Analysis
Every business with a website has the same problem hiding in plain sight: most of the content is old, nobody is checking it, and Google notices. A pricing page that lists last year's numbers. A "how-to" guide that points to a tool that's since changed its interface. A blog post that used to rank on page one and has quietly slid to page three. None of it is broken enough to set off alarms, which is exactly why it sits there rotting.
The idea behind a tiered refresh system is simple. Not every page deserves the same attention. Your top earners need watching closely; the page three pieces don't. So instead of trying to keep an entire site evergreen by hand, you sort pages into three buckets by how much they matter, then update each bucket on its own clock. The busy pages get checked constantly. The middle gets a weekly pass. The rest gets a proper review four times a year.
What makes this practical now is that the boring parts can be handed to automation. Pulling traffic numbers, ranking pages, deciding which bucket each one falls into, and running the actual updates can run on a schedule with tools like n8n (n8n workflow automation) doing the orchestration and an AI agent doing the writing. You set the rules once and the system keeps your site honest in the background.
The rest of this guide is the build. Fair warning before you copy anything: the tier percentages, the refresh intervals, and the scoring weights below are a recommended setup, not gospel. An hourly refresh on hot content in particular is aggressive, and most teams won't need it that often. Start with the structure, then dial the numbers to your own site.
Analysis
Prerequisites
- Google Analytics 4 or similar analytics
- Content management system (any)
- n8n or similar automation tool
- Claude Code for content generation
- Airtable or database for tracking
Step-by-Step Framework
Step 1: Content Inventory and Scoring
Start by building a full inventory of your pages with the metrics attached. You can't sort pages into tiers until you know how each one actually performs. This script pulls the numbers straight from GA4 and assigns a tier to every page:
# content_inventory.py
import pandas as pd
from google.analytics.data_v1beta import BetaAnalyticsDataClient
from google.analytics.data_v1beta.types import RunReportRequest
PROPERTY_ID = "YOUR_GA_PROPERTY_ID"
def fetch_content_metrics():
client = BetaAnalyticsDataClient()
request = RunReportRequest(
property=f"properties/{PROPERTY_ID}",
dimensions=[
{"name": "pagePath"},
{"name": "pageTitle"}
],
metrics=[
{"name": "sessions"},
{"name": "activeUsers"},
{"name": "averageEngagementTimePerSession"},
{"name": "bounceRate"},
{"name": "conversions"}
],
date_ranges=[{"start_date": "30daysAgo", "end_date": "today"}]
)
response = client.run_report(request)
rows = []
for row in response.rows:
rows.append({
'url': row.dimension_values[0].value,
'title': row.dimension_values[1].value,
'sessions': int(row.metric_values[0].value),
'users': int(row.metric_values[1].value),
'avg_engagement': float(row.metric_values[2].value),
'bounce_rate': float(row.metric_values[3].value),
'conversions': int(row.metric_values[4].value)
})
return pd.DataFrame(rows)
def assign_tiers(df):
"""Assign tiers based on percentile rankings."""
df['session_score'] = df['sessions'].rank(pct=True)
df['conversion_score'] = df['conversions'].rank(pct=True)
df['engagement_score'] = df['avg_engagement'].rank(pct=True)
# Composite score
df['composite_score'] = (
df['session_score'] * 0.5 +
df['conversion_score'] * 0.3 +
df['engagement_score'] * 0.2
)
# Assign tiers
df['tier'] = pd.cut(
df['composite_score'],
bins=[0, 0.5, 0.9, 1.0],
labels=['cold', 'warm', 'hot']
)
return df
# Run
metrics = fetch_content_metrics()
tiered = assign_tiers(metrics)
tiered.to_csv('content_inventory.csv', index=False)
print(tiered['tier'].value_counts())
# hot 45
# warm 180
# cold 225The GA4 calls here are accurate: BetaAnalyticsDataClient and RunReportRequest live in the google.analytics.data_v1beta package and run exactly as shown (Google Analytics python-docs-samples quickstart.py). The dimensions (pagePath, pageTitle) and metrics (sessions, activeUsers, averageEngagementTimePerSession, bounceRate, conversions) are all valid GA4 names too (GA4 Dimensions and Metrics Complete Reference). If you need to set up the API itself, Google's Analytics Data API quickstart covers the auth.
