Playbook: Create a chatbot from a website

This playbook is written for AI agents and automation scripts. The goal is:

Given a website URL, create a useful SiteGPT chatbot end-to-end with knowledge, branding, icons, persona, instructions, starters, followups, and initial verification.

Use this page to choose the right setup path. For the full step-by-step flow, use the path-specific playbook:

New customer onboarding

Build a temporary preview chatbot before signup, then share the onboarding URL for preview and claim.

Existing account setup

Create or update a chatbot inside an authenticated SiteGPT account.

Choose the setup path

Use the right path before running authentication checks:

Situation	Path
The user does not have a SiteGPT account yet or wants to try SiteGPT first	Agent-first onboarding chatbot
The user already has a SiteGPT account, token, or saved profile	Existing account chatbot setup

For agent-first onboarding, start with sitegpt onboarding start. It creates a temporary chatbot and returns a temporary token. Use that token for the rest of the setup, then share the onboarding URL after the chatbot is useful. PROFILE_NOT_CONFIGURED is expected in this flow and should not stop setup. For existing accounts, log in first and create the chatbot directly in the account.

Configuration differences

Both flows should produce a well-configured chatbot, but the safety posture is different:

Area	Agent-first onboarding	Existing account
Knowledge	Add the best initial source and use `onboarding status` checklist before sharing.	List existing docs/jobs before updating; avoid deletes or bulk resync unless asked.
Brand settings	Make the preview feel branded immediately before sharing the onboarding URL.	Read current appearance first; preserve existing choices unless creating a new bot or user asked for refresh.
Persona	Create one strong default and activate it.	List existing personas before replacing active behavior.
Instructions	Create one active grounded instruction set.	List current instructions and avoid clobbering production behavior without approval.
Starters/followups	Add broad visitor prompts so the preview is easy to test.	Add/refine prompts; do not remove existing prompts without approval.
Lead/support	Enable only when purpose and visible contact details justify it.	Respect existing notification/routing settings.
Handoff	Share onboarding URL and claim path.	Share dashboard/install links and summarize changes.

Required access for existing accounts

Use a profile with these scopes:

sitegpt login \
  --profile setup-agent \
  --scope account:read \
  --scope chatbots:read \
  --scope chatbots:write \
  --scope knowledge:read \
  --scope knowledge:write \
  --scope settings:read \
  --scope settings:write \
  --scope personas:read \
  --scope personas:write \
  --scope instructions:read \
  --scope instructions:write \
  --scope starters:read \
  --scope starters:write \
  --scope followups:read \
  --scope followups:write

If you also need to delete mistakes during setup, include the matching delete scopes. For agent-first onboarding, no login is needed before the first command. The temporary token returned by sitegpt onboarding start is scoped to the new chatbot.

Inputs

Ask for or infer the chatbot purpose before creating the chatbot. If the user did not say, ask one concise question:

What should this SiteGPT chatbot optimize for: customer support, marketing/site
guide, lead generation, docs/help, onboarding, or a mix?

If the user is unavailable and the task should proceed, infer the purpose from the prompt and website, then state that assumption in the final report. Capture:

Input	Example
Website URL	`https://example.com`
Chatbot purpose	Customer support, lead generation, documentation assistant
Target audience	Customers, prospects, internal team
Support email	`support@example.com`
Brand preference	Friendly, formal, concise, technical

1. Inspect the website

Use browser/fetch tools for prose understanding, but use raw HTML for structured signals such as colors, icons, manifest, and sitemap links.

curl -sL https://example.com -o /tmp/sitegpt-page.html

Extract candidate colors:

grep -oE '#[0-9a-fA-F]{6}' /tmp/sitegpt-page.html | sort | uniq -c | sort -rn | head -30
grep -oE 'name="theme-color"[^>]*' /tmp/sitegpt-page.html

Extract candidate icons and images:

grep -oE '<link[^>]*rel="[^"]*icon[^"]*"[^>]*' /tmp/sitegpt-page.html
grep -oE '<link[^>]*apple-touch[^>]*' /tmp/sitegpt-page.html
grep -oE '<link[^>]*manifest[^>]*' /tmp/sitegpt-page.html
grep -oE 'property="og:image"[^>]*content="[^"]*"' /tmp/sitegpt-page.html

Look for sitemap hints:

curl -sL https://example.com/robots.txt
curl -I https://example.com/sitemap.xml

Capture:

Brand/product name
Value proposition
Support topics
Docs/pricing/contact URLs
Best brand color, preferably from raw HTML, CSS variables, manifest, or a prominent CTA
Best icon/logo file, preferably apple-touch-icon, a logo PNG, or a clean square icon
Sitemap URL if available

Do not choose a generic color like #2563EB unless the site actually uses it. If color or icon falls back to a default, treat setup as incomplete and inspect raw HTML/assets again.

