Use AI as a metadata draftsperson
Suggest with AI can turn a long page into a Category, Summary, Topics, FAQs, Key Facts, Related Pages, and exclusion suggestions. It is deliberately separated from saving and synchronization. The model proposes; the administrator edits, approves, saves, and publishes.
When the drafting assistant earns its keep
Section titled “When the drafting assistant earns its keep”Use it when a source is substantial, accurate, and difficult to summarize quickly; when you need metadata in one of the supported languages; or when an older page needs an editorial starting point. Manual editing is usually faster for a short contact page, one precise policy, or a record that only needs two alternate Topics.
Do not use generation to discover what your organization’s policy should be. It can summarize available text, but it cannot approve prices, deadlines, legal positions, medical advice, or operational promises.
What leaves WordPress when you ask for a draft
Section titled “What leaves WordPress when you ask for a draft”The request requires the regular OpenAI API key. SmartSite prepares the page’s processed content, capped at 24,000 characters, along with existing metadata, optional guidance, the requested language, and up to 12 candidate internal related pages. That material is sent to OpenAI to produce structured JSON. The request uses the plugin’s configured metadata-generation implementation; it is not a local-only analysis.
The model can only choose Related Pages from the candidate list it receives. A missing relation may simply be outside that candidate set; it does not prove the page is irrelevant or broken.
The drafting controls, without mystery
Section titled “The drafting controls, without mystery”| Control | What it changes | Editorial use |
|---|---|---|
| Language | Requests English, Dutch, German, Spanish, French, Croatian, Italian, or the plugin default. | Choose the language visitors use for questions. Keep official product names unchanged where translation would be misleading. |
| Guidance | Adds up to 400 characters of context to this request. | State scope or vocabulary: “Focus on commercial customers; keep plan names in English.” Do not paste new policy here. |
| Generate / retry | Sends a new OpenAI request and can incur cost. | Retry after improving the source or guidance, not repeatedly in hope that one answer will look more favorable. |
| Preserve existing values | Fills empty single-value fields and merges list-style suggestions into current form values. | Best when an editor already created trusted metadata and wants additions without discarding it. |
| Replace existing values | Applies the reviewed suggestion in place of the current form values. | Use for a deliberate rewrite after comparing the old record. Copy anything you must retain before applying. |
| Apply to Form | Moves the chosen draft into the editable metadata form. | This is not Save and not Sync. Continue editing before committing it. |
Preserve or Replace: make this choice before applying
Section titled “Preserve or Replace: make this choice before applying”Preserve existing values is conservative. Existing Category and Summary remain when already filled, while list-style fields such as Topics, FAQs, Key Facts, and Related Pages are merged with suggestions. It is the usual choice when a human has already curated the record.
Replace existing values is a reset. Choose it when the current metadata is obsolete or poor, and only after recording anything worth keeping. Replace is not inherently “cleaner”; it can throw away valuable alternate vocabulary or carefully qualified facts.
Neither option saves to the database. It changes the form. You can still correct, remove, or rewrite every applied value before selecting Save Metadata.
The output has boundaries—not a guarantee of quality
Section titled “The output has boundaries—not a guarantee of quality”| Field | Output limit | What the limit does not guarantee |
|---|---|---|
| Category | 80 characters | That the grouping matches your editorial taxonomy. |
| Summary | 650 characters | That every claim is supported or that the summary is concise enough for your readers. |
| Topics | Up to 12; each up to 60 characters | That the phrases are real visitor language rather than repetitive synonyms. |
| FAQs | Up to 6; question up to 160 and answer up to 320 characters | That an answer preserves every exception or remains current. |
| Key Facts | Up to 8; each up to 180 characters | That a number keeps its unit, date, market, or eligibility condition. |
| Related Pages | Up to 6, chosen from prepared candidates | That all useful relationships were available in the request. |
| Exclude Content | Up to 600 characters | That suggested phrases match processed text or are safe as redaction. Exclusions remain literal text removal. |
These are technical output constraints, not editorial targets. A page may need three Topics and no FAQ. More generated material is not automatically more retrievable.
