AI Chatbot: Training Best Practices Thread

How about starting a thread on best practices to train the AI chatbot?

Here’s what seems to be working so far — we’d love to hear what others are doing.

On Which Website Pages Should You Install the AI Chatbot?

On our ecommerce site, we’ve installed the AI chatbot on every page except checkout and forms pages — specifically, not where customers enter information like their address, payment details, or order tracking requests.

The idea is to avoid distractions and maintain visitor focus on those pages.

Are other stores following a similar approach? What’s your strategy?


What Chatbot Greeting Are You Using?

For our AI chatbot, we’re currently testing this simple greeting:

Got questions?
I’m here to help!

Has anyone found a different short, compelling greeting that works well? We’d love to hear what you’ve tried.


AI Chatbot Instruction Box Tips

The AI chatbot seems to ignore general directives when they’re placed in the headings of uploaded training documents.

However, it definitely does seem to follow new directives added to the Instructions box under the Train Your Assistant section.

For example, after appending the instruction below to the existing text in that box — without removing any of the original content — the chatbot began displaying all page titles, including the titles of products it recommends, as clickable hyperlinks:

When mentioning the name of any web page — such as a product page, contact page, or FAQ — always display it as a Markdown-style hyperlink using the page title as the anchor text. Example: [contact us](https://www.example.com/contact-us.html). Do not display plain URLs or unlinked page titles. All references to web pages must be formatted as clickable links.

Important: If entering a similar instruction, replace the sample title and URL with an actual page title and URL from your own website.

Looking forward to any feedback, tips, or suggestions!

4 Likes

Hi there, @Paul_D :waving_hand:

Thanks so much for starting this thread - it’s a fantastic idea! I’m sure a lot of users will find your insights really helpful.

As for the Instruction Box, you’re spot on. The “Instructions” section is definitely the best place to add those general guidelines (like formatting preferences, tone of voice, etc.). I love the example you gave about clickable links — such a great tip!

On the topics of chatbot greetings and where to install the bot, I’m also excited to hear what other Elfsighters have tried. I’m sure we’ll get some awesome ideas and strategies from the community.

Feel free to keep the conversation going, and let’s make this thread as insightful as possible :wink:

1 Like

Some websites give their chatbot a friendly name to help increase visitor engagement.

The name usually stands for something. For example, a chatbot might be called EVA, which might stand for Expert Virtual Assistant.

It’s a good idea to create your own unique, short, friendly name for your chatbot. AI tools like ChatGPT can help suggest names that fit your site, along with meaningful abbreviations.

You’d configure a greeting like:

Hi, I'm EVA.
How can I help?

If you set this up, it’s important to add a line like the following to the Instructions box:

You are an AI virtual assistant called EVA, because your full name is Expert Virtual Assistant.

If you don’t add this to your Instructions box, and someone asks, “Why are you called EVA?” the chatbot will reply, “My name is not EVA.”

4 Likes

Just a question? Whats the hierarchy in the chatbot
1 st Instructions
2d Q& A
3d Files
4 Text blocks?
Can you explain me the flow of an answer?

3 Likes

Armando,

I don’t know if there is a hierarchy. I believe these elements all work together to train the LLM. Instructions are important to give the bot guidance like, “Where available always include a URL.” That’s how I got the bot to include our product page URLs in product descriptions.

2 Likes

In our experience, scraping our web pages for chatbot training wasn’t very effective.

We started by scraping a single, comprehensive FAQ page from our site. While the content is SEO-optimized and works well for human visitors, it repeatedly confused the AI.

We had to heavily edit and shorten the training document created from the scraped page, removing any entries that overlapped with other training materials — even subtle or partial overlaps — before the chatbot gave satisfactory answers again.

Our takeaway: content that works for people doesn’t always work for the AI.

What works better for us is monitoring chatbot history emails. When we spot an unsatisfactory answer, we upload that chat to ChatGPT along with our training documents, explain the correct response, and ask ChatGPT to suggest specific edits to the training files.

Since the chatbot is newly installed, we’re currently making a few edits per day.

At this stage, we’re impressed by the quality of nearly all responses.

2 Likes

Hi there, @Armando_Espino :waving_hand:

Q&A, Files, and Text Blocks

All these blocks serve the same basic purpose — they let you add knowledge base content to your widget. We offer these different formats so you can pick whichever one suits your needs best, but they have the same prioritization level.