2. Start onboarding or create the chatbot

Agent-first onboarding

sitegpt onboarding start https://example.com --agent-name "<agent-name>" --json

Save the returned values:

WORKSPACE_ID=<data.workspace.id>
CHATBOT_ID=<data.workspace.chatbotId>
export SITEGPT_API_TOKEN=<data.apiToken>

The export is optional, but it keeps the rest of this playbook readable. If you do not export it, prefix each setup command with SITEGPT_API_TOKEN=<temporary-token>. Check the onboarding workspace:

sitegpt onboarding status $WORKSPACE_ID --json

Existing account

sitegpt chatbots create "Example Support" \
  --description "Answers customer questions about Example products, pricing, documentation, and support." \
  --json

Save the returned chatbot ID:

CHATBOT_ID=<returned-chatbot-id>

Skip this create command for onboarding because onboarding start already created the temporary chatbot.

3. Add knowledge

Prefer a sitemap when available because it usually gives cleaner coverage than a blind crawl.

Sitemap preferred
Website crawl fallback
Selected links

sitegpt knowledge sitemap add \
  --chatbot $CHATBOT_ID \
  https://example.com/sitemap.xml \
  --max-links 500 \
  --only-main-content true \
  --sync WEEKLY \
  --scan WEEKLY \
  --json

sitegpt knowledge website add \
  --chatbot $CHATBOT_ID \
  https://example.com \
  --depth 3 \
  --max-links 200 \
  --only-main-content true \
  --json

sitegpt knowledge links add \
  --chatbot $CHATBOT_ID \
  https://example.com/pricing \
  https://example.com/docs \
  https://example.com/contact \
  --only-main-content true \
  --json

If the website has noisy navigation, cookie banners, or repeated layout blocks, add selectors:

sitegpt knowledge website add \
  --chatbot $CHATBOT_ID \
  https://example.com \
  --include-selector main \
  --exclude-selector nav \
  --exclude-selector footer \
  --only-main-content true

4. Download and upload brand assets

Download the chosen icon/logo to a local file:

curl -sL https://example.com/logo-icon.png -o /tmp/sitegpt-logo.png
file /tmp/sitegpt-logo.png

Upload it as the bot icon and chat bubble icon when it is suitable:

sitegpt icons upload --chatbot $CHATBOT_ID bot /tmp/sitegpt-logo.png
sitegpt icons upload --chatbot $CHATBOT_ID chat-bubble /tmp/sitegpt-logo.png

If the source icon is tiny or visually poor, use the best clean logo mark available rather than a blurry favicon.

5. Configure appearance

Use the extracted brand color and readable contrast.

sitegpt settings appearance update \
  --chatbot $CHATBOT_ID \
  --title "Example Support" \
  --welcome "Hi! I can help with Example products, pricing, documentation, and support." \
  --placeholder "Ask about Example..." \
  --brand-color "#F43F5E" \
  --brand-text-color "#FFFFFF" \
  --link-color "#F43F5E" \
  --icon-shape CIRCLE \
  --icon-position RIGHT

6. Create persona

Write persona text to a temporary file:

cat > /tmp/sitegpt-persona.md <<'EOF'
You are Example's helpful customer support assistant.
Use a clear, friendly, and practical tone.
Prioritize accurate answers from SiteGPT knowledge.
When the visitor asks about pricing, support, integrations, setup, or troubleshooting, answer directly and offer the most relevant next step.
EOF

Create and activate it:

sitegpt personas add --chatbot $CHATBOT_ID --title "Example support specialist" --file /tmp/sitegpt-persona.md --json
sitegpt personas use --chatbot $CHATBOT_ID <persona-id>

7. Create instructions

cat > /tmp/sitegpt-instructions.md <<'EOF'
Answer using the chatbot knowledge first.
If the answer is not available in the knowledge base, say that you do not have enough information and suggest contacting support.
Keep answers concise unless the visitor asks for detail.
Do not invent prices, policies, or guarantees.
When useful, suggest a follow-up question the visitor can ask next.
EOF

sitegpt instructions add --chatbot $CHATBOT_ID --title "Grounded support instructions" --file /tmp/sitegpt-instructions.md --temperature 0.3 --json
sitegpt instructions use --chatbot $CHATBOT_ID <instruction-id>

8. Add starters and followups

Create conversation starters from the website’s main user intents:

sitegpt starters add --chatbot $CHATBOT_ID --title "What can you help with?" --message "What can you help me with?"
sitegpt starters add --chatbot $CHATBOT_ID --title "Pricing" --message "Tell me about pricing."
sitegpt starters add --chatbot $CHATBOT_ID --title "Getting started" --message "How do I get started?"

