Find the screen that owns the problem
Begin with what the visitor experienced.
“The greeting is wrong,” “the policy is old,” and “the contact request fired unexpectedly” sound like chat problems, but they belong to three different screens. Use this tour to reach the owner without disturbing working configuration.
Begin with the symptom, not the menu
Section titled “Begin with the symptom, not the menu”| What you observe | Open this area first | Reason |
|---|---|---|
| Wrong public name, greeting, tone, or scope | Assistants | Display Name, Welcome Message, model, and Instructions live in the active profile. |
| Answer misses or uses old website facts | Knowledge Base, then Analytics | Check authority/current state first, then see whether retrieval found it. |
| Email, webhook, or live data call is wrong | Tools | Definitions, parameters, assignment, result messages, and Enabled kill switch live there. |
| Conversation moves to the wrong specialist | Agent Flows, then Analytics trace | Published workflow routing, handoffs, and specialist scopes own Advanced behavior. |
| WhatsApp fails but website chat works | Channels | Meta credentials, webhook state, and channel configuration are separate from the web widget. |
| Widget looks wrong or is misplaced | Design | Colors, layout, position, viewport controls, and design preview live there. |
| Landing title, default language, avatar, or access is wrong | Settings | General chat content and visibility are not Assistant profile fields. |
| Need conversation, cost, performance, or gap evidence | Analytics | Charts, Chat History, OpenAI usage/costs, and Knowledge Gaps are operational views. |
| Visitor is blocked, abusive traffic passes, or a message is flagged | Security | Rate, length, detection, blocked IP/word, flags, and whitelist controls live there. |
| Models do not load or OpenAI credentials need changes | OpenAI API | Regular runtime key and optional Admin usage credentials are managed there. |
The ten tabs at a glance
Section titled “The ten tabs at a glance”Knowledge Base curates published pages/posts, metadata, documents, the vector store, and synchronization state. It supplies retrievable facts.
Assistants creates local profiles and selects one active profile. It owns internal/public identity, welcome message, model, and behavior instructions.
Tools creates callable data and action capabilities. Disabled is the immediate per-Tool runtime kill switch.
Agent Flows designs Advanced specialist routing. The tab remains visible when Advanced Agents is off, but its editing interface is locked and explains how to enable it. Saved workflows remain stored.
Channels connects external conversation delivery. Currently this includes WhatsApp through Meta.
Design controls the widget’s visual system and preview: colors, size, positioning, responsive behavior, and appearance choices.
Analytics helps you observe what happened: conversation logs, model/token/duration data, knowledge diagnostics, gap heuristics, and Advanced route/handoff traces.
Security governs safeguards before AI processing: rate and length limits, blocked IPs/words, automated detection, flagged conversations, and whitelisting.
OpenAI API stores the regular key, optional organization Admin key and Project ID, Advanced Agents runtime toggle, and a read-only active Assistant reference.
Settings owns general widget/chat content, language and translation choices, avatars, access/visibility, and other plugin-wide behavior.
The separate Setup Wizard submenu is onboarding, not an eleventh configuration store. The hidden Assistant edit screen appears when you select Edit from a card.
Three labels that are easy to confuse
Section titled “Three labels that are easy to confuse”Assistant Welcome Message versus landing content
Section titled “Assistant Welcome Message versus landing content”The Welcome Message is the Assistant’s first conversation message. Landing title/description and other pre-chat copy live in Settings → Chat Content. Editing one does not replace the other.
Selected page versus Current knowledge
Section titled “Selected page versus Current knowledge”A selected page is local intent. Sync must publish the processed version to the vector store. Saving metadata can make a page Pending. Always read the state before testing.
Route decision versus handoff
Section titled “Route decision versus handoff”Advanced entry routing can choose the best specialist before the Base Assistant answers. That is logged as a route decision, not necessarily a runtime handoff. A handoff occurs when an active agent transfers to another agent under a handoff/switch rule.
