GoHighLevel AI Knowledge Base: Setup Guide (2026)
GoHighLevel AI Knowledge Base: Setup Guide (2026)
By Dr Priya Jaganathan, GoHighLevel Certified Admin · HL Growth Partner, Australia · Updated 1 September 2026 · 8 min read
A GoHighLevel AI knowledge base is the pile of source material your bot is allowed to answer from — and nothing else. It is not a personality, not a prompt, and not a magic layer that reads your website in real time. It is a set of stored, indexed snippets that Conversation AI and Voice AI search through before they compose a reply. If the answer is not in there in plain words, the bot will either guess badly or fall back to a holding message.
Most of the bots I get asked to fix in Australian sub-accounts are not broken bots. They are starved bots. Someone pointed the crawler at a homepage full of stock photography and three lines of copy, clicked Train, and expected it to quote pricing and open hours. This guide covers what the training sources actually do, how to structure material so answers come out clean, how to test before you go live, and where the credit costs land.
What a GoHighLevel AI knowledge base actually is
Inside a sub-account, the knowledge base lives under the AI Agents / Conversation AI area (Settings > AI Agents in most builds, or via Agent Studio if your agency has that enabled). You add training sources, HighLevel chunks and indexes them, and at run time the bot retrieves the closest matching chunks and writes an answer grounded in them.
Three things follow from that architecture, and they explain nearly every complaint I hear:
The bot retrieves, it does not reason about your business
Retrieval is similarity matching. If a customer asks "do you bulk bill?" and your document says "we are a private billing practice with no gap for concession card holders", there is no shared vocabulary for the matcher to latch onto. Add an FAQ pair using the customer's words, not yours.
Training is a snapshot in time
A URL training source crawls the page when you add it. Change your pricing page and the bot keeps quoting the old figures until you re-crawl. This is the single most common source of embarrassing answers — a bot quoting a $99 setup fee that went to $249 in March.
Intents and the knowledge base do different jobs
Intents (including the appointment booking intent) drive actions — collecting a name, offering calendar slots, tagging the contact. The knowledge base drives information. A bot with a beautiful knowledge base and no booking intent will explain your services eloquently and never book anyone.
Training sources for a HighLevel AI knowledge base
You have three real training source types, plus contact data that behaves like a fourth. Pick deliberately — each has a different refresh burden and a different way of failing.
| Training source | Best used for | Refresh effort | How it fails |
|---|---|---|---|
| FAQ pairs (question + answer) | Pricing, hours, location, parking, cancellation policy, "do you treat X", objections | Manual edit, 10 minutes a month | Written in staff jargon instead of customer phrasing, so retrieval misses |
| URL / web page crawl | Service pages, About, long-form blog explainers already written for humans | Re-crawl after every site edit — nothing is automatic | Pulls navigation, cookie banners and footer text as if it were content; goes stale silently |
| Uploaded documents (PDF, DOCX, TXT) | Price lists, service menus, onboarding packs, policy documents | Re-upload the whole file per version | Tables and multi-column PDFs chunk into nonsense; scanned PDFs contain no extractable text at all |
| Custom fields and custom values | Contact-specific and account-specific facts — business name, address, booking link, trading hours | Set once, edit centrally | Not knowledge base content — if referenced in prompts but left blank, the bot outputs an empty gap mid-sentence |
Custom values deserve a specific mention. Store your booking URL, phone number and business address as custom values and reference them in the bot's prompt rather than hard-coding them into forty FAQ pairs. When the number changes you edit one field. If you have not set these up cleanly, work through a proper GoHighLevel custom values setup before you train anything.
Structuring and chunking your source material
Chunking is where good knowledge bases are won. HighLevel splits long documents into segments and retrieves segments, not whole files. A 40-page policy PDF becomes dozens of fragments, and the fragment containing your refund terms may not contain the words "refund policy" anywhere in it — because the heading sat two pages earlier.
Write self-contained blocks
Every block should make sense read alone by someone with no context. Repeat the subject in the body: not "It costs $180 and runs for 50 minutes" but "An initial physiotherapy consultation at our Parramatta clinic costs $180 and runs for 50 minutes."
