Content Decay: The 31% Refresh Threshold and the Dual Decay Curve

Content half-life, the time from peak organic performance to 50% traffic decline, has compressed from 12-18 months to 3-6 months for competitive topics. The compression is not uniform. Technology content decays fastest (30-40% annual decline) because framework versions and API changes create hard obsolescence dates. Foundational how-to content decays slowest (5-10%) because the underlying processes rarely change.

Ahrefs (2024) found that 66% of pages older than two years experience declining organic traffic. This aligns with their 1.3M keyword study showing the average #1-ranking page is 5 years old. The pages that survive are the minority that received ongoing investment. Only 1.74% of new pages reach the top 10 within a year, down from 5.7% in 2017. The barrier to entry is rising, and the cost of not maintaining what you already have is rising with it.

The pages that survive over years are the ones that keep taking investment. Content built for refresh costs less over its life than content built with an expiration date. The ROI section below quantifies the gap.

A zero-authority domain collapses in under three months

A 16-month SE Ranking experiment (March 2026) quantified one extreme of the decay curve. SE Ranking published 2,000 AI-generated articles across 20 new domains. 70-75% of all impressions and clicks occurred in the first 2.5 months. Rankings collapsed around month 3 (from 28% of pages in the top 100 down to 3%), with no meaningful recovery at month 16. Total clicks across all 2,000 pages over 16 months: 1,381.

The YMYL differential was stark. Finance pages retained only 9 of 100 indexed at month 16, Health only 14 of 100, while broad niches retained near-complete indexing. This suggests decay on zero-authority domains is not uniform. YMYL content faces an accelerated de-indexing curve on top of the ranking collapse.

This pattern represents the fastest observed decay trajectory: content that never accumulates the authority or engagement signals to sustain initial rankings. The 3-6 month half-life cited above applies to content on established domains. On new domains with no link equity, no behavioral signals, and no content differentiation, the effective half-life compresses to under 3 months.

Google evaluates freshness through at least four independent methods, and cross-checks them against each other to catch manipulation. Freshness is not a single timestamp check. Each layer operates on its own, and conflicts between layers degrade the freshness signal rather than amplify it. The 2024 API leak and patent US8549014B2 confirm the architecture.

freshByDocFp reads the content, not the timestamp

freshByDocFp detects a change in the content itself, and ignores a change to the timestamp. lastSignificantUpdate tracks the timestamp of the last substantive revision. Google stores only the last 20 versions of a document, which means high-frequency trivial updates consume version slots without generating freshness benefit.

Three date signals cross-check each other

Google cross-references three date signals: bylineDate (explicitly stated), syntacticDate (extracted from URL structure or title), and semanticDate (NLP analysis of whether information, sources, and data are current). When these conflict, Google overrides explicit dates with semantic analysis. bylineDateConfidence scores how much Google trusts the stated date. Changing the byline date without changing the content does not improve freshness scores and may degrade them.

Link velocity dates the page from outside

freshdocs applies a link value multiplier favoring newer pages. The rate of new and disappearing incoming links signals whether content is becoming stale or freshly relevant.

Renewed clicks re-date a stale document

CTR acceleration on a previously stale document signals renewed relevance. This feeds into NavBoost's click-based re-ranking system.

The freshness infrastructure above operates at the document level. QDF (Query Deserves Freshness) operates at the query level. It detects a shift in what a query needs, then overrides relevance scoring in favor of recency for as long as the shift lasts.

Breaking news triggers the strongest override. Recurring events, such as annual reports and seasonal topics, trigger predictable freshness windows. Queries about frequently changing information, such as pricing, best-of lists, and technology comparisons, trigger a persistent recency preference. The 2011 Freshness Algorithm Update extended the logic to about 35% of all queries.

The diagnostic distinction matters. A page that held position and then dropped may have lost to QDF-triggered results rather than to its own staleness. QDF displacement calls for something newer. Genuine staleness calls for a refresh of what exists. A team that confuses the two wastes the budget. In the clinical diagnostic framework, this maps to whether the binding constraint is at the competition layer (L7) or the content layer (L4-L5).

The most actionable finding from the RepublishAI study (n=14,987 URLs, 20 verticals, 76-day observation window) is that content refresh effectiveness has a threshold. Treatment group pages (n=6,819) received updates of varying magnitude. Control group pages (n=8,168) received no updates.

For a typical 1,500-word article, the 31-100% threshold means adding 500-1,500 words of new content. The net difference versus control: +7.96 positions (+5.45 against a -2.51 decline for non-updated pages). Minor tweaks and moderate expansions do not clear the threshold. The moderate category performed worse than no update at all. A small update may signal low-quality refresh intent to Google's evaluation system.

This aligns with Google's freshness infrastructure. freshByDocFp detects actual content changes, and the age-distribution scoring in patent US8549014B2 weights the ratio of new-to-old content within a document. Minor tweaks do not shift the ratio enough to register as a significant update.

Four changes register as significant

• Substantive content change detected by freshByDocFp
• Shift in the document's age distribution (adding a paragraph to a 3,000-word article changes less than rewriting 40%)
• Semantic date shift detectable by