Unstale - AI Content Refresh, Rewrite & Republish for SEO
| 开发者 | webandseofr |
|---|---|
| 更新时间 | 2026年8月4日 22:31 |
| PHP版本: | 8.1 及以上 |
| WordPress版本: | 7.0 |
| 版权: | GPLv2 or later |
| 版权网址: | 版权信息 |
详情介绍:
- Real AI in the free plugin. Analysis, live web research, fact-checking of numbers and prices, rewrite, SEO metas, JSON-LD when no SEO plugin already emits it, key-takeaway callouts.
- No credits, no quota. Unstale sells you nothing per refresh. You pay your AI provider directly, at its own rates. Unlimited refreshes.
- Your content never touches our servers. Your WordPress talks to the AI provider you connected, directly. We never see your content or your credentials.
- Nothing to configure twice. Since WordPress 7.0, providers and credentials are managed by WordPress itself in Settings > Connectors. Unstale uses whatever you set up there, so no API key is ever stored by this plugin.
- Review queue with a side-by-side diff. Nothing is published without your approval (an optional full-auto mode exists). Native WordPress revisions give you one-click rollback when revisions are enabled on your site.
- Internal linking. Each refresh adds up to 10 relevant internal links with natural, varied anchors, one link per target page.
- SEO-safe republication. Optional publish-date bump with position control, randomized timing, allowed days and time windows. Stable dated permalinks via the %unstale_year%, %unstale_monthnum% and %unstale_day% permalink tags.
- Full journal. Every step, token count and cost is logged per article.
- Automation-ready. REST API (unstale/v1) and WP-CLI (wp unstale refresh) included.
安装:
- Install and activate the plugin. WordPress 7.0 or newer is required.
- Go to Settings > Connectors and connect an AI provider if you have not already. Unstale will tell you if none is connected.
- Go to Unstale > Settings, pick a default model and refresh level.
- In the Posts list, click "Refresh with Unstale" on any published post. Review the estimated cost, confirm, then approve the proposed rewrite in Unstale > Review queue.
屏幕截图:
常见问题:
How much does a refresh cost?
You pay your AI provider directly, per token. Unstale shows an estimate BEFORE each launch and logs the real cost after. For a 1000-word article at the Balanced level the estimate runs from roughly $0.50 to $1.50 depending on the model you pick. That estimate bounds the worst case (it assumes every web search the level allows is spent), so the cost logged after the run is usually lower. Web search fees are part of the estimate but only appear in the logged cost when your provider reports how many searches it ran. WordPress does not standardise that number, so the logged cost can understate searches on some providers.
Does my content go through your servers?
No. WordPress calls the provider you connected, directly. The only request to unstale.ai is the public model catalog (no content, no personal data).
Which API key does it use?
None of its own. Since WordPress 7.0, credentials live in Settings > Connectors and are managed by WordPress. Unstale never sees, stores or transmits them. If you used an older version of this plugin, the API keys it had stored are deleted automatically when you update.
Can it publish without my approval?
Not by default. Every refresh waits in the review queue with a side-by-side diff. You can enable full-auto publishing in the settings if you want it.
My permalinks contain the post date. Will the date bump break my URLs?
No: the bump is skipped automatically when your permalink structure uses the core date tags. To keep dated URLs AND bump dates, switch your structure to the %unstale_year%, %unstale_monthnum% and %unstale_day% tags: they freeze the original date in the URL forever. If you deactivate Unstale later, the plugin automatically restores the core date tags so your site keeps routing (posts whose date was bumped will then use the bumped date in their URL). Deleting the plugin files manually without deactivating skips that safety net: re-save your permalink structure in that case.
Which models are supported?
Whichever models the provider you connected exposes. Unstale ships model cards for Claude Sonnet 5 (default), GPT-5.6 Terra and Gemini 3.5 Flash, with pricing and ratings to help you choose. GPT-5.6 Terra is still in limited preview: if you pick it, Unstale runs GPT-5.5 instead and states the substitution on the model card and in the job log. The catalog is updated remotely: when a provider ships a new model, it appears in your settings without a plugin update.
更新日志:
- Fixed: a refresh could die on an article that has nothing to fact-check. The model was answering correctly, but explaining in a sentence that it had found no figures to verify before giving its answer, and the plugin could not read a reply written that way. Three such replies in a row and the refresh gave up. Those answers are now understood.
- Fixed: a model that answered with reasoning only, and no readable content, killed the refresh outright. The journal said nothing useful about it and the call it had already paid for was never counted. That answer is now treated like any other step failure: the reason appears in the journal, the tokens appear in the cost, and the step is tried again.
- Fixed: the cost shown for a refresh could be lower than what your AI provider actually billed you. A step that failed after the model had already answered dropped the tokens it had just spent, and because a step is tried up to three times, a refresh could pay for three calls and report none of them. Every call that reaches the provider is now counted, on its own line in the journal.
- Fixed: when a model replied with something the plugin could not read, the journal only said the answer was invalid. It now quotes what the model actually sent back, so a failed step can be diagnosed from the journal instead of guessed at.
- Fixed: a refresh that died in the middle of a step always reported the same generic reason, whatever had really happened. The journal now records what the worker actually reported, so a provider error, a memory limit and a scheduler timeout no longer look alike and a failed refresh can be understood from the journal alone.
- Fixed: a refresh could fail for good when the AI provider stopped responding. All three attempts used to fall inside a 90-second window, so a short provider outage exhausted them and the job died. Retries are now spaced 2 and 10 minutes apart, which carries a refresh through an outage instead of losing it.
- Initial release: pipeline on the WordPress AI Client (analyze, web research, fact-check, rewrite, enrich, internal links, cleanup, review, publish), review queue with diff, journal, SEO-safe republication, targeted cache purge, REST API and WP-CLI.