One topic per FAQ pair
Resist the mega-answer. "Hours, parking and payment methods" as a single pair will be retrieved for a parking question and dump irrelevant payment detail into the reply. Split it into three.
Use the customer's vocabulary in the question
Write the question the way a stressed person types it at 9pm on a phone: "how much is a first appointment", "do you do after hours", "are you open Saturdays". Add two or three phrasings as separate pairs where the wording genuinely differs.
Convert tables to prose before uploading
A pricing table in a PDF frequently loses its column alignment during extraction, and the bot confidently pairs the wrong price with the wrong service. Rewrite each row as a sentence. It is dull work and it is the difference between a bot you can trust and one you cannot.
Strip the marketing
Crawled pages carry hero headlines, testimonials and calls to action. Those pollute retrieval. Where possible, point the crawler at content-dense pages — a detailed FAQ page or service page — rather than a visual homepage. If you are also generating page copy inside the platform, the same discipline applies to GoHighLevel Content AI setup: specific, factual copy trains better than adjectives.
Testing before you switch the bot on
Use the bot trial / preview panel inside the AI agent settings before the widget goes anywhere near a live visitor. This runs conversations without touching real contacts or consuming your live channels.
Build a 25-question test script
Pull the last 100 real conversations from your Conversations tab and list the questions people genuinely asked. That is your test set. Run every one through the preview, and log three outcomes per question: correct, wrong, or fell back. Anything wrong gets a new FAQ pair written in the exact phrasing that failed.
Test the awkward ones deliberately
Ask it for a discount. Ask for medical or legal advice. Ask something about a competitor. Ask a question you have deliberately not trained. A knowledge base is only safe once you know what it does when it has nothing to say.
Start in Suggestive mode
Suggestive mode drafts a reply and waits for a human to approve or edit before sending. Auto-Pilot sends without review. Run two weeks in Suggestive on live traffic, correct the drafts, and feed every correction back in as an FAQ pair. Then switch to Auto-Pilot channel by channel — web chat first, SMS last. SMS mistakes are the ones that get screenshotted.
Check the channel plumbing
The widget under Sites > Chat Widget needs to be published on the live domain. SMS replies need LC Phone provisioned and A2P registration approved. Email replies need Mailgun connected and the sending domain verified. A perfectly trained bot that cannot send is a very common and very silly failure — and if messages are queuing rather than sending, work through standard GoHighLevel workflow troubleshooting steps before blaming the AI.
Guardrails, fallbacks and escalation to human
Every knowledge base needs a defined edge. Write an explicit instruction in the bot prompt: if the answer is not in the knowledge base, say so plainly, offer the booking link, and hand off. Never let it improvise.
Escalation to human should be triggered by more than confusion. Configure handoff on: a request to speak to a person, three consecutive fallback replies, any mention of a complaint, refund, cancellation or legal matter, and any clinical or financial advice request. In a Workflow, use the Conversation AI Bot action to run the bot for a defined window, then branch to a human notification if no booking is made or a handoff tag is applied.
For after-hours coverage the same knowledge base can back a voice agent, which is why teams building a GoHighLevel Voice AI after-hours receptionist should train once and reuse. Voice is less forgiving than chat, though: keep spoken answers under about thirty words, because a caller will not sit through a paragraph.
Per sub-account versus snapshot-level knowledge
Knowledge bases are sub-account scoped. If you run twelve clients, you have twelve knowledge bases. The useful split is between generic and specific.
Generic material — how booking works, what happens after a form is submitted, rescheduling and cancellation mechanics, privacy wording — belongs in your snapshot as a starter set. Specific material — prices, practitioner names, suburb, opening hours, parking — must be written per sub-account. Bear in mind that snapshot behaviour around AI configuration has changed repeatedly and not every AI asset transfers cleanly; verify what actually landed in the child account rather than assuming. Our GoHighLevel snapshots guide covers how updates propagate and where they do not.
