendly this item is no longer available message is, to a fetching system, a page that exists and contains almost no information. It stays in indexes, it can still be retrieved, and it contributes an answer that amounts to your product not being available.
Documents that get a new filename at every revision
Technical documentation has a habit of being re-uploaded rather than updated. Each new revision arrives under a slightly different filename, the previous file is deleted, and every reference anyone ever made to it breaks on the same day. Nothing about the content changed meaningfully, but the address did, and addresses are what citations are made of.
What to check on your own site
This is falsifiable in an afternoon and needs nothing but a browser.
- Collect the old addresses. Your previous sitemap, your analytics export from before the migration, or the list of files your team removed. Any of the three is a starting point.
- Open a sample of them. Not all, a spread: the best-selling product, a certification page, a datasheet, a page in each language. What you want to see is a permanent redirect landing on the direct equivalent, not the homepage and not an error.
- Watch for pages that load but say nothing. If a retired page returns a polite empty shell, decide deliberately: redirect it to the successor product, or let it return a genuine not-found status. The ambiguous middle is the worst option.
- Ask an assistant for the source. Query something your site should answer, ask which page it is drawing on, and open the link yourself. This is the fastest way to discover that the address in circulation is one you retired.
- Repeat in every language. Localised sites are usually migrated by different people at different times, and the language versions are where the redirect map most often has holes.
How to move a page without losing what pointed at it
Plan the address before the design. The single most useful habit in technical publishing is a stable URL per product that survives redesigns, with revisions handled inside the document rather than in its filename. When a move is unavoidable, map each old address to its closest equivalent and apply one permanent redirect, not a chain. Keep the old addresses answering for far longer than feels necessary: absorbed references decay slowly, and the cost of keeping a redirect alive is close to nothing.
For downloadable documentation, keep the file at one address and replace its contents when the revision changes, rather than publishing a new file next to it. If your process requires an archive of past revisions, give the archive its own addresses and leave the canonical one where it has always been.
What this fixes and what it does not
Repairing a broken address restores the possibility of being cited. It does not create citations. An assistant may have moved on, may be leaning on a source it now considers better, or may simply not surface your category the way it did before, and no redirect changes that. Being honest about this boundary matters: the fix is real but modest, and anyone promising that a redirect map will restore your presence in AI answers is selling something.
What you can do is watch. PSentry records which sources the assistants lean on when they answer questions about your brand, across ChatGPT, Claude, Gemini and Perplexity and across the languages you sell in, so a source that quietly stops appearing after a migration becomes visible instead of being guessed at. It is measurement, on a scheduled cadence rather than continuous, and it makes no claim to influence what any assistant says.
Frequently Asked Questions
Does a permanent redirect preserve an existing citation?
It preserves reachability, which is the part you control: a system following the old address arrives at your content instead of an error. Whether the assistant keeps naming you is a separate matter, decided by machinery you cannot instruct.
How long do models keep quoting an address I retired?
Longer than you would like, and unpredictably. An address absorbed during training persists until the model is retrained, and references to it on other sites persist until those sites are updated, which for directories and trade listings can be a very long time.
Should I keep discontinued product pages online?
Usually yes, in some form. A page explaining that a product is superseded, naming its successor and keeping the original specifications visible, is far more useful to both buyers and retrieval than a deleted page or an empty placeholder.
Is it acceptable to redirect everything to the homepage during an emergency?
As a temporary measure during an incident, yes. As a migration strategy it is the main way companies lose the specific pages that made them citable, because the equivalence between old and new is never recorded anywhere.
My datasheets are hosted on a separate document platform. Same problem?
The same, and often worse, because those platforms tend to mint a new address per upload by design. If you cannot control the address there, publish the specifications as ordinary text on a page you do control, and treat the platform copy as a convenience rather than the canonical source.