Why the engine keeps using the old URL
AI engines cite from an index they already hold, plus a few pages fetched at answer time. A URL that was retrievable when indexed stays a candidate until the index is refreshed for that topic. If the page is now a 404, the engine may still cite it, because the index entry, not the page, is what was retrieved.
Search Engine Land's August 2026 description of ChatGPT's retrieval stack shows the mechanism: an index, a cache of previously opened pages, and live fetches. A page that was opened and cached is cited far more often than one only retrieved. An old page with a cache entry can outlive a new page nobody has opened yet.
The two cases need different fixes
| Case | What the reader sees | What fixes it |
|---|---|---|
| Dead URL (404, removed, moved without redirect) | A source link that goes nowhere | A live page on the same topic the engine can re-retrieve, at the same URL or a 301 from it |
| Outdated URL (live, but stale) | A source link to a page with old numbers or an old product name | Update the page in place, keeping the URL, so the trusted index entry inherits the new content |
The second case is common and easier. The engine already trusts the URL. Edit the page it cites; keep the URL and the H1's job, change the body.
What to do, in order
- Confirm the citation. Run the prompt three times on the engine and record the cited URL each time. If the old URL appears on all three, it is the index, not a one-off.
- Check what the old URL returns now: a 404, a soft redirect to the homepage, or a live but stale page.
- If dead and the topic still matters, put a live page back at that URL, or 301 it to the live equivalent. A redirect to the homepage does not count.
- If live and stale, edit in place. Fix the numbers, product names and first paragraph. Leave the URL and the H1's job alone.
- If a third-party page links to your dead URL, ask that site to update the link. Engines often reach your page through theirs.
- Re-run the same prompt weekly. The swap shows as a change in the cited URL column, usually within one to three crawl cycles. Record the date.
What does not force the swap
- Putting the year in the title. "Guide 2026" tells a human it is new. David Konitzny's September 2026 analysis of ChatGPT's search behaviour found a freshness preference that influences retrieval without constraining it; a stale page can still be retrieved, and a date in the H1 does not make a page fresh.
- An llms.txt file. A courtesy index for crawlers, not an instruction to drop a cached URL.
- Publishing a near-duplicate at a new URL. Now the engine has two candidates, one of which it already trusts.
- Blocking the old URL in robots.txt. The engine cannot re-fetch it, so it cannot learn it is gone. Serve a 404 or 410.
How long a citation persists: one observation
In a registered trial on this site, a Perplexity citation for one fixed prompt was absent on the first six observations, then present on nine of the next ten, with one later valid miss. Once an engine has a page in its index it tends to keep citing it, which is why the old page is sticky. The swap is visible only if you run the same prompt repeatedly and log the URL. The result is engine-specific; ChatGPT and Gemini never cited the same prompt in the window.