Where a superseded answer is still coming from

You updated a page days ago, the site shows the new wording, and the assistant keeps giving the old one.

In rough order of how often each one turns out to be the answer. Work down rather than across: each carries a way to tell whether it is yours before you change anything.

• 1

  The answer cache is still holding the old reply

  Why

  Each assistant keeps a small cache of previous answers, matched against new questions by meaning at a very high similarity, holding at most 200 entries and expiring after a day. It is cleared automatically whenever a source finishes indexing something new, so the window this can happen in is between editing your page and re-indexing it, not after.

  How to confirm it is this one

  Ask the same thing again in substantially different words. The match has to be very close, so a genuine rephrasing misses the cache. If the reworded question gives the new answer and the original wording gives the old one, it is the cache and nothing else.

  Fix

  Re-index the source, which clears it, or wait out the day. Note the exact condition: the cache is cleared when a sync actually indexes new text, so a re-crawl finding nothing changed leaves it in place.

• 2

  The re-index has not run since you made the change

  Why

  A crawl indexes a site as it was on the day it ran. Re-crawls are scheduled weekly by default, can be set to daily, and can be turned off entirely, so an edit made the day after a run waits for the next one.

  How to confirm it is this one

  The line under the source name on the Knowledge tab carries both facts: when it last synced, and the schedule it re-crawls on. Compare the last sync against when you made the edit.

  Fix

  Press Resync on that source. Unchanged pages are skipped by their text, so a re-crawl of a large site after a small edit is close to free, which is why the default is weekly rather than never.

• 3

  Both versions are indexed, and the old one is still being retrieved

  Why

  Documents are identified by a hash of their extracted text. A re-crawl of an edited page produces text that does not match, so it is stored as a new document, and the previous one stays where it is. The assistant now holds two versions of one page and retrieval picks whichever matches the question better, which is often the old one.

  How to confirm it is this one

  Compare the source's document count before and after a resync. If it went up while you added no new pages, the increase is edited pages arriving alongside their older selves. The numbered source under the answer in the widget links to the page, so you can also watch it point at the right address while quoting text that is no longer on it.

  Fix

  Delete the source and re-crawl it. Deleting a source removes its documents with it, so the rebuilt source holds one copy of every page. This is the one case where deleting and starting again is the correct move rather than a superstition.

• 4

  The edit did not change the extracted text at all

  Why

  Extraction keeps headings, paragraphs, lists, tables, quotes and code, and drops navigation, scripts, comments and anything marked hidden. An edit confined to what gets dropped leaves the text identical, the hash identical, and the document skipped as already indexed.

  How to confirm it is this one

  The opposite of the previous cause: the document count does not move after a resync. View source and check where your change actually landed. Edits inside collapsed panels, banners marked hidden, or navigation labels are the usual ones.

  Fix

  Put the change where the extractor can see it, in the body of the page. If the content genuinely lives in a collapsed panel, move it out or add the same text as an uploaded file.

• 5

  What did not update is appearance, not knowledge

  Why

  The widget's configuration, meaning colours, labels, the launcher, the allowed sites and whether the assistant is active, is cached for a minute. Answers do not come from it at all, which is worth stating because a minute of delay on one thing teaches people to expect it on everything.

  How to confirm it is this one

  Ask what actually changed. If it is something on the Appearance or Install tab, that is the one-minute cache and it resolves itself. If it is the wording of an answer, that cache is not involved and the four causes above are the list.

  Fix

  Wait a minute and hard refresh. Nothing else to do, and knowing this saves you from re-crawling a site to fix a button colour.

If the old wording survives a delete and a full re-crawl, the page is probably still serving it to the crawler. Fetching carries no session and no cookies, so a copy cached in front of your site, or a version served only to signed-in visitors, is what gets read. Load the page in a private window with no cookies and read what comes back before assuming the index is at fault.

How long can a cached answer survive?

A day at most, and usually far less, because indexing anything new for that assistant clears it. The cache holds at most 200 entries per assistant and only ever holds answers that cited a source.

Does re-crawling cost me pages against my plan?

Only for pages whose text has actually changed, since unchanged ones are skipped before anything is stored. That is what makes a daily schedule affordable on a site that changes weekly.

Is there a way to remove one page without deleting the source?

Not from the source card, which is why the delete-and-re-crawl route is the honest answer for a page whose old copy is still being quoted. On a large site, narrowing the crawl and rebuilding it is quicker than it sounds, because most of the work is skipped.