# How to Control Chatbot Sources on Your Website

> Learn how to control what your website chatbot knows: pick approved sources, block pages, keep prices current, set fallbacks and test answers before visitors see them.

[Back to blog](https://konvuno.com/blog)October 25, 2026

Learn how to control what your website chatbot knows: pick approved sources, block pages, keep prices current, set fallbacks and test answers before visitors see them.

![](https://media.babylovegrowth.ai/blog-images/organization-42511/1792897231204_Owner-selecting-website-sources-for-chatbot.jpeg)

You control a website chatbot’s sources by deciding, in writing, which content it may answer from, then connecting only that content, excluding everything else, keeping it synced, and testing the answers. Four levers do most of the work: an approved list of pages and FAQs, a live product feed, a clear rule for what happens when the sources are silent, and a regular check of real conversations. This guide walks through each one for small business owners, with no developer needed.

Konvuno

Keep Chatbot Answers Tied to Your Content

Konvuno answers visitor questions from your FAQs, website content and product catalog, with product feeds syncing automatically.

[Explore Konvuno](https://konvuno.com/)

## TL;DR

- A chatbot is only as reliable as the material you let it use. Start with a short, approved source list: FAQs, key pages, policies and your product feed.
- Robots.txt and meta tags control crawlers. They do not control your own assistant, which reads only what you connect to it.
- Sync product data from a feed so prices and availability update by themselves, instead of retyping them into a knowledge base.
- Decide what the bot says when it has no answer: admit it, offer a human, and capture the lead.
- Test with real questions before launch and review conversations every week or two.

## Table of Contents

- [Where a website chatbot gets its answers](https://konvuno.com/blog/control-chatbot-sources#where-a-website-chatbot-gets-its-answers)
- [Two different things people call “control”](https://konvuno.com/blog/control-chatbot-sources#two-different-things-people-call-control)
- [Choose the sources before you connect anything](https://konvuno.com/blog/control-chatbot-sources#choose-the-sources-before-you-connect-anything)
- [Keep product prices and availability current](https://konvuno.com/blog/control-chatbot-sources#keep-product-prices-and-availability-current)
- [Block pages and information you do not want used](https://konvuno.com/blog/control-chatbot-sources#block-pages-and-information-you-do-not-want-used)
- [Decide what the bot does when the sources are silent](https://konvuno.com/blog/control-chatbot-sources#decide-what-the-bot-does-when-the-sources-are-silent)
- [Test before visitors do](https://konvuno.com/blog/control-chatbot-sources#test-before-visitors-do)
- [Privacy and security: give it only what it needs](https://konvuno.com/blog/control-chatbot-sources#privacy-and-security-give-it-only-what-it-needs)
- [Perspective: fewer sources, better answers](https://konvuno.com/blog/control-chatbot-sources#perspective-fewer-sources-better-answers)
- [Setting this up with Konvuno](https://konvuno.com/blog/control-chatbot-sources#setting-this-up-with-konvuno)
- [A short checklist](https://konvuno.com/blog/control-chatbot-sources#a-short-checklist)
- [FAQ](https://konvuno.com/blog/control-chatbot-sources#faq)
- [Sources](https://konvuno.com/blog/control-chatbot-sources#sources)

## Where a website chatbot gets its answers

A modern website assistant does not “know” your business. It retrieves material you supplied and writes an answer from it. That material usually comes from four places:

1. **Your FAQs and help content**, typed in or imported.
2. **Your website pages**, such as shipping, returns, services and pricing pages.
3. **Uploaded documents**, such as PDFs, price lists or policy files.
4. **Your product catalog**, often pulled from a store plugin or a product feed.

A general-purpose AI model also has background knowledge from its training. That is why a poorly configured bot can sound confident about things you never told it, such as a refund window you do not offer. The UK Government’s guidance on generative AI says outputs [can be biased or misleading and may vary between repeated prompts](https://www.gov.uk/government/publications/guidance-to-civil-servants-on-use-of-generative-ai/guidance-to-civil-servants-on-use-of-generative-ai), so they need checking. Controlling sources is how you shrink the room for that kind of error.

## Two different things people call “control”

The phrase covers two separate jobs, and mixing them up causes most of the confusion.

**Job one: controlling outside AI crawlers.** Other companies’ bots visit your site to index it or gather content. You manage that with robots.txt and page-level tags. A robots.txt file [tells AI crawlers and search engines which parts of a site they can and cannot access](https://docs.prerender.io/docs/how-to-control-crawling-and-indexing), and meta tags handle the same decision for individual pages.

