SEO.co
// blog

Automated SEO Decay Detection With Search Console and AI Agents

Wire Google Search Console to AI agents that flag decaying pages, draft refreshes, and queue approvals so content never quietly bleeds traffic again.

Samuel EdwardsSamuel Edwards
Automated SEO Decay Detection With Search Console and AI Agents

Refreshing content by hunch, pulling up whichever post someone remembered in a meeting, is how teams spend quarters touching pages that never mattered while their ranked pages bleed out. A decay-detection pipeline driven by Search Console data and gated by human review shifts the work to the pages actually losing money.

What follows specifies one: a scheduled pull from the GSC API, a scoring function that flags pages with real revenue exposure, an AI agent that drafts the refresh, and a human approval gate before anything ships.

What Decay Actually Looks Like in the Data

Decay is rarely a cliff. The pattern that eats most organic programs is slower: a post that held position 3 for nine months slides to 6, then 11. Impressions hold for a few weeks before the SERP reshuffles. By the time someone looks at the traffic chart, the page has shed two thirds of its clicks and a competitor has taken its featured snippet.

A Topdog analysis of the top articles across five Swedish content sites found the median piece lost 61% of its traffic one year after publication, and roughly six in ten lost at least half.

The implication for pipeline design: a monthly snapshot will not catch decay in time for a refresh to matter. You need week-over-week deltas against each page's own history, and enough rows to catch the long tail where most B2B conversions live. Pair that with a model for rankings volatility so the scorer separates a real slide from a one-week reshuffle.

Six in Ten Top Articles Lose Half Their Traffic in a Year
Six in Ten Top Articles Lose Half Their Traffic in a YearLost at least half their traffic within a year: 59%; Did not lose half within a year: 41%Lost at least half…59%Did not lose half w…41%
Count the squares: 59 of every 100 top articles shed at least half their traffic within twelve months of publication. Source: Topdog / inCrev, 2023

The Pipeline, End to End

The architecture is small. Six components, each doing one job:

  • A scheduled extractor hitting the GSC Search Analytics API for the last 16 months of page and query data.
  • A warehouse, with BigQuery the obvious one, storing daily snapshots and computing rolling windows.
  • A scoring job that flags pages against a decay function.
  • An orchestration layer that feeds flagged pages to an AI agent.
  • The agent, which drafts a refresh plan per page: new sections, outdated claims to pull, missing schema, internal links to add.
  • A review queue where a human approves, edits, or rejects each draft before publish.

Nothing unusual there. The engineering value is in the plumbing: idempotent runs, proper backfills, clean joins between GSC data and CMS metadata so the agent knows what page it is proposing to change. The shape sits inside what agentic SEO describes: narrow agents with tool access, running against a known pipeline, not an open-ended copywriter loose on the blog.

Pulling the Signal From Search Console

The GSC UI export caps at 1,000 rows. For any site past the brochureware stage, the rows you care about, the long-tail queries where a position shift from 8 to 14 cost you twelve conversions, never surface.

The Search Analytics API lifts that ceiling to 50,000 rows per day per site per search type, which the Search Console team documented in 2022. For enterprise sites, even that is small. Bulk export to BigQuery, announced the following year, ships ongoing daily exports with no row cap.

Search Console Data Access by Interface
Search Console Data Access by InterfaceUI export (per report): 1,000 rows; Search Analytics API (daily, per site, per search type): 50,000 rowsUI export (per report)1,000 rows50,000 rowsSearch Analytics API(daily, per site, per…
The API lifts the 1,000-row UI cap by 50x; BigQuery export removes the daily row limit entirely. Source: Google Search Central, 2022

Pick the lowest tier that holds your actual query space. A 10,000-URL site usually sits in the API. A 100,000-URL site needs BigQuery, both for row count and for the joins. Then set the window: 16 months is the ceiling of what GSC will return, and most decay functions want at least 90 days of rolling comparison.

Scoring Decay So the Agent Doesn't Chase Noise

Not every traffic drop is decay. Seasonal queries fall. Algorithm updates reshuffle. Headlines lose freshness without rank losing ground. The scoring function has to isolate pages where a refresh will actually move something.

A workable score has four inputs:

  • Position slope across the last 90 days, weighted by the inverse of starting position (a drop from 4 to 11 matters more than 44 to 51).
  • Impression decline against the same page's prior-year window, to control for seasonality.
  • Click-through-rate collapse holding position constant, which usually means a competing SERP feature took the clicks.
  • Revenue weight pulled from your analytics stack, so a 20% drop on a BOFU page outranks a 50% drop on a glossary entry.

Treat 404 errors on previously ranked URLs as a hard signal, not something the slope detector should average away. Published refresh work tends to reward pages that were still earning and just started to slip.

Share of Blog Views and Leads From Old vs New Posts
Share of Blog Views and Leads From Old vs New PostsMonthly blog views: 76%; Monthly blog leads: 92%Old postsNew postsMonthly blog views76%24%Monthly blog leads92%
HubSpot reported 76% of monthly blog views and 92% of monthly blog leads came from older posts, which is why refresh programs matter. Source: HubSpot, 2024

Where the AI Agent Fits

The agent's job is narrow: for each flagged page, produce a refresh plan a reviewer can approve in under two minutes. That means fetching the live page, pulling current top-ranking competitors, diffing them for coverage gaps, checking schema and internal link density, and writing a change proposal with before-and-after excerpts.

Model Context Protocol servers are the useful pattern. One MCP server exposes your GSC data, another your CMS, another a crawler, another your analytics stack. The agent calls each tool, assembles context, and produces structured output a human can scan.

Teams already building this end up close to the architecture covered in SEO MCP servers: tools first, prompts second, nothing the agent is allowed to ship without a review step.

The Approval Gate

Automation without review publishes slop. The queue is where the program earns or loses the right to scale.

What to check in each refresh:

  • Does the proposed change actually address the decay signal, or something the agent found convenient?
  • Are claims sourced to pages that exist, or is the agent hallucinating stats and links?
  • Does the restructure create duplication with sibling pages? A refresh competing with another page you own is a cannibalization problem in disguise.
  • Is the URL being preserved, or is this secretly a migration with redirect debt attached?

Approval should take minutes, not hours. If reviewers rewrite every draft, the scoring function is wrong, not the agent.

The One Metric That Tells You It Works

The number to watch: median position delta across refreshed pages, measured 28 days after publish, holding query intent constant. Pair it with a hold-out control of flagged pages you did not refresh, so the signal is the program, not the algorithm.

Backlinko's content relaunch case study moved its target keyword from position 7 to position 4 inside two weeks, with a 260.7% traffic lift on the post. That is the upper end of what a single well-chosen refresh produces. For a program covering 40 or 80 pages a month, the healthy signal is more modest: a two-position median gain within a month, applied consistently.

If after 90 days your refreshed set is not beating the control in median rank, the scoring function is picking the wrong pages. If the control is winning, stop refreshing until you figure out why. A badly scored program can cost more than no program at all.

Samuel Edwards
// written by
Samuel Edwards
In his 15+ years as a digital marketer, Sam has worked with countless small businesses and enterprise Fortune 500 companies and organizations including NASDAQ OMX, eBay, Duncan Hines, Drew Barrymore, Washington, DC based law firm Price Benowitz LLP and human rights organization Amnesty International. As a technical SEO strategist, Sam leads all paid and organic operations teams for client SEO services, link building services and white label SEO partnerships. He is a recurring speaker at the Search Marketing Expo conference series and a TEDx Talker. Today he works directly with high-end clients across all verticals to maximize on and off-site SEO ROI through content marketing and link building. Connect with Sam on Linkedin.