GoHighLevel Conversation AI tutorial: train a chatbot that sounds human (2026)

This is where most bots go wrong. HighLevel splits a bot into three parts, and each one has a job.

The Knowledge Base is the facts. In my words from the real estate build: it's the information the bot refers to, to answer questions, to know things about the business, to know what it should be doing. Prices, hours, services, policies, listings.

The prompt is the behavior. Who it is, how it talks, what it must never do, and what to do when it doesn't know. HighLevel's own docs say FAQs don't belong in Additional Instructions, and that's right. If you paste your whole price list into the prompt, you get a long, messy prompt and answers that drift.

Bot Goals and actions are the doing: booking, collecting the email, triggering a workflow, handing off. The prompt can say how to talk about booking, but the booking itself is set in Bot Goals.

The web crawler is the fastest way to train a bot, and super underrated for demos. Before a sales meeting I'll scrape the prospect's website and show them a bot that already knows their business. It blows them away, because they've never seen it answer questions about their own pool or their own listings.

HighLevel's advice is not to train every page, and I agree for live bots. Navigation text, footers, legal boilerplate and a three year old promo all end up as chunks the bot can pull from. Pick the service pages, pricing, FAQ, about and contact pages, and leave the rest out.

After it crawls, open the content review icon next to the URL and read what it actually pulled. You can delete junk and add missing context there, and it doesn't change the live website. One catch: refreshing or re-crawling that URL can overwrite your edits, and that includes Auto Refresh. So for a page you've edited by hand, either skip Auto Refresh or copy your edits somewhere first.

For pages that change often, Auto Refresh re-crawls on a daily, weekly or monthly schedule (AI Agents > Knowledge Base > Web Crawler, the circular arrows icon). HighLevel says re-crawls don't cost extra credits, and if one fails the old content stays live.

FAQs are for the nuance questions. When I'm meeting with the owner I ask: what are the questions people ask you that I can't find on your website? Those become FAQs. In my hotel concierge build I added a deliberately weird one, what kind of chlorine do you use in the pool, just to see if the bot would find it.

If the business has no website, or the good info isn't on it, make a Google Doc with everything about the business (services, listings, common scenarios), share it so anyone with the link can view, and add it as a web crawler source with Exact URL. That's the hack from the video above. You can also import Docs, PDFs and Word files through the Google Drive source, and there's a daily sync in Labs that re-checks a Drive file every 24 hours.

For prices and service menus, use a Table. Upload a CSV (up to 500 columns, and you pick the 20 most relevant for retrieval). If a bot keeps getting a price wrong from a CSV, check which columns you selected and whether the table finished processing, then test that exact question in the Retrieval Tester.

When I built an Airbnb concierge for hosts, the knowledge base was very specific to the property: where the remote is, which streaming services the TV has, how to get into the pool, the host's rules. Get it from the owner with a simple form. The more specific it is, the more the bot sounds like someone who actually knows the place.

A bot sounds like a bot when it replies in half a second with three paragraphs and a bullet list to a one-line text. Fix the length, the speed and the personality and most of the robot goes away.

Length: turn on Response Settings in the bot and pick Concise for SMS and DMs. It aims for about 30 words. Balanced is about 80, Detailed about 200, which can work for a support chat on a website. These are targets, not hard caps. Don't also write "always reply in 20 words" in the prompt, HighLevel says competing length rules make results less predictable.

Speed: set a minimum and maximum response delay in the bot's advanced settings. HighLevel suggests 2 to 5 seconds to mimic reading time. In my early SMS builds I added a wait step for the same reason, so it would seem human.

Personality: write it like you're briefing a new hire at the front desk. Then add the rules in Additional Information. HighLevel also has Brand Voice, which you can create from text or a URL, if the business already has a tone guide.

For a demo you don't have to go crazy with the prompt. Tell ChatGPT the business, the objective, the tonality and how it should speak, and paste the result in. Save the website as a PDF and give it that as context so the prompt knows the business.

The line I add to my prompt framework: if there's anything unclear, ask me questions first. It comes back with questions about boundaries and restrictions (should it quote prices? what is it not allowed to do?), and the second version is a lot better.

When you close the deal, make it far more detailed for the real client. The more you put into the prompt the better the output. Just watch the word count shown in the prompt area, and read the whole thing once for rules that fight each other. HighLevel's example of a conflict is "always keep replies under one sentence" next to "explain every service in detail".

You can also insert custom values (contact name, business details) with Add Custom Value, so one prompt works across a snapshot of client accounts.

Two ways. The normal one: Sites > Chat Widget > New, pick Live Chat (that's the real-time Web Chat channel, which is the one Conversation AI answers), style it, then copy the code onto the client's site or pick the widget inside a HighLevel funnel or website. Then in the bot's Deploy tab, turn on Chat Widget and select that widget.

The fast one, if the site is built in HighLevel: open the funnel or website in the builder, Add Elements > AI And Chat Widget, create a Live Chat widget, then on the AI agents tab click Setup Conversation AI and Scan & auto-configure. HighLevel reads the funnel pages and writes a Knowledge Base, Personality, Goals, Instructions and an opening message for you. It's a good first draft. Click Manage agent and read every field before you publish, because it only knows what's on those pages.

One thing to know: the widget's welcome message is the same on every page, it can't change per page. Put page-specific stuff in the knowledge base instead.