**Job two: controlling your own assistant.** The chatbot on your site reads only what you connect inside its dashboard. Robots.txt does not govern it, and a page you hide from crawlers is not automatically hidden from your assistant if you added that page as a source.

Both matter, but they are configured in different places. The rest of this guide covers your own assistant first, then crawlers.

## Choose the sources before you connect anything

The temptation is to point the assistant at your whole site and let it sort things out. Resist that. A site accumulates old blog posts, expired promotions, draft pages and contradictory policies, and a bot will happily quote any of them.

Build the list in a spreadsheet first. For each source, write down:

- what it is and where it lives,
- who owns it (the person who fixes it when it is wrong),
- how often it changes,
- whether it is safe for any visitor to read.

### What belongs on the list

Start with the content that answers the questions your team already gets every day: shipping times, return rules, opening hours, service areas, how booking works, what is included in each package. Add a single page per policy, not three versions of it.

### What should stay off

- Old promotions and archived blog posts with outdated prices.
- Internal notes, staff procedures and supplier information.
- Pages with customer data, such as order confirmations or account areas.
- Anything legal or medical you would not want paraphrased loosely. Link to the original instead and let a person handle the nuance.

A good test: if a new employee on their first day would be wrong to read the page aloud to a customer, the bot should not use it.

### Write the FAQ answers yourself

For the 20 to 30 questions that make up most of your inbox, write short, direct answers in your own words. Curated FAQ entries give you precise wording and one place to edit when something changes. Website pages then act as the backup layer for less common questions.

## Keep product prices and availability current

Product answers go stale faster than anything else. If a visitor asks whether a jacket is in stock in medium and the bot answers from a page cached last month, you have either lost a sale or created a complaint.

The fix is not more careful typing. It is removing the typing. Connect the assistant to the same structured data your store already maintains:

- **WooCommerce:** sync the catalog directly so titles, prices, variations and stock status flow through.
- **Google Merchant or Facebook product feeds:** if you already publish one for ads or shopping listings, the same file can feed the assistant, so you maintain one source of truth.

Three habits keep this reliable:

1. **Treat the feed as the owner of product facts.** If a price is wrong in the assistant, fix it in the store or feed, not in a separate FAQ entry that will drift.
2. **Keep prices out of free-text pages where you can.** A blog post that says “our best-selling kit is $49” will outlive the price.
3. **Spot check after every big change.** After a sale, a catalog import or a platform update, ask the bot about three products you changed.

The same logic applies to services. If you publish a price list, keep one current version and remove copies.

## Block pages and information you do not want used

When a page should not feed an answer, you have layered options. Use the one that matches the problem.

### Inside your assistant

Remove the page or document from the source list, or exclude its section. This is the strongest control because it acts directly on what the assistant can retrieve. Also check any setting that crawls your whole domain automatically, and exclude folders such as staging areas, account pages and archives.

### On your website, for outside crawlers

Use robots.txt for broad rules and meta tags for single pages. One detail that trips people up: each AI crawler has its own token, and settings are independent. OpenAI, for example, documents that a site can [allow its search crawler while disallowing its training crawler](https://developers.openai.com/api/docs/bots), because each setting is separate. A [robots.txt template for AI crawlers](https://dblscanner.com/blog/ai-crawler-robots-txt-template-2026) shows how those separate groups are written, so a rule for one bot does not silently fail for another.

![Separate robots.txt controls for AI crawlers](https://media.babylovegrowth.ai/blog-images/organization-42511/1792897279525_Separate-robots.txt-controls-for-AI-crawlers.jpeg)

After editing the file, run it through a [robots.txt validator](https://babylovegrowth.ai/free-tools/robots-txt-checker) to catch syntax mistakes. A typo can block your whole site from search or leave a private folder open.

### What robots.txt does not do

It is a request that well-behaved crawlers follow. It is not a lock. Anything truly confidential belongs behind a login, not behind a disallow line.

## Decide what the bot does when the sources are silent

Every assistant eventually gets a question your sources cannot answer. What happens next is a configuration choice, and it is the one that most affects trust.

A bot that guesses is dangerous. Research on retrieval-based answering shows the trade-off clearly: in one TREC 2025 report, a system that [constrained answers to cited evidence reached citation precision of 0.77 and recall of 0.76](https://trec.nist.gov/pubs/trec34/papers/ii_research.rag.pdf), while its coverage of information was lower. In plain terms, answers tied to sources are more trustworthy but sometimes less complete. For a small business, that is usually the right trade.

