Skip to content
Chatbotscape
Editorial flat-vector illustration for How to Fix a Chatbot That Gives Outdated Answers (Prices, Policies, Hours)
Published 6 min read

How to Fix a Chatbot That Gives Outdated Answers

Quick answer: First find which copy of the information the bot is reading, because the bot rarely invents an old price; it repeats one. Then re-sync right after each edit, make sure deleted pages leave the index, give every page an owner and a review date, and re-test a saved list of questions after each change. Most fixes take minutes once you know where the stale text lives.

Step 1: Find where the old answer comes from

Ask the bot the question that went wrong and look at what it cites, if it shows sources (see showing sources in chatbot answers). Then check, in this order:

  1. The live page. Is your website itself correct? Fix it first.
  2. The bot's copy. Does the platform's knowledge area show the old text, or an old sync time?
  3. Duplicates. Is the same fact in two places, such as a help article and an uploaded PDF, with only one updated?
  4. The prompt. A price typed into the system prompt never changes when the website does.

If the answer comes from none of these and no source supports it, the cause is not freshness; go to the hallucination guide.

Step 2: Re-sync right after every change

Waiting for a scheduled crawl leaves a gap. Intercom's help article on syncing public URLs says automatic re-syncs usually run weekly depending on source size, that very large or JavaScript-heavy sources move to 14 days, and that a source stops re-syncing automatically once none of its pages has been used in a Fin answer for 90 days, although the old pages stay available. A manual re-sync can be run from the source's settings at any time.

So the rule is simple: when you change a price, policy or hours page, trigger a re-sync yourself and then check the bot. Make this the last item on whoever's change checklist edits the page. Platforms differ; look up your own platform's schedule and manual option rather than assuming.

If you run your own pipeline, re-index only the changed pages where you can (compare each page's modified date or a content hash with what you stored) and run a full crawl on a slower schedule as a safety net. The mechanics of the index are in the build a RAG chatbot guide.

Step 3: Make sure deleted pages really leave

Updates are the easy half. A removed page may stay searchable. Microsoft's documentation for Azure AI Search indexers on Azure Storage says the indexer detects new and changed blobs by their last-modified time but does not track deleted objects, so without a deletion strategy the search documents for deleted files remain in the index as orphans; it recommends a soft-delete approach, and says the policy needs to be in place from the first run. That is one product, but the lesson is general.

For any setup:

  • Find out whether your platform or pipeline removes content when a source page disappears. If the documentation does not say, test it.
  • Test with a harmless page: publish it, let it sync, ask the bot about its unique phrase, delete the page, re-sync, ask again.
  • Keep a list of retired pages (old plans, discontinued products) and spot-check that the bot cannot quote them.
  • When you replace a document rather than edit it, remove the old file and any vectors made from it so old and new passages do not compete. See vector database.

Step 4: Give every page an owner and a review date

Pages that never changed can still be wrong. Add two fields to each article: owner (a named person or role) and last reviewed (a date). Set review periods by how fast the content moves:

ContentSuggested review limit
Prices, plans, stock, hoursCheck on every change; review monthly
Policies (returns, shipping, terms)Check on every change; review quarterly
How-to and product guidesEvery six months
Company backgroundYearly

These limits are starting points for you to adjust, not standards. A monthly meeting that lists pages past their limit is enough for most small teams; the chatbot maintenance cadence guide covers the wider routine.

Step 5: Keep fast-changing facts out of free text

Where you can, store one source of truth for volatile facts and have the bot read it, instead of copying the number into many articles. Examples: one pricing page the bot always retrieves, or a small table the bot looks up through a tool call rather than recalling. Then a price change is one edit. If a platform cannot do this, at least link articles to the single pricing page and avoid repeating figures in blog-style content. The knowledge base basics are in building a chatbot knowledge base.

Step 6: Re-test with saved questions

Keep a short list of the questions where freshness matters (returns window, price of the main plan, opening hours, a discontinued product) with the answer you expect today. After each content change, and weekly regardless, ask them all and compare. Mark any miss and trace it back through Step 1. This is a small golden dataset; the regression testing guide shows how to run it. Customer thumbs-downs help too, but few people rate, so see collecting feedback on chatbot answers for sampling unrated ones.

Check the arithmetic

An invented example, not a measurement. A shop has 50 help articles and a weekly re-sync. On a Tuesday it shortens the returns window from 30 days to 14, one day after the last sync, so the next automatic sync is six days away. The bot answers about 40 returns questions a day.

  • Without a manual re-sync: 6 days × 40 = 240 answers quote the old 30-day window.
  • With a manual re-sync right after the edit: the gap shrinks to the minutes the sync takes, so near zero.

A review audit then finds 12 of the 50 articles not reviewed in 180 days: 12 ÷ 50 = 24%. The team assigns owners and clears those 12 over the next month.

Common mistakes

  • Editing the website and assuming the bot knows.
  • Fixing the article but leaving the old PDF or system-prompt copy.
  • Never testing whether deleted pages are removed.
  • Same price in ten places.
  • No owner, so nobody feels responsible for review.
  • Treating every page alike instead of by how fast it changes.

FAQ

Why does my chatbot give old information? Most often the changed page has not been re-read yet, an old copy of the fact still exists in another document or in the prompt, or a deleted page remains searchable.

How do I update a chatbot knowledge base after a policy change? Fix the live page, trigger a re-sync or re-index, confirm the bot's copy changed, remove any duplicates, and ask a saved test question to confirm.

How often should the knowledge base be refreshed? Right after every change for prices and policies, and on a review schedule for everything else. Automatic schedules, such as weekly on one platform, are a safety net, not a substitute.

Will fine-tuning fix outdated answers? Not reliably. Facts that change belong in content you can edit. See fine-tuning versus RAG.

Do I need to delete old pages from the bot? Yes, and test it. Some systems keep removed content unless a deletion step runs.

Is this the same as the model's training cutoff? No. The cutoff concerns the model's own training text; this guide concerns the documents you supply.

Sources