Review the draft like an editor
Section titled “Review the draft like an editor”- Read the source first. Know the approved answer before judging the summary.
- Choose language and short guidance. Explain audience or terminology, not desired facts.
- Generate once. Read the entire review; do not jump straight to Apply.
- Fact-check nouns and numbers. Product names, dates, prices, locations, units, and exceptions deserve deliberate comparison.
- Challenge every FAQ. Ask where its answer appears in the source. Remove an answer that relies on assumption.
- Inspect Related Pages. Keep only connections that help complete the visitor’s journey.
- Treat exclusions skeptically. Verify exact matching text and never use the field to protect sensitive data.
- Choose Preserve or Replace. Make the decision based on the existing record, then Apply to Form.
- Edit in the form. Remove repetition, restore qualifiers, and make the language sound like your organization.
- Save, sync, and test. The draft becomes useful only after both Save Metadata and a successful synchronization.
Long pages need a sharper human brief
Section titled “Long pages need a sharper human brief”The processed source included in one request is capped at 24,000 characters. On a long page, a footer, repeated navigation, product grid, or several unrelated sections can crowd the material the draftsperson sees. Do not respond by putting the missing conclusion into Guidance and accepting it as fact. Split an over-broad page into authoritative subjects, improve its opening and headings, or write the metadata manually from the approved source.
Guidance is most useful when it resolves editorial ambiguity: audience, terminology, geography, or which part of the page deserves focus. “Write an FAQ saying everyone qualifies” is not guidance if the page does not establish that rule. The clearer the approved source and the narrower the brief, the less time a reviewer spends removing confident-looking invention.
A review example: attractive but unsafe
Section titled “A review example: attractive but unsafe”Imagine a suggestion says: “Customers can cancel free of charge within 24 hours.” The page actually says cancellations are free when received at least 24 hours before the scheduled appointment, except for a named service. The draft sounds natural but changes the rule. Rewrite it with the timing direction and exception intact.
Likewise, a generated Key Fact of “Delivery takes 3–5 days” is incomplete if the source specifies business days, one country, and in-stock products. Retrieval often lifts concise facts because they look useful. Concise metadata therefore needs more care around qualifiers, not less.
Why a suggestion can be disappointing
Section titled “Why a suggestion can be disappointing”| What you see | Likely reason | Better response |
|---|---|---|
| Generic summary and topics | The source is thin, repetitive, or its opening text contains little substance. | Improve the WordPress page; generation cannot recover facts that are not present. |
| Wrong language or translated brand terms | Language/default choice or guidance did not match editorial expectations. | Select the intended language and say which names must remain unchanged. |
| A useful related page is absent | Only a bounded set of candidate pages was supplied. | Add the relationship manually after Apply, provided it is genuinely relevant. |
| The request errors or returns malformed content | The API key, service response, source content, or structured output failed. | Keep the existing form untouched, check API configuration, shorten guidance, and retry once with a meaningful source. |
| Good draft disappears after closing | Apply only changed the form, or the form was never saved. | Apply, edit, Save Metadata, then synchronize. |
Cost and repetition
Section titled “Cost and repetition”Every generation request is an OpenAI API operation. The cost of one draft may be small, but bulk retries without editorial purpose add up. A sound workflow is “improve source → generate → review,” not “generate six times → choose the friendliest answer.” If multiple pages share the same structural weakness, fix your source template or editorial process rather than asking the model to compensate page by page.
The review screen
Section titled “The review screen”- Capture
- Show the AI metadata review state with a fictional summary, two topics, one FAQ, and the merge/replace choice visible.
- Show
- Language, guidance, suggested fields, application mode, Apply button
- 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
Take the draft to production carefully
Section titled “Take the draft to production carefully”- Learn the manual field meanings in Edit pages like a knowledge librarian.
- Publish the reviewed record through Sync and status meanings.
- Check whether it improved a real question with evidence-led troubleshooting.