Edit pages like a knowledge librarian
A visitor rarely uses the same wording as your navigation menu. Page metadata helps the right source meet the visitor’s language—but the source itself remains the authority. Your goal is not to fill every box. Your goal is to leave a compact, truthful trail toward a useful answer.
The source is the truth; metadata is the signpost
Section titled “The source is the truth; metadata is the signpost”Suppose a page is titled “Right of Withdrawal,” while visitors ask, “How long do I have to return something?” The policy should still live in the page. Metadata can add “returns window,” a short summary, and a precise FAQ so retrieval recognizes the connection.
Metadata is a bad place to repair an unapproved or contradictory policy. It is also a bad place for tone instructions such as “always apologize first.” Put factual content in the source; put behavior in the Assistant or specialist Agent instructions.
Selecting pages without swallowing the whole website
Section titled “Selecting pages without swallowing the whole website”The Pages tab offers published WordPress pages and posts. Select durable, visitor-facing sources: services, pricing explanations, policies, location details, support material, and real editorial FAQs. Be suspicious of archives, tag pages, expired campaigns, duplicate language variants, thin landing pages, and anything whose main content is rendered by an external system rather than stored in the post.
Click a page to open its metadata editor. The Title, URL, and Content Preview are automatic context. Read the preview before writing metadata: it shows the processed text the knowledge pipeline can work with more accurately than the visual page builder does.
Saving Metadata also selects the page if it was not already selected. If the saved data changes the page’s current hash, its status becomes Pending and requires synchronization.
The complete metadata desk
Section titled “The complete metadata desk”| Field | What it contributes | A good entry | What to avoid |
|---|---|---|---|
| Title (automatic) | Names the source and travels with its knowledge document. | A specific page title such as “Zagreb Clinic Cancellation Policy.” | Changing metadata to compensate for a vague source title when you can improve the WordPress title itself. |
| URL (automatic) | Preserves the public source address as retrieval context. | A live canonical page visitors are allowed to open. | Selecting redirects, staging URLs, or private pages and assuming the URL grants access control. |
| Content Preview (automatic) | Shows a readable sample of processed page text. | Enough meaningful body text to answer the intended questions. | Ignoring missing text because the visual page looks complete; text in images may not become usable content. |
| Category | Adds one broad grouping signal. | “Appointments,” “Shipping,” or “Commercial solar installation.” | Internal acronyms, several categories squeezed into one value, or a category more specific than the page. |
| Summary | Explains the page’s scope in a compact editorial statement. | “Eligibility, deadlines, fees, and rescheduling rules for in-person appointments.” | Marketing claims, new facts, long introductions, or instructions to the AI. |
| Topics / Keywords | Connects visitor vocabulary, service names, synonyms, and concepts to the source. | “reschedule visit,” “missed appointment,” “cancellation fee.” | Keyword stuffing, unrelated popular terms, spelling every singular/plural variation, or hidden promises. |
| FAQs | Pairs a realistic question with an answer grounded in this source. | “Can I move my appointment?” followed by the approved deadline and method. | Invented edge cases, legal interpretation, duplicated questions, or answers that conflict with the page. |
| Key Facts | Surfaces short facts whose qualifiers matter. | “Cancellations within 24 hours may incur the published late fee.” | Fragments such as “24 hours” without explaining what the number means. |
| Related Pages | Links this source with another relevant internal page. | Policy → appointment booking; service overview → eligibility page. | Creating a web of every page to every other page, or linking merely because pages share a menu section. |
| Exclude Content | Removes each entered literal text fragment from processed content before upload. | An exact, stable boilerplate sentence that appears in the preview and should not become knowledge. | Privacy rules, CSS selectors, vague concepts, secrets, or “ignore the footer” as if the field understood page structure. |
Exclude Content is a text eraser, not a security rule
Section titled “Exclude Content is a text eraser, not a security rule”This field deserves special care. Each nonempty line is applied as a case-insensitive literal replacement to the processed page content. If the exact text is found, that text is removed. The field does not understand “header,” “sidebar,” “old prices,” or “private information” as concepts. It does not select a DOM element, enforce visitor permission, or redact similar wording.
