Refresh Decaying Blog Posts With an AI Agent and n8n

The short answer: If last year's most-read posts are losing clicks and nobody has the hours to work out which ones and why, give the finding and the first draft to an agent and keep the decision with an editor. On n8n Starter at $24 a month billed monthly (n8n pricing) and NEURONwriter Gold at $69 a month, the first plan with its API (NEURONwriter pricing), the setup costs $93 a month plus the model, checked on 26 October 2026. For example, if picking and planning one refresh by hand takes 90 minutes and you refresh 8 posts a month, the agent takes 8 x 90 = 720 minutes of hunting off someone's month, and the 8 yes-or-no calls stay with a person.

Disclosure: I may earn a commission through some links on this site. It does not change what this page says.

A post that ranked and then slid still has its URL, its links and its history with Google. What changed is the field around it: newer pages answer the question more fully, or the facts in it have aged. Search Console already shows which pages are slipping. Reading it every week is the chore that slips first.

Before you build:

1. Work out the meter. n8n counts one execution each time a schedule fires, whatever happens inside the run (executions), so a Monday run uses 4 or 5 of Starter's 2,500 a month. NEURONwriter is the meter that matters: every /new-query call counts as one analysis (API FAQ) against Gold's 75 documents a month, so five pages a week for four weeks spends 5 x 4 = 20.
2. Start from what you have. A Search Console property you can read, the CMS your editor already publishes from, a Slack channel for approvals, and a one-page refresh policy: what the agent may suggest, what it must flag for a person, and what it never touches, such as prices, quotes and legal lines.
3. Take the before picture. In the Performance report, open the Date filter and choose Compare (Search Console Help), set the last three months against the three before, and export the Pages tab; exports come as Google Sheets, Excel or CSV (Google Search Central). Every later number is read against that file.

The workflow, run every Monday

1. Trigger. A Schedule Trigger set to Weeks fires on Monday morning. The workflow has to be published for the schedule to run.
2. Read Search Console. n8n has no Search Console node, so this step is glue. An HTTP Request node with a generic OAuth2 credential (HTTP Request credentials) posts to the Search Analytics endpoint, which accepts only OAuth 2.0 and offers a read-only scope, webmasters.readonly (authorising requests). Make two calls grouped by page: the last 28 days, and the same 28 days one quarter earlier. The query reference caps a response at 25,000 rows, sorts by clicks and returns the top rows rather than every row, so the smallest pages fall away; for this job that is the right bias.
3. Pick five. A Code node joins the two windows, computes each page's change in clicks and keeps the five largest losers that still earn impressions. My rule of thumb: skip anything refreshed in the last 90 days, because its new version has not settled.
4. The agent reads the page and its competition. An AI Agent node takes the refresh policy as its system message, and each candidate with its top queries from a third Search Console call, grouped by query and filtered to that page. Its tools are HTTP Request nodes, which n8n lists among the tools an agent can call (Tools Agent), pointed at NEURONwriter's API with the X-API-KEY header. The agent runs /new-query on the page's main query, polls /get-query until the status reads ready, usually about a minute later, and sends the live URL to /evaluate-content, which scores the page without saving anything (API FAQ).
5. It writes a plan, not a page. The output is short: the queries the page no longer answers, the recommended terms and subtopics it lacks, every dated fact a person must check, and one proposed title.
6. The editor says yes or no. The only tool that writes is /import-content, which saves a new revision into the NEURONwriter editor, and it sits behind n8n's human review step in Slack (human-in-the-loop). The editor sees the page, its click loss and the plan before deciding. A denial cancels the call and tells the agent, and the system message says what comes next: stop, and log the page as skipped.
7. A draft lands, and a person publishes. On approval the agent drafts the revised sections and imports them as a revision. The editor rewrites what reads wrong, checks every fact, and publishes to WordPress through Gold's integration. NEURONwriter's MCP page states that its agent connection cannot publish to WordPress, and the API's method list has no publish call, so the last step stays human by construction.
8. The log. A Google Sheets row per page: the date, clicks before, the decision and who made it.

The Slack credential and an error workflow are set out in how to use n8n. Working a term list by hand, page by page, is in how to use NEURONwriter.

The free version: Search Console, a sheet and an hour a month

None of the tools above is needed to start. Search Console is free (About Search Console), and the baseline export is already a candidate list: sort it by the change in clicks and take the five largest losers that still show impressions. For each one, filter the Performance report to that page and read its queries, which are the questions Google currently sends it. Open the page beside that list and mark what it no longer answers. Then search the main query and read the three results above you. Write the plan in the sheet, let the editor approve it, rewrite, republish.

A free chat assistant can draft the gap list if you paste in the page and its queries. Treat that draft as you would the agent's: a person checks every claim. Google's questions for site owners include whether the date of a page is changed to make it seem fresh when the content has not substantially changed (helpful content guidance), so the date moves only when the substance did.

Done monthly, that is enough on its own. The paid route adds the habit: a list that arrives every Monday without anyone remembering to make it, with the ranking pages' terms and subtopics already pulled.

What it costs, and the two shortcuts

Two vendors sell the detection as a feature rather than a workflow. Frase's Content Guard watches pages through Search Console and, when one drops, drafts better titles, headings and internal links, but only for articles created in Frase and published to a live URL. That suits a library built there, not a back catalogue written somewhere else. Surfer lists Rank Drop Detection from its Standard plan, and the in-tool route is in how to use Surfer.

The agent route comes to $24 + $69 = $93 a month before the model. Frase Starter alone is cheaper at $49 a month (Frase pricing), provided your posts were written in Frase.

Frase's Content Guard settings include an "Auto-apply safe fixes only" mode that ships meta titles and descriptions without asking (Content Guard), and a platform priced by pages watched and articles shipped looks most useful when its changes go live without waiting for anyone. Keep "Always ask me" on. A title is a promise to the searcher, and a person should make it.

How to know it worked

The measure is clicks per refreshed page over 28 days, read six to eight weeks after republishing and set against the baseline export. Position alone misleads, because a page can climb two places on a query nobody clicks.

Take a holdout, because seasons and Google updates move every page at once. Each month, leave every second candidate on the list untouched for eight weeks. If the refreshed pages recover and the held-back ones do not, the refresh did it. If both recover, it was the calendar. The newest days in the report can be preliminary and still change (Performance report), so end each comparison window a few days back.

Read the editor's verdicts as well. When more plans come back denied than approved, the refresh policy needs work, not the editor.

When the wiring is not the hard part

The workflow above is a handful of nodes. The harder question sits one level up: which content jobs a team should hand to agents at all, and who signs off on each. That is the agentic marketing work I sell, so this is the one paragraph on this page that is in my interest.

FAQ

Can the agent publish the refreshed post itself?

Not in this design. Its only writing tool saves a revision inside NEURONwriter, behind a human review step in Slack, and NEURONwriter says its agent connection cannot publish to WordPress (MCP page). Publishing stays the editor's click.

How many posts should it flag each week?

As many as your editor can rewrite properly; five is my starting point. At five a week, Gold's 75 documents a month (NEURONwriter pricing) still leave room for new articles.

Does a refreshed post need a new published date?

Only if the content changed in substance. Google's helpful content guidance asks site owners whether they change dates to make pages seem fresh. Show an updated date on a real rewrite and leave it alone for a typo fix.