A fixed page can show up correctly in an AI answer within minutes, after weeks, or not until a new model is trained. Which one you get depends on where the answer comes from: live retrieval, a search index, or the model's trained knowledge. The public sources we can cite here give no per-URL timeline for ChatGPT or Perplexity. Treat any single number you see for "how fast ChatGPT updates" or "Perplexity recrawl" as unverified, and measure your own page instead.

The three layers an answer can draw on

An AI answer reflects one or more of these. Which ones a given product uses for a given question is not something you can assume.

LayerWhat it isWhat changes itWhat you control
Trained knowledgeWhat the model absorbed during training, up to a cutoff dateA new model versionNothing on a schedule you can plan around
Search indexA stored copy of pages, built by crawlingThe engine recrawling and reprocessing your URLCrawl access, sitemaps, internal links, honest dates
Live retrievalThe system fetching a page at question timeThe fetch itself, so your page as served right nowServer response, access rules, page clarity

A fix to your page cannot change trained knowledge. If an assistant answers from memory with no sources, a corrected page will not change that answer until the model is replaced. If it cites sources, it is using a search index or live fetches, and your page can matter.

Live retrieval is the fastest path, because the system reads the current page. An index adds a delay. The answer can only be as new as the last crawl the index has processed.

What is documented, and what is not

The only timing figures in our sources come from Google, and they describe Google Search crawling. They do not describe ChatGPT, Perplexity, Claude or any other engine. They are still useful as a picture of how crawl-and-index lag works:

  • Google says repeated crawls can happen every few minutes for breaking news homepages, while pages that look unchanged for years may be recrawled about once a month. [1]
  • Google says most sites should expect indexing to take three days or more, and that same-day indexing should not be expected unless the site is newsy or highly time-sensitive. [2]
  • Google says requesting a crawl does not guarantee immediate inclusion in search results, and repeated requests for the same URL do not make it crawl faster. [3]
  • Google says there are no fixed crawl frequencies. Crawl speed depends on site size and how fast crawling is possible, and a medium-sized site move can take a few weeks or more to surface. [4]

The range is wide even within one engine. The same site can have a page recrawled in minutes and another untouched for a month.

For ChatGPT and Perplexity, we have no sourced recrawl interval or freshness guarantee to quote. In our reading, the vendor pages we looked at describe features like live search and cited sources, not a turnaround time for a changed URL. If you need a number, test it.

Why the lag often sits outside the model

When an answer cites a page, the delay between your edit and the answer usually comes from the crawl and index step, not from the language model. That has practical consequences:

How to test with a dated change

You are testing one question: after I change this page, how long until an engine's answer reflects it? Run the test on a page the engine already cites or retrieves.

  1. Pick a testable fact. Choose a harmless, specific detail that a user could ask about, such as a support hours line, a plan limit, or a version number. Avoid anything that would mislead customers if it were wrong.
  2. Record the baseline. Ask the same fixed prompts in each engine and save the full answers, cited URLs and the date. Repeat each prompt a few times, because one run proves little. See why single-run citation percentages are dishonest.
  3. Make the change and stamp it. Change the fact to a new, distinctive value. Add a visible "Last updated" date that matches the edit. Update dateModified and sitemap lastmod only if the content really changed.
  4. Verify what is served. Fetch the page the way a crawler would and confirm the new text is in the response:

``bash curl -sI https://example.com/page | head -n 5 curl -s https://example.com/page | grep -i "new distinctive value" ``

If the text only appears after JavaScript runs, a crawler that does not execute scripts may never see it.

  1. Watch your logs. Note the first request to that URL from each crawler you can identify, with its date.
  2. Re-ask on a schedule. Use the same prompts at intervals, for example daily for the first week and then weekly. Log the date when each of these first appears: - the new value in the answer text, - the page as a cited source, - the old value disappearing.
  3. Write down the result per engine. You get a range for your site, not a rule for the engine.

Interpretation: if the answer is correct but cites nothing, you cannot tell whether the engine fetched your page or guessed. If the old value persists after your logs show a fresh crawl, the answer may be coming from memory or from a cached copy. Neither is a failure of your edit.

Realistic expectations

  • Answers with citations to your page: these can update as soon as the engine fetches or reindexes the page. Plan for a range from minutes to weeks, as the Google figures above show for one index. [1][2][4]
  • Answers with no sources: these will not change from a page edit. Fixing the page helps the next time the system retrieves or the model is retrained, and you cannot schedule either.
  • Large or low-priority sections of a site: expect slower pickup than a single important page. Google ties crawl speed to site size and crawl capacity. [4]
  • Requests for faster handling: a recrawl request is a hint, not a promise. [3]

Do not promise a client or a manager a date. Promise the test and report the measured range.

What to do next

  1. Confirm the engines you care about can reach your pages, using the AI crawler check.
  2. Make sure the corrected text is in the served HTML and the date signals are honest.
  3. Run the dated-change test above on one page and log the results per engine.
  4. Re-run it after your next significant fix. Timelines shift, and your own measurements are the only ones that apply to your site.