Your AI chatbot not finding answers that are clearly on your website is one of the most frustrating problems after launch. A visitor asks about delivery to Norway, the page exists, and the assistant says it does not know and hands the chat to your team. In almost every case the model is not the problem: the content never reached the assistant, reached it in a form it could not use, or was outranked by something vaguer.
This guide is a troubleshooting checklist. It walks through the common reasons pages get missed, how to find which one applies, and the fix for each, using Talkmio’s assistant, Mio, as the example.
How an AI Chatbot Gets Its Knowledge
A grounded assistant does not know your business by itself. It answers from a knowledge base built from your sources. In Talkmio, those sources are your website (Mio reads every public page on the same domain, respecting robots.txt), single pages, Q&A pairs, free text and uploaded files in PDF, DOCX, TXT, MD, CSV or HTML up to 15 MB. When a visitor asks something, the assistant looks for the most relevant pieces of that knowledge and writes an answer only from them. If nothing relevant is found, it should say so and hand over rather than guess.
So when an answer is missing, there are three possible failures:
- The content is not in the knowledge base at all.
- It is there but not found for this question, because it is buried, vague or contradicted.
- It is found but not usable, because it is out of date, incomplete or ruled out by your instructions.
The checklist below goes through them in order. For background on grounding, see our guide to RAG in AI chatbots.
Step 1: Reproduce the Question in a Test Tool
Start by asking the same question yourself, in the same words the visitor used. In Talkmio, Try Mio shows the answer, which knowledge pieces were used and how long it took. That one view usually tells you which of the three failures you are dealing with:
- No relevant pieces used: the content is missing or not found.
- Wrong or outdated pieces used: a competing source is winning.
- The right piece used but a weak answer: the content itself needs work.
Keep a list of the questions you test. They become your regression set after every content change.
Why an AI Chatbot Is Not Finding Answers: The Common Causes
| Cause | Typical symptom | Fix |
|---|---|---|
| Page blocked in robots.txt | A whole section is never used | Unblock it, or add the content as a file or Q&A pair |
| Content behind a login | Member or customer-area answers missing | Upload the key documents or write Q&A pairs |
| Content on another domain | Help centre or shop on a separate domain ignored | Add it as a separate source or upload exports |
| Information only in images or video | Size charts or price tables never quoted | Add the same facts as text |
| Linked PDFs not uploaded | Terms or price lists missing | Upload the files to the knowledge base |
| Page changed after the last read | Old answer keeps appearing | Re-read the source |
| Plan page limit reached | Newer pages missing on a large site | Prioritise important pages or move to a larger plan |
| Vague or scattered wording | Right page used, weak answer | Rewrite with direct, specific sentences |
| Contradictory sources | Different answers to the same question | Remove or correct the outdated source |
| Instructions too strict | Hand-off even though content exists | Adjust the business instructions |
Content That Never Reached the Knowledge Base
Blocked by robots.txt
Mio respects robots.txt, the standard file that tells crawlers which paths they may read. If a section such as /help/ or /shipping/ is disallowed, perhaps from an old SEO decision, it is not read. Check your file against the robots.txt rules and remember that a change also affects search engines. If you want a page kept out of search but available to the assistant, add its content as a file or Q&A pair instead.
Behind a login
Only public pages are read. Customer portals, member areas and partner pages are invisible to a crawler without an account, which is usually what you want. If visitors ask questions answered only there, decide whether that information should be public in chat at all. If yes, upload the relevant documents or write Q&A pairs; if not, let the assistant hand those questions to your team.
On another domain
The website source covers pages on the same domain. A help centre hosted on a different domain, a shop on a platform subdomain or documentation on a separate site are different sources. Add the important content separately, for example by uploading exported articles as files.
Locked in images, video or downloads
A size chart saved as an image, a price table inside a banner or delivery rules explained only in a video are hard or impossible for a text-based assistant to use. Put the same facts on the page as text, or upload them as a document. PDFs linked from your pages are best uploaded directly too, so you know they are in the knowledge base.
Over the plan’s page limit
Each plan limits how many knowledge-base pages the assistant can keep: 20 on Free, 200 on Pro, 1,000 on Ultimate and more on Enterprise, according to the pricing page. A large site with blog archives, tag pages and product variants can fill the limit with pages that rarely answer questions. Check that your policy, delivery, pricing and product pages are included, and use Q&A pairs for the essentials.
Content That Is There but Not Found
The answer is buried
A delivery rule hidden in the fourth paragraph of a long “About us” page competes with everything else on that page. Assistants find answers more reliably when each page has one clear topic and the answer is stated directly: “We deliver to Norway in 4–6 working days. Shipping costs €19.” Our guide to what to put in a knowledge base explains how to structure content for retrieval.
The visitor uses different words
Customers say “postage”, “shipping”, “delivery” and “how long until it comes”. Your page may only say “dispatch”. Good assistants handle synonyms well, but the more your pages use the words customers use, the better. Read real chat transcripts and borrow their phrasing. Questions in one language can find answers written in another in Talkmio, so a Lithuanian question can use an English FAQ, but translated pages give more natural answers.