The failure I see most in agencies: a snapshot loaded into a Brisbane client still carrying the demo clinic's Melbourne address in its FAQ pairs. Audit every new sub-account's knowledge base for leftover demo facts before launch.
Keeping the knowledge base fresh
Set a recurring monthly task. Fifteen minutes, four steps: re-crawl any URL source whose page has changed; re-upload any document with a new version; read the last month of bot conversations for fallbacks and wrong answers; add the fixes as FAQ pairs. Tie it to the calendar, not to goodwill.
Add an unscheduled review whenever prices change, a practitioner joins or leaves, hours change over Christmas or Easter, or you add a service. Australian public holiday hours are a reliable source of wrong answers every December.
Cost and credit implications
Conversation AI, Voice AI and the wider AI Employee suite bill on usage — per message for chat, per minute for voice — drawn from agency wallet credits, with rebilling available to sub-accounts. Training and re-training the knowledge base is not the expensive part; conversation volume is. Check the current published rates on the GoHighLevel pricing page and the setup detail in the official HighLevel help centre, because these figures are revised more often than most agencies notice.
The practical consequence: a thin knowledge base costs more, not less. Bots that cannot answer generate long back-and-forth threads before escalating, and you pay per exchange. Ninety minutes spent writing sixty solid FAQ pairs usually pays for itself inside a month of live traffic.
What a GoHighLevel AI knowledge base cannot do
It cannot read your calendar availability from a document — live slots come from the booking intent and the calendar integration, not from training material. It cannot look up an individual customer's order, invoice or appointment unless that data is exposed to it through fields or a workflow. It does not browse the live web at answer time. It cannot reliably interpret scanned or image-only PDFs. It will not give safe clinical, legal or tax advice, and you should explicitly instruct it not to try. And it does not learn on its own from conversations — every improvement is something a human adds.
Common mistakes to avoid
- Training on the homepage URL alone and assuming the crawler found the pricing, hours and service detail buried three clicks deeper.
- Uploading a scanned or image-based PDF, seeing no error, and never realising zero text was extracted from it.
- Writing FAQ answers in internal terminology — "initial biopsychosocial assessment" — that no customer would ever type.
- Going straight to Auto-Pilot on SMS on day one, with no Suggestive-mode period and no human reviewing drafts.
- Never re-crawling after a website update, leaving the bot quoting prices that changed six months ago.
- Cloning a snapshot into a new client and shipping the demo account's address, hours and practitioner names with it.
If you want your GoHighLevel AI knowledge base built and trained so the bot actually answers correctly, book a strategy call with the HL Growth Partner team.
Frequently asked questions
How many FAQ pairs does a GoHighLevel AI knowledge base need before it is useful?
For a single-location service business, 40 to 60 well-written pairs covering pricing, hours, location, services, booking, cancellation and the five most common objections will handle the large majority of enquiries. Quality beats volume — sixty specific, self-contained pairs outperform three hundred vague ones, because vague content makes retrieval ambiguous.
Does the URL training source keep my bot updated automatically when I edit my website?
No. A URL source captures the page content at the moment you train it. If you edit the page, you must re-crawl that source for the bot to see the change. Build a re-crawl step into your website update process, otherwise the bot will quote outdated pricing indefinitely with complete confidence.
Can one knowledge base serve both Conversation AI and Voice AI?
Within the same sub-account, yes — the trained material can back both text and voice agents. Adjust the prompt instructions per channel though: voice answers must be short and spoken-sounding, while chat can carry longer replies and links. Same facts, different delivery instructions.
Should I use Suggestive or Auto-Pilot mode when I first launch?
Suggestive, for at least two weeks on real traffic. It drafts replies for a human to approve, which surfaces every gap in your training material at zero risk. Once your approval rate is consistently high and corrections are rare, move to Auto-Pilot one channel at a time, starting with web chat.
Do I have to rebuild the knowledge base for every sub-account?
Knowledge bases are sub-account scoped, so each client needs their own. Keep generic process content in your snapshot as a starter set and write the client-specific facts fresh each time. Always verify what actually transferred after loading a snapshot, and delete any demo details before the bot goes live.