Instruction

The Instructions section is meant for general guidelines like how the bot should respond, its communication style, adding contact details, etc, not for storing actual knowledge base content.

Note: Just be careful not to overload Instructions, as that can make the bot’s responses less accurate.

If you notice the bot isn’t using (or is misusing) information from Q&A, Files, or Text Blocks, you can increase their priority by adding key phrases in the Instructions section.

You can find some useful examples in this guide :slightly_smiling_face: - Basic Tips For Your AI Chatbot Instruction

1 Like

Oh, interesting insight—thanks for sharing, @Paul_D! :+1:

I checked with the devs, and they confirmed it’s just a coincidence. All Q&As have the same priority, regardless of their order in the list :slightly_smiling_face:

2 Likes

Using Chatbot Logs to Improve Your Instructions Text

We’ve found that the Instructions field is the key to determine how the AI interprets and responds to user questions. It defines the chatbot’s overall behavior and can be used to establish general rules that guide how the AI uses your training documents.

Use your chatbot logs to guide targeted improvements to your Instructions.

Avoid over-editing or adding unnecessary detail — too much Instructions text can reduce response quality and slow performance. Instead, focus on concise, high-impact rules that address specific issues you observe in your chatbot logs.

When you find an unsatisfactory chatbot reply, one approach is to upload your current Instructions text, training documents, and the user question with the chatbot response to a tool like ChatGPT to get suggestions for refining the Instructions.


Examples – for an imaginary restaurant food delivery website
(Note: These examples are for illustration only. Don’t use them without adapting to your use case.)


1. Watch for Overly Broad or Overly Restrictive Responses
If the AI interprets questions too narrowly, the chatbot may respond with “I don’t know” too often. Or if it’s too loose, it might give irrelevant or misleading answers.

When you notice these patterns in your chat logs, adjust your Instructions to better guide the AI. You may need a few iterations to find the right balance.

Example Instructions Rule:

MATCH ALL QUESTIONS BY INTENT AND CLOSE MEANING

- Always check for a closely related FAQ before responding that the information is unavailable.
- Don’t rely on exact wording — match based on meaning and user intent.
- For a restaurant delivery site, match based on:

* Restaurant type, delivery area, price range, or delivery estimate
* How to view menus, place orders, or track deliveries
* Availability of specific restaurants or cuisine types
If a relevant FAQ exists, provide that answer.
* Only respond with "I don’t have that information" if no entry matches the user’s intent.
* Prioritize meaning over phrasing.


2. Watch for Confusing Words
Some training documents may include words with multiple meanings that confuse the AI. Use the Instructions to clarify the intended meaning based on how users typically ask questions.

Chat log issue:
Users asking about a “menu” get an answer about the website navigation menu instead of a food menu.

Example Instructions Rule:

“MENU” QUESTIONS
Assume "menu" refers to a restaurant food menu unless the user clearly means a website UI element.
Only refer to a UI menu if:
* The user asks how to navigate the website, or
* The matching FAQ or document describes a menu item in the UI.


3. Flag Important Special Cases
If your training materials include marginal or advanced options that don’t apply to most users, and the AI suggests them too frequently, consider adding a rule to limit when those items are shown.

Example Instructions Rule:

EXPENSIVE RESTAURANT RULE
Do not suggest "Jacob’s Grillhouse" or "Bella’s Bistro" unless the user asks for them by name.

This helps the AI avoid promoting less suitable choices while still keeping them in the training material.


I’d love to hear how others are using Instructions to improve accuracy and avoid misfires.

Please share your own examples or ideas.

4 Likes

Our website serves mainly business customers, but we occasionally see inappropriate or irrelevant chatbot questions (like “tell me a joke”).

We want to ensure the chatbot’s reply doesn’t invite more off-topic input, so we’re currently testing this additional Instructions text:

HANDLING NONSENSE OR OFF-TOPIC INPUT
If the user submits input that is clearly unrelated to our products or services (e.g., random words, nonsense, or irrelevant phrases), respond with a brief, non-confrontational message that avoids inviting further questions:

“I’m here to help with questions about our products and services. If you have a specific project or problem in mind, just let me know!”

Do not respond with links, apologies, or suggestions to contact support unless the input shows some connection to our products or services. Avoid saying “I don’t have that information” for input that has no discernible meaning or relevance.

Would be interested to hear what others have tried!

3 Likes