For example, entering Book a free consultation removes that literal phrase wherever it appears, regardless of case. Entering remove the navigation will only remove those exact words if they happen to exist in the content; it will not discover a navigation block.
Use it for stable boilerplate you can see in Content Preview. After saving, inspect the JSON view or processed preview, synchronize, and test. If information must never leave the site, remove it from the source or do not select the source. Exclude Content is not an acceptable privacy boundary.
Build one excellent metadata record
Section titled “Build one excellent metadata record”- Name the page’s job. In one sentence, write what questions this source is authorized to answer.
- Read the preview. Confirm the decisive facts actually appear and that stale or irrelevant text is not dominating it.
- Add the visitor’s vocabulary. Enter a small set of Topics that people use but the source does not.
- Clarify scope. Write a Summary that says what is included and, when useful, what is not.
- Capture high-value questions. Add FAQs only for common questions with answers you can point to in the source.
- Protect qualifiers. Use Key Facts for limits, units, dates, audiences, locations, and conditions that must travel together.
- Relate sparingly. Add only pages that genuinely complete the visitor’s next step.
- Save, sync, and ask. Metadata sitting in the form is not live knowledge. Save Metadata, run Sync, and test in a new conversation.
A worked example: cancellation policy
Section titled “A worked example: cancellation policy”Imagine the source states that appointments can be rescheduled without charge up to 24 hours before the start time. A clean record might use:
- Category: Appointments
- Summary: Rules, deadline, and contact method for cancelling or moving an appointment.
- Topics: reschedule visit; change booking; cancellation fee; missed appointment
- FAQ: “Can I change my appointment?” → “Yes. Use the contact method on this page at least 24 hours before the scheduled start to avoid the late-cancellation charge.”
- Key Fact: The 24-hour deadline is measured from the scheduled appointment start.
- Related Page: Book an Appointment
The metadata does not say the assistant may waive a fee, cancel on the visitor’s behalf, or access a booking. Those are capabilities, not knowledge. They would require a deliberately designed Agent Tool.
Quality review: five questions before Save
Section titled “Quality review: five questions before Save”| Question | Pass condition | If it fails |
|---|---|---|
| Can every claim be traced to the approved source? | Yes; metadata summarizes or rephrases without adding policy. | Correct the source or remove the unsupported statement. |
| Would a visitor naturally use these Topics? | Yes; they bridge real vocabulary rather than search-engine keyword lists. | Use phrases from support tickets, sales calls, or Knowledge Gaps. |
| Do numbers keep their units and conditions? | Yes; “24 hours before the scheduled start,” not simply “24.” | Rewrite the fact or FAQ so it cannot travel without its qualifier. |
| Is any excluded material genuinely gone from the processed version? | The exact text is absent after the literal exclusion is applied. | Use the exact fragment, remove it at source, or drop the page. |
| Does this duplicate another authority? | No, or both sources say the same current thing for a clear reason. | Consolidate, update, or deselect the weaker source. |
Saving is not the final step
Section titled “Saving is not the final step”Save Metadata writes the editorial record in WordPress and selects the page. It does not finish the remote update. A changed source should report Pending; then Sync updates the page’s structured knowledge document. If you stop after Save, the chat can continue retrieving the previous synchronized copy.
After synchronization, start a new conversation. Ask one direct question, one question using a synonym from Topics, and one question the page should not answer. Good metadata improves discovery without making the source appear broader than it is.
See the full editor
Section titled “See the full editor”- Capture
- Open the metadata editor for a fictional published page with every metadata section expanded and safe example values filled in.
- Show
- Content preview, Category, Summary, Topics, FAQs, Key Facts, Related Pages, Exclude Content, Save Metadata
- 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
Next editorial move
Section titled “Next editorial move”- Want a first draft? Use AI-assisted metadata with a human review.
- Changed a record and see Pending? Go to Sync and status meanings.
- The right source is still missed? Follow the knowledge investigation.