A solid fallback has four parts:

1. **Say so plainly.** “I don’t have that information” beats a plausible invention.
2. **Offer a next step.** A contact form, a phone number, or a handoff to a person during business hours.
3. **Capture the lead.** Ask for an email so someone can follow up, with a clear note of what happens to it.
4. **Log the gap.** Every unanswered question is a missing FAQ entry. Add it, and the same question gets answered next time.

Set hard limits for sensitive topics too: medical, legal, financial and safety questions should route to a qualified person, not to a paraphrase.

## Test before visitors do

Testing a chatbot is not the same as testing a form. The same question can produce different wording on different tries, and the UK guidance cited above stresses that [plausible language is not evidence that an answer is correct](https://www.gov.uk/government/publications/guidance-to-civil-servants-on-use-of-generative-ai/guidance-to-civil-servants-on-use-of-generative-ai). So test deliberately.

### A simple pre-launch test plan

- **Real questions.** Pull 30 questions from your inbox, chat logs or phone notes. Ask each one twice, worded differently.
- **Source check.** For each answer, find the line in your content it came from. If you cannot, treat it as a failure.
- **Trap questions.** Ask about things you do not offer, a discount that does not exist, or a competitor’s policy. The right response is a polite refusal or a handoff.
- **Stale-data check.** Ask about three products whose price or stock you recently changed.
- **Edge cases.** Misspellings, a question in another language, and a long rambling message.

NIST’s 2025 crosswalk frames AI testing as [broader than accuracy alone](https://airc.nist.gov/documents/1/20250527-Crosswalk_NIST_600-1_IMDA_AI_Verify.pdf), covering data governance, transparency, robustness, safety, fairness and accountability. For a small site that translates to: does it stay on topic, resist odd prompts, avoid leaking data, and have a person responsible for fixing mistakes?

### Fixing a wrong answer

Find the cause before you patch it:

- **Wrong source used:** remove or correct that page.
- **Right source, wrong wording:** add a curated FAQ entry that states it exactly.
- **No source at all:** add the missing content, or set the fallback.
- **Conflicting sources:** keep one and delete the other.

Then ask the same question again, twice, to confirm.

### Review on a schedule

After launch, read a sample of real conversations every week for the first month, then every two weeks. Look at what visitors ask, where the bot declined, and where a human had to step in. That analytics trail is the most useful part of the whole setup, because it shows what your customers actually want to know.

## Privacy and security: give it only what it needs

Source control is also data control. Before connecting anything, ask what the provider does with it. The UK’s National Cyber Security Centre advises considering whether prompts are [visible to the provider, stored, or used to develop its service](https://www.ncsc.gov.uk/sites/default/files/pdfs/blog/chatgpt-and-large-language-models-whats-the-risk.pdf), and recommends against putting sensitive information into public language models. Read the terms and privacy policy of any tool you use, and ask about retention.

The UK Government’s generative AI framework adds two practical principles: an AI tool should [access only the organisational data needed for its task](https://assets.publishing.service.gov.uk/media/656ee89a0f12ef07a53e01fe/Generative_AI_Framework_for_HMG.pdf), and you should understand where data is processed and who else can reach it. The same framework recommends technical safeguards such as content filtering and validation checks so responses are accurate and do not leak data.

In practice:

- Do not upload customer lists, invoices or anything with personal data as a knowledge source.
- Keep internal documents out of the source list entirely.
- Limit who in your team can edit sources and change settings.
- Tell visitors they are talking to an AI.

That last point is increasingly a legal matter. The European Commission says that from 2 August 2026, [EU transparency rules help people recognize when they are interacting with AI](https://commission.europa.eu/strategy-and-policy/priorities-2019-2024/europe-fit-digital-age/excellence-and-trust-artificial-intelligence_en) and require specified AI-generated content to be labeled. If you serve EU visitors, check how that applies to you. Guides on [disclosing AI interactions](https://marcy.com/ai-chatbot-for-local-business-2026-integration-guide) explain simple ways to label a chat widget for local businesses. A clear greeting such as “I’m an AI assistant” costs nothing and builds trust anywhere.

## Perspective: fewer sources, better answers

We build Konvuno for owners who do not have a support team, and the pattern we see in the setup is consistent: the businesses that get the best results are not the ones that connect the most content. They are the ones that connect the right content and keep it honest.

