Why do AI answers still cite pages I deleted?

99Visibility Research · Updated October 21, 2026 · 5 min read

Short answer

AI answers cite deleted pages for two reasons. First, the model may have learned the page during training, and that knowledge stays frozen at its cutoff date. Second, a search index the engine draws on may still hold the old copy because its crawler hasn't revisited the URL. You can fix the second cause by returning a proper status code and keeping crawlers able to see it. You can't edit the first.

Key takeaways

A deleted page can live on in a model's training data, which no publisher can change until a newer model ships.
A deleted page can also live on in a search index that hasn't recrawled the URL, and that part is fixable.
Return a real error status or a redirect, never a 200 page that says 'not found'.
Keep retrieval crawlers allowed at both robots.txt and the firewall, or they can't see that the page is gone.
Check Google and Bing separately, because each engine only cites what its own index holds.

Why a deleted page keeps showing up

An AI answer can come from two places, and a deleted page can survive in either one.

The first is parametric knowledge: what the model learned in training, with no web access. That knowledge is frozen at a cutoff date. If your page existed before the cutoff, the model may still repeat it, with or without a link.

The second is a search index. When an engine searches the web for an answer, it pulls from an index its crawler built earlier. If the crawler hasn't come back to your URL since you deleted it, the index still holds the old copy, and the engine can still cite it.

The two causes need different responses. The index problem is in your control: you tell crawlers the page is gone and make sure they can hear you. The training problem isn't. You wait for a newer model and make sure the current facts are easy to find when the engine does search. Our playbook on correcting outdated facts past a knowledge cutoff covers that second path in detail.

Cause one: the model learned the page before its cutoff

OpenAI's model documentation lists a knowledge cutoff of 16 February 2026 for the GPT-5.6 family. The Instant path (chat-latest) carries an older cutoff of 31 August 2025. So "what ChatGPT knows" isn't one fixed thing. It depends on which model answers.

When an engine answers without searching, the cutoff is the only thing that governs the answer. Nothing you publish or delete today reaches that answer. Retrieval is also discretionary: ChatGPT decides per query whether to search, and Claude.ai's search is a user toggle. You often can't tell which mode produced a given answer unless it shows citations.

What that means for you: if the answer has no live link to your deleted URL, it's probably parametric. The fix is to make the current version of the fact easy to retrieve, so answers that do search get it right. Our guide to ChatGPT's trained knowledge versus search explains how to tell the two apart.

Cause two: an index still holds the old copy

If the answer links to your deleted URL, an index is the likely source. Google states that a page must be "indexed and eligible to be shown in Google Search with a snippet" to appear as a supporting link in AI Overviews or AI Mode. The reverse also holds: once the page leaves the index, those surfaces lose the basis for citing it.

Otterly's 2026 experiment shows how separate the indexes are. Google deindexed two test sites, and Google's AI surfaces almost stopped citing them. Bing kept both sites, and Microsoft Copilot, which grounds on Bing, kept citing them. Otterly tracked 2,718 of capybara-world.com's 3,934 AI citations to Copilot.

The lesson for deleted pages: one engine can drop your URL while another keeps quoting it. Each index updates on its own crawler's schedule. Our answer on whether Copilot can cite a site that isn't in Bing's index goes further into the Bing side.

What to do next

These steps cover the part you control.

Return a proper status code. If the page is gone, the URL should return an error status such as 404 or 410. If a replacement exists, use a 301 redirect to it. A page that says "not found" but returns 200 looks like a live page to a crawler.
Use noindex for pages you keep but don't want shown. Google and Bing document noindex, nosnippet and max-snippet:0 as the on-page controls that keep content out of AI answers.
Keep retrieval crawlers allowed. A crawler that can't fetch the URL can't see that it's gone. Check robots.txt per bot: OpenAI states that each of its crawler settings is independent, and OAI-SearchBot controls ChatGPT search, not GPTBot.
Check the firewall too. A CDN or firewall can return 403 to a crawler that robots.txt allows. The AI Crawler Checker requests a page as each AI crawler and shows which ones get turned away.
Check each index separately. Use Google Search Console and Bing Webmaster Tools. One being clean tells you nothing about the other.
Look for copies elsewhere. If an answer cites a third-party page that quoted your old content, the fix is on that page. See why AI cites weak or outdated sources about your brand.

Edge cases and what you can't control

Some fetchers behave differently. As of October 2026, OpenAI says robots.txt "may not apply" to ChatGPT-User, and Perplexity says Perplexity-User generally ignores robots.txt. These fetch pages when a user triggers them, so they'll hit your error status directly.

No engine publishes how often it recrawls a URL. You can't force a recrawl schedule, and you can't force a newer model release. What you can do is make the gone signal clear, keep the door open for crawlers, and verify each index. For a quick snapshot of whether AI answers name you correctly today, the free checker on this site runs up to three buyer questions. If Perplexity keeps quoting a dead source, our guide to correcting Perplexity answers covers its specifics.

Frequently asked questions

Will blocking AI crawlers make them forget a deleted page faster?

Blocking works against you here. A crawler that can't fetch the URL can't see the 404, 410 or redirect, so the index has no new signal to act on. Blocking also doesn't touch what a model learned in training. Keep retrieval bots such as OAI-SearchBot, PerplexityBot and Claude-SearchBot allowed, and let them see that the page is gone.

Should I use a 404, a 410 or a redirect?

Choose by intent. Use a 301 redirect when a replacement page covers the same topic. Use 404 or 410 when the content is simply gone. No source we rely on measures whether one code clears an AI index faster than another, so don't pick on speed claims. The real failure to avoid is a 200 response on a page that says it's missing.

Does pinging IndexNow make AI engines drop the page faster?

There's no evidence it does. IndexNow notifies Bing and a handful of other engines, and Google doesn't take part. The specification never mentions AI, and a successful response only confirms the ping was received. It's harmless Bing-side hygiene, but don't expect it to change AI answers on a schedule.

How long until AI answers stop citing the deleted page?

Nobody publishes a timeline. Index-based citations change after each engine's crawler revisits the URL, which differs by engine and isn't disclosed. Answers from training data change only when a model with a later cutoff replaces the current one. Check Google Search Console and Bing Webmaster Tools to see where each index stands.

Sources

This page is AI-assisted. It was written from a verified evidence base and checked against our editorial policy. Engine behaviour changes often; facts are as of October 21, 2026.