Content decay: detecting and fixing SEO traffic decline with n8n and AI

A page that ranked in position 4 for a year and slowly slides to position 9, with no penalty ever showing up in Search Console: that's the most common, and most silent, form of content decay. Nobody touched anything, nothing technically "broke," and yet clicks keep shrinking month after month. The issue isn't that an algorithm changed — it's that the rest of the web kept moving while the page stayed frozen. This guide builds an n8n pipeline that detects this decline before it becomes critical, diagnoses it with an AI Agent, and routes it into a refresh backlog reviewed by a human rather than publishing automatic fixes.

Content decay is measurable, not a feeling

The idea that content "ages" on the web isn't just an SEO writer's intuition. A landmark study by Ntoulas, Cho, and Olston, What's New on the Web? The Evolution of the Web from a Search Engine Perspective (WWW 2004), tracked the weekly evolution of several hundred sites over a full year: roughly 8% of pages get replaced every week, and nearly a quarter of links change over the same period. In practice, a page that hasn't moved since it was published doesn't just stay "stable" — it mechanically falls behind an environment that keeps renewing itself.

This isn't just a contextual effect: search engines explicitly bake freshness into their relevance models. The study by Dong et al., Towards Recency Ranking in Web Search (WSDM 2010, Yahoo Labs team), describes a classifier that detects recency-sensitive queries and adjusts ranking accordingly using dedicated temporal signals. Freshness isn't just a vague editorial best practice, then — it's a ranking signal explicitly modeled for a meaningful share of queries.

Content decay can be defined precisely from this: a progressive decline in clicks, impressions, or average position on a page, measured between two comparable periods — not a one-off shock tied to an algorithm update, which hits a whole site rather than a single page.

Step 1 — Pull two comparable periods from the Search Console API

The workflow starts with a monthly Schedule Trigger, followed by two calls to the Search Console API's searchanalytics.query endpoint: one for the last 90 days, one for the 90 days before that, both with the page dimension and a high rowLimit to cover the whole site. OAuth2 credential setup and API quotas are covered in our weekly SEO report with Search Console guide — this pipeline reuses that same authentication base without repeating it here.

One thing to watch: both calls need to use exactly the same filters (search type, country, device) to stay comparable. A filter left off one of the two calls silently skews the whole decline calculation that follows.

Step 2 — Calculate the decline and filter out the noise

A Code node merges the two result sets on the page URL, then computes the change in clicks, impressions, and average position between the two periods for each one. Two filters keep the workflow from just surfacing statistical noise:

• an impressions floor on the baseline period (100 is a reasonable starting point): below that, natural variance far outweighs any real decline signal;
• an actual decline threshold, for example -25% clicks between the two periods.

It's worth splitting two distinct decline signatures at this stage, since they call for different fixes:

• impressions and clicks drop together → the page is genuinely losing visibility in results, a sign of a real content or ranking gap;
• impressions stay stable but clicks drop → the page is still shown but clicked less, often a sign that the title or meta description has become less compelling next to refreshed competitor snippets. Our guide on generating meta descriptions with AI covers this specific fix, which is faster than a full rewrite.

Step 3 — Diagnose with an AI Agent before touching the content

For each page that crosses the threshold, an HTTP Request node (or a dedicated scraper as described in our AI web scraping guide) pulls the page's current content. An AI Agent then receives that content, the decline metrics, and, where possible, an excerpt of the snippet currently ranking first for the target query. Its job isn't to rewrite the page — it's to produce a structured diagnosis: an outdated statistic cited in the text, a section that's grown too thin next to more thorough competitors, cannibalization by another page on the same site, or no clear cause, which flags the page for manual review instead.

A Structured Output Parser (see our dedicated guide) forces the agent to return a consistent format — verdict, likely cause, suggested concrete actions — usable page after page without manual reformatting.

Step 4 — Route to a refresh backlog, never straight to publishing

Each page's diagnosis then flows into a dedicated Notion or Airtable base, acting as a refresh backlog sorted by potentially recoverable clicks (baseline-period clicks minus current clicks). A condensed summary goes out on Slack in parallel, with a Wait node that holds any action until an editor has approved the priority and the angle of the refresh — the same human-approval mechanism detailed in our Wait and Slack guide. This point isn't negotiable: letting an AI pipeline automatically republish a page that's already generating real traffic turns a useful diagnostic tool into an uncontrolled regression risk.

Common pitfalls

• Confusing seasonality with real decline: always compare the recent period to the same period a year earlier, in addition to the 90-day comparison, to rule out normal cyclical swings (a "back to school" guide that naturally dips every summer, for instance).
• Threshold set too low: without an impressions floor, the workflow drowns real alerts under dozens of false positives from marginal-traffic pages.
• Isolated decline vs. site-wide decline: check the overall site click trend before analyzing page by page — a widespread decline points to an algorithm update or lost domain authority, not a content problem to fix one page at a time.
• Forgetting internal cannibalization: a page that "declines" may actually have been replaced in results by another, more recent page on your own site. Check for this before launching an unnecessary refresh.

Going further

If your content is already indexed in a pgvector store for RAG use, as described in our Supabase guide and used in the workflows of the RAG Assistant Pack (€119), that same index can also automatically surface evergreen pages on your site worth linking to during a refresh, with no manual search needed. To also track what competitors are publishing on the same topics and enrich the AI Agent's diagnosis, our automated competitive monitoring guide pairs naturally with this detection pipeline.