A competing source wins
If an old PDF says returns are accepted within 14 days and the website says 30, the assistant may pick the wrong one. Try Mio shows which piece was used. Remove or correct the outdated source rather than adding a third version beside it.
Content That Is Found but Not Usable
Stale content
Sources are re-read on demand. If you changed a page yesterday and did not press Re-read, the assistant may still use the old version. Make re-reading part of every content change. For a routine, see our article on auditing AI chatbot accuracy.
Incomplete answers
A page that says “we ship worldwide, contact us for rates” gives the assistant nothing to answer “how much is shipping to Canada?”. It will rightly hand over. If the question is common, publish the rates or ranges.
Instructions that block the answer
Business instructions tell Mio what it must never promise, such as “never quote delivery dates”. That is valuable, but an instruction written too broadly can stop it answering a general question your page covers well. If the right piece was used but the assistant still handed over, reread your instructions for a rule that is wider than you meant.
A Worked Example
A furniture shop notices that visitors keep asking whether a sofa fits through a 76 cm door, and the assistant hands every one of those chats to the team. The product pages do list dimensions, so the content seems to be there.
In Try Mio, the question returns no relevant pieces. Looking closer, the dimensions appear only in a product image with measurement arrows, and the package sizes sit in a PDF data sheet linked from each page but never uploaded. The fix takes an afternoon: add the assembled and packaged dimensions as text on each product page, upload the data sheets, re-read the website source and ask the question again. This time Try Mio shows the product page as the source, and the answer quotes the largest box size. The next week’s hand-offs on that question drop to the cases that really need a person, such as unusual staircases.
A Weekly Routine to Catch Gaps
- Review hand-offs. In the Inbox, look at conversations Mio handed to the team and note the questions that were answerable from your content.
- Check reports. Talkmio’s reports show the share answered by Mio alone and hand-offs; a sudden drop often follows a site change.
- Test the top questions in Try Mio after every significant website update.
- Fix the source, then re-read it. Edit the page, upload the file or add a Q&A pair, then re-read and test again.
- Keep a log of what you changed and why, so the next person does not reintroduce the problem.
When a Hand-Off Is the Right Answer
Not every missing answer is a bug. A grounded assistant should hand over when the question is about a specific order, when the visitor asks for a person or complains, and when your content genuinely does not cover the question. In Talkmio, those chats get a “needs human” badge, your team is notified, and once a person replies, Mio stays out of that conversation. When nobody is online, the visitor leaves an e-mail and your reply goes out by e-mail. An honest “I’ll connect you with the team” is far better than a confident wrong answer. Our guide to fallback messages covers what the assistant should say in those moments.
Frequently Asked Questions
Why is my AI chatbot not finding answers that are on my website?
Usually the content never reached the knowledge base, or it is hard to retrieve. Common causes are pages blocked in robots.txt, content behind a login or on another domain, facts only shown in images or video, linked PDFs that were not uploaded, pages changed since the last re-read, or a plan page limit filled by less useful pages.
How can I see which content the chatbot used for an answer?
Use a test tool that shows sources. In Talkmio, Try Mio displays the answer, which knowledge pieces were used and how long it took. If no relevant piece appears, the content is missing or not found; if an outdated piece appears, remove or correct that source and re-read it.
Does robots.txt affect what an AI chatbot can read?
It can. Talkmio’s assistant reads every public page on your domain while respecting robots.txt, so disallowed paths are not read. Remember that changing robots.txt also affects search engines. If a page should stay out of search but be available to the assistant, add its content as an uploaded file or Q&A pair instead.
Why does the chatbot still give an old answer after I updated the page?
The source may not have been re-read since your change, or an older document with the previous information is still in the knowledge base. In Talkmio, sources are re-read on demand, so press Re-read after every content change, and delete outdated files rather than adding new versions beside them.
Can an AI chatbot read content inside images, videos or PDFs?
Text-based assistants work best with text. Facts shown only in images, banners or videos are hard or impossible to use, so put them on the page as text as well. PDFs are supported as uploads in Talkmio, along with DOCX, TXT, MD, CSV and HTML files up to 15 MB, so upload important documents directly.
Is it a problem when the chatbot hands a question to my team?
Not when the question is about a specific order, the visitor asks for a person, complains, or your content really does not cover it. That is the assistant working correctly. It is a problem only when the answer exists in your content, so review hand-offs weekly and fix the sources behind repeated questions.
The Bottom Line
An AI chatbot not finding answers is almost always a content problem you can fix: unblocked and public pages, facts written as text, files uploaded, sources re-read after changes and outdated versions removed. Test the questions customers really ask, check which pieces were used, and review hand-offs every week. The assistant can only be as good as what it reads. Try Talkmio free and use Try Mio to see exactly where your content has gaps.