![Perspective: fewer sources, better answers — overview diagram](https://media.babylovegrowth.ai/blog-images/organization-42511/1792897261266_Perspective-fewer-sources-better-answers-overview-diagram.jpeg)

Our view is that control should sit with the owner, not with a developer. That is why everything is configured from a hosted dashboard, with unlimited configuration and a preview before you pay. You can see exactly what the assistant answers, change a source, and ask again, without touching code. We also believe product facts belong in the feed, not in anyone’s memory, which is why catalogs from WooCommerce or a Google Merchant or Facebook feed stay in sync on their own.

And we think a good assistant knows when to stop. Handing off to a human, and capturing the lead on the way, is not a failure of automation. It is what a good front desk does.

> _— Konstantin_

## Setting this up with Konvuno

If you want these controls without building them yourself, Konvuno gives you a single script tag for your site and a dashboard for sources, conversations, leads and analytics on what visitors ask. Every plan includes full conversational catalog capabilities. Plans are Starter at €29 per month, Growth at €79 per month and Business at €199 per month, with no per-seat pricing. You can [see the plans and start configuring](https://konvuno.com/) and preview your assistant before you pay.

## A short checklist

1. List your approved sources and their owners.
2. Write curated answers for your top 20 to 30 questions.
3. Connect your product feed instead of retyping product facts.
4. Exclude archives, drafts, account pages and internal files.
5. Set robots.txt and meta tags for outside crawlers, and validate the file.
6. Configure the fallback: admit, hand off, capture the lead.
7. Run the test plan, then review real conversations on a schedule.
8. Disclose that visitors are talking to an AI.

## FAQ

### Can a chatbot use only the information I give it?

Yes, if it is configured to answer from your connected sources and to decline when they do not cover a question. Underlying models still have general background knowledge, so test with trap questions about things you do not offer. If the bot invents an answer, tighten the sources and the fallback rule.

### Does robots.txt stop my own website chatbot from reading a page?

No. Robots.txt tells outside crawlers what they may access. Your own assistant reads whatever you connect in its settings. To keep a page out of its answers, remove it from the source list or exclude it there.

### How do I keep chatbot answers about prices and stock accurate?

Sync the assistant to your product catalog, such as a WooCommerce store or a Google Merchant or Facebook feed, rather than typing prices into FAQs. Then spot check a few changed products after sales or imports. One source of truth means one place to fix mistakes.

### What should a chatbot say when it does not know?

It should say it does not have that information, offer a person or a contact route, and ask for the visitor’s details if appropriate. Then add the missing answer to your sources so the gap closes. Guessing is the worst option.

### How often should I review what the chatbot says?

Read a sample of conversations weekly for the first month, then every two weeks or after any major change to your site, prices or policies. Look for declined questions, wrong answers and repeated topics that deserve their own FAQ entry.

## Sources

- [How to control crawling and indexing](https://docs.prerender.io/docs/how-to-control-crawling-and-indexing)
- [Overview of OpenAI Crawlers](https://developers.openai.com/api/docs/bots)
- [Guidance to civil servants on use of generative AI, GOV.UK](https://www.gov.uk/government/publications/guidance-to-civil-servants-on-use-of-generative-ai/guidance-to-civil-servants-on-use-of-generative-ai)
- [IIUoT at TREC 2025 Retrieval-Augmented Generation Track](https://trec.nist.gov/pubs/trec34/papers/ii_research.rag.pdf)
- [NIST crosswalk, NIST 600-1 and IMDA AI Verify](https://airc.nist.gov/documents/1/20250527-Crosswalk_NIST_600-1_IMDA_AI_Verify.pdf)
- [ChatGPT and large language models: what’s the risk? (NCSC)](https://www.ncsc.gov.uk/sites/default/files/pdfs/blog/chatgpt-and-large-language-models-whats-the-risk.pdf)
- [Generative AI framework for UK Government](https://assets.publishing.service.gov.uk/media/656ee89a0f12ef07a53e01fe/Generative_AI_Framework_for_HMG.pdf)
- [Artificial Intelligence, European Commission](https://commission.europa.eu/strategy-and-policy/priorities-2019-2024/europe-fit-digital-age/excellence-and-trust-artificial-intelligence_en)
- [A robots.txt template for AI crawlers in 2026](https://dblscanner.com/blog/ai-crawler-robots-txt-template-2026)
- [Robots.txt Validator & Checker](https://babylovegrowth.ai/free-tools/robots-txt-checker)
- [AI Chatbot for Local Business: 2026 Integration Guide](https://marcy.com/ai-chatbot-for-local-business-2026-integration-guide)

---

Canonical: https://konvuno.com/blog/control-chatbot-sources
Index: https://konvuno.com/llms.txt