Test in two layers. First the knowledge base on its own: AI Agents > Knowledge Base, open it, and use the Retrieval Tester. Type a real customer question and look at which chunks come back. If the right chunk isn't there, no prompt will save you. Fix the source first.

Then the whole bot in Test your agent: a common question, the same question worded differently, one of your FAQs, and something the business doesn't do. Hit the reset icon between scenarios. Use Preview as to see how it looks on each channel, since Live Chat shows buttons and Instagram and Facebook show quick replies. Nothing in the test panel gets sent to a real contact.

Then Prompt Optimizer. If you have it on in Labs, an Optimize Prompt button shows up in Test your agent once the bot has replied. It builds scenarios from that chat and your current prompt (even unsaved edits), runs them against a copy of the agent, and suggests a better prompt. Applying it only changes the builder, so read it and click Save yourself. Actions can still run during these tests, so point the bot at a test calendar and test workflows first. On my voice agents I keep running these simulations until it passes almost everything (see Build an AI voice agent), and Conversation AI works the same way.

The last test is you, on your own phone, on every channel the bot is on. In my hotel build the bot got the gym and pool right, said it wasn't sure about the chlorine and offered to have the team follow up, which is exactly what you want when it doesn't know. It also didn't pull my FAQ that time, which is why you test and double check.

Don't just rewrite the prompt and hope. Find out where the answer came from.

In Conversations, click Response Info next to the AI message. It shows the knowledge chunks (up to three), FAQs and other sources behind that reply, and you can edit some knowledge base entries straight from there. Agent Logs go deeper: each turn broken into steps, so you can see whether it searched the knowledge base, which action it called and what came back.

Then fix the cause. Wrong price: there's probably an old page or a second version of the price list in the knowledge base, so delete the outdated one. Made something up: the answer isn't in the sources, so add an FAQ, and make sure the prompt says not to guess. Right info, wrong tone: that's the prompt or Response Style. Ignored an instruction: look for a rule that conflicts with it, or something that belongs in Bot Goals.

Retest the exact question that failed, in the Retrieval Tester and in Test your agent. For a new client, run the bot in Suggestive mode for the first stretch so a person approves each reply, and you'll catch these before a customer does.

Every bot needs a way out. Human Handover (a Labs action in Bot Goals) passes the chat to a person when someone asks for a human, when the bot doesn't have the information, or when it fails to resolve something after two tries. The setup is in my appointment setter guide. What I'd add here: write the escalation cases in the prompt too, in plain words, the same way I tell a concierge bot to escalate emergencies to the property manager.

To keep the bot away from existing customers, use tags. With Channel Management on (Settings > Labs, then each bot's Deploy tab), every channel can have Has tags and Doesn't have tags conditions. Tag customers as customer and set the sales bot to Doesn't have tags: customer. You can even route two bots off the same widget, one for leads and one for customers.

For one contact, use the bot status icon in the Conversations message box: Active, Inactive, or Inactive with a reactivate timer. The bot also goes inactive on its own when a contact replies STOP, when it hits its max messages, or after a manual message if that sleep setting is on.

If a contact is getting two replies to one message (people report this on Instagram), check that two bots can't both match that contact, and that a workflow isn't also sending a reply or calling a bot at the same time as the channel deployment. A bot assigned to a contact directly, by a workflow action or a bot transfer, takes priority over channel rules.

The Add Contact Info action can now fill radio, checkbox, single and multi-select dropdown and phone custom fields, plus Contact Source and Contact Type. The customer doesn't need to use your exact wording: "I'm interested in consulting and training" gets saved as those two options. It works on prompt-based bots in Auto-Pilot. That's how the qualifying answers end up on the contact, where your workflows and pipeline can use them (see GoHighLevel workflows).

HighLevel's prompts article still lists GPT-4.1 as the recommended default. Since Oct 27, 2026 you can also turn on Claude and Gemini support in Settings > Labs and pick Claude Opus 5.5, Sonnet 5.5, Haiku 5.5 or Haiku 4.5, or Gemini 3.8 Flash, 3.7 Flash or 3.5 Flash Lite from the same Model menu. Each one shows a relative cost from $ to $$$$, and you can set a secondary model from another provider in case the main one goes down.

For a website FAQ bot or SMS replies, a fast cheap model is usually plenty. Use a top model for longer sales conversations where the wording really counts. Switching models keeps your prompt, knowledge base and actions, but each model writes differently, so run your tests again before you save. If a bot was still on GPT 5 or GPT 5 Mini, move it, those retired Oct 30.

Building the Knowledge Base is free (fair use), and HighLevel says web page re-crawls are included. The replies are what you pay for: pay per use by tokens, or an AI Employee plan on the sub-account. Growth is $50 a month with 1,000 Conversation AI responses and 500 Prompt Optimizer messages, and Unlimited is $97 a month with fair use. Without a plan, Prompt Optimizer runs bill at the normal Conversation AI per-use rates, so a long optimization run on a top model isn't free. Prices as of Oct 31, 2026. The full cost breakdown, with HighLevel's token math, is in the appointment setter guide.

AI sells itself, but only if they can see it. Crawl the prospect's website, build the bot before the meeting, and let the owner ask it questions about their own business. Then ask them for the nuance questions it got wrong, and that list is your onboarding.

This is a real business model and I run my agency on it. You could set this up for a med spa, a hotel or a property manager and charge a setup fee plus a monthly fee you set to keep the knowledge base current and the bot tuned. How I find and pitch those clients is in my guide to selling AI automation to local businesses.