Create followups:

sitegpt followups add --chatbot $CHATBOT_ID --title "Talk to support" --escalation
sitegpt followups add --chatbot $CHATBOT_ID --title "View docs" --link https://example.com/docs --type LINK
sitegpt followups add --chatbot $CHATBOT_ID --title "Compare plans" --message "Can you compare the plans?"

9. Configure lead and support settings when relevant

For lead-generation sites:

sitegpt settings lead-form update \
  --chatbot $CHATBOT_ID \
  --enabled true \
  --collect-name true \
  --collect-phone true \
  --trigger intent \
  --notification-email sales@example.com

For support-heavy sites:

sitegpt settings human-support update \
  --chatbot $CHATBOT_ID \
  --enabled true \
  --request-prompt "Would you like me to connect you with support?" \
  --notification-email support@example.com

10. Verify the chatbot

sitegpt dashboard --chatbot $CHATBOT_ID --json
sitegpt knowledge documents list --chatbot $CHATBOT_ID --json
sitegpt messages send --chatbot $CHATBOT_ID "What can you help me with?" --json
sitegpt messages send --chatbot $CHATBOT_ID "How do I contact support?" --json

Check for failed documents:

sitegpt knowledge documents list --chatbot $CHATBOT_ID --status FAILED --json

If failures exist:

sitegpt knowledge documents resync --chatbot $CHATBOT_ID --state failed

For onboarding, also inspect the setup checklist:

sitegpt onboarding status $WORKSPACE_ID --json

Fix PENDING, WARNING, or UNKNOWN checklist items when possible before sharing the onboarding URL. For existing-account setup, give the user the dashboard link:

sitegpt dashboard --chatbot $CHATBOT_ID

For agent-first onboarding, give the user the onboarding URL returned by onboarding start after the chatbot is useful. If the user wants to claim it, ask for email, plan, and interval:

sitegpt onboarding claim $WORKSPACE_ID \
  --email user@example.com \
  --plan GROWTH \
  --interval MONTH \
  --json

Plans are STARTER, GROWTH, and SCALE. Intervals are MONTH and YEAR.

Quality bar

Before calling the setup complete, confirm:

Chatbot title and description match the website.
Knowledge was added from sitemap, website crawl, links, files, or text.
Brand color is extracted from the real site, not guessed.
Bot/chat bubble icon is uploaded when a suitable brand asset exists.
Persona and instructions are active.
Starters and followups match likely visitor questions.
Lead/human support settings match the user’s goal.
At least one test message returns a useful answer.
dashboard and documents list show no obvious setup failures.

Parallelization guidance

Agents can parallelize independent discovery tasks:

Fetch raw HTML, robots.txt, and sitemap.xml.
Inspect prose/content in a browser while raw HTML is searched for assets.
Download candidate icons while creating persona/instruction drafts.
After chatbot creation, add knowledge, upload icons, and prepare settings in parallel when the CLI environment supports safe concurrent commands.

Do not parallelize commands that depend on returned IDs, such as personas use before personas add returns a persona ID.

New customer onboarding

Existing account setup

​Choose the setup path

​Configuration differences

​Required access for existing accounts

​Inputs

​1. Inspect the website

​2. Start onboarding or create the chatbot

​Agent-first onboarding

​Existing account

​3. Add knowledge

​4. Download and upload brand assets

​5. Configure appearance

​6. Create persona

​7. Create instructions

​8. Add starters and followups

​9. Configure lead and support settings when relevant

​10. Verify the chatbot

​11. Share or claim

​Quality bar

​Parallelization guidance

Choose the setup path

Configuration differences

Required access for existing accounts

Inputs

1. Inspect the website

2. Start onboarding or create the chatbot

Agent-first onboarding

Existing account

3. Add knowledge

4. Download and upload brand assets

5. Configure appearance

6. Create persona

7. Create instructions

8. Add starters and followups

9. Configure lead and support settings when relevant

10. Verify the chatbot

11. Share or claim

Quality bar

Parallelization guidance