A tour for three common jobs
Section titled “A tour for three common jobs”- Improve a poor answer. Start in Chat History, identify active profile/runtime and retrieval evidence, then move to Assistants or Knowledge Base based on what failed.
- Change who can use chat. Open Settings → Access. Test both an allowed and denied visitor; do not use Assistant instructions as access rules.
- Launch a new action. Create it disabled in Tools, test its receiver, test Basic behavior, connect it to an Advanced specialist only if needed, then monitor Analytics.
- Investigate abuse. Preserve the log, use Security to inspect the exact rule/score/action, and avoid weakening Assistant instructions or knowledge.
- Change branding. Separate identity wording (Assistant/Settings) from visual styling (Design), then inspect desktop and mobile.
Admin navigation and permission context
Section titled “Admin navigation and permission context”The WordPress sidebar uses AI Website Chat as the main menu. Setup Wizard is available as a submenu, and some workflows open hidden internal pages such as Edit Assistant. Normal plugin management requires an administrator with the appropriate capability, commonly manage_options; plugin installation requires the WordPress capability to install plugins.
Do not share a broad administrator account simply because another person needs to edit wording. Use organizational WordPress role/access practice, protect OpenAI and webhook secrets, and keep production changes attributable.
Which screen changes the AI reply directly?
Section titled “Which screen changes the AI reply directly?”| Area | Directly shapes generated content? | Its real influence |
|---|---|---|
| Assistants / specialist Agents | Yes | Instructions and model shape response behavior and scope. |
| Knowledge Base | Yes, when retrieved | Supplies approved website-specific context. |
| Tools | Yes, when called | Returns data or action results the response can use. |
| Design / visibility | No | Changes presentation and who can access chat, not the factual model context. |
| Security | Before generation | May allow, flag, warn, or block a message before it reaches the model. |
| Analytics | No | Observes and explains the result; it does not repair it. |
This distinction helps when an administrator sees a low-confidence answer and changes colors, or sees a hidden widget and rewrites Instructions. Supporting controls matter greatly, but they operate at a different point in the request journey.
When a tab is not the cause
Section titled “When a tab is not the cause”A valid setting can still fail because hosting blocks outbound HTTPS, PHP cannot run a dependency, mail delivery is unconfigured, Meta rejects a webhook, a reverse proxy caches assets, or a remote API enforces its own authorization. Once the relevant tab shows correct saved configuration, follow the transaction across its external boundary instead of toggling unrelated options.
Likewise, a reasonable AI answer can be wrong because the organization has no single approved source. The plugin cannot reconcile policy ownership through settings. Assign a content owner, correct the authority, synchronize, and retest.
Build a repeatable admin handoff
Section titled “Build a repeatable admin handoff”When training another administrator, give them five exercises instead of a tour of every field: change a Welcome Message and verify it, synchronize one harmless page edit, disable and safely test a synthetic data Tool, find that conversation in Analytics, and restrict access to a test user. Each exercise joins configuration with evidence.
Then show the emergency controls: restrict visibility, disable an individual Tool, turn Advanced Agents off to return to Basic runtime, and preserve a suspicious conversation log before making changes. Knowing how to contain a problem is more valuable than memorizing tab order.
The navigation bar you should recognize
Section titled “The navigation bar you should recognize”- Capture
- Show the AI Website Chat header and complete horizontal tab navigation at a width where no label is clipped.
- Show
- All ten main tab labels and active-tab indicator
- Viewport
- Desktop, 1440 × 900
- Annotate
- Use numbered callouts only for controls referenced in the procedure.
- Redact
- OpenAI keys, tokens, secrets, personal information, private URLs, IP addresses, and conversation text
Open the owner guide
Section titled “Open the owner guide”- Behavior and identity: Assistants
- Facts and sources: Knowledge Base
- Capabilities: Agent Tools
- Specialist routing: Agentic Flows
- Evidence-led system checks: Diagnostics