The scoring weights (sessions at 0.5, conversions at 0.3, engagement at 0.2) and the bin cutoffs are my call, not a standard. On a 450-page site those bins land you at roughly 45 hot, 180 warm, and 225 cold, which is where the comment numbers come from. Change the weights if conversions matter more to you than raw traffic.
Step 2: Define Refresh Rules per Tier
With pages sorted, decide what "refresh" actually means for each tier. A hot page needs its prices and stats checked; a cold page needs someone to ask whether it should still exist. Spelling that out in a config keeps the automation honest:
# refresh_rules.yaml
tiers:
hot:
refresh_interval: "1h"
max_age_hours: 2
actions:
- check_price_accuracy
- update_statistics
- verify_links
- refresh_related_content
agent: "content-refresher-v2"
approval_required: false
warm:
refresh_interval: "1w"
max_age_days: 14
actions:
- update_outdated_facts
- refresh_images
- optimise_for_new_keywords
- add_related_articles
agent: "content-optimiser"
approval_required: false
cold:
refresh_interval: "3M"
max_age_days: 120
actions:
- full_content_audit
- seo_analysis
- merge_or_redirect_recommendation
- archive_if_irrelevant
agent: "content-auditor"
approval_required: true
signals:
freshness_degradation:
- bounce_rate_increase: 10
- ranking_drop: 5
- traffic_drop_percent: 20
escalation:
cold_to_warm: "traffic increases 300% over 7 days"
warm_to_hot: "traffic increases 200% over 3 days"
any_tier_refresh: "on manual editor request"Two things worth flagging. The escalation rules matter as much as the schedule: a cold page that suddenly catches fire should jump tiers automatically rather than wait for its quarterly slot. And note that cold-tier actions carry approval_required: true, because "archive this page" or "redirect it" is the kind of call you want a human signing off on. The thresholds themselves are starting points, not numbers handed down from anywhere.
Step 3: Build the Refresh Agent (Claude Code)
This is where you wire up the agent that does the actual rewriting. One correction before you build on this code: the snippet below uses an export default defineSkill({...}) pattern in a .ts file, but that isn't how Claude Code skills actually work. Real Claude Code skills are SKILL.md markdown files inside a directory under .claude/skills/, with a description that drives when the skill runs (Extend Claude with skills - Claude Code Docs). There's no documented defineSkill TypeScript helper. Likewise, the claude.generate({prompt: ...}) call is pseudocode, not a real Anthropic SDK surface. Treat the code below as a structural sketch of a generic refresh agent or script rather than a Claude Code skill you can drop in as-is.
// .claude/skills/content-refresh.ts
export default defineSkill({
name: 'content-refresh',
description: 'Refresh content based on tier rules',
input: z.object({
url: z.string(),
tier: z.enum(['hot', 'warm', 'cold']),
currentContent: z.string(),
lastRefreshed: z.string().datetime(),
metrics: z.object({
sessions: z.number(),
bounceRate: z.number(),
avgTimeOnPage: z.number()
})
}),
async execute({ url, tier, currentContent, lastRefreshed, metrics }) {
// Fetch latest data for hot content
const latestData = tier === 'hot'
? await fetchLatestData(url)
: null;
// Generate refreshed content
const refresh = await claude.generate({
prompt: `Refresh this ${tier}-tier content.
Last refreshed: ${lastRefreshed}
Current metrics: bounce ${metrics.bounceRate}%, avg time ${metrics.avgTimeOnPage}s
Current content:
${currentContent.slice(0, 3000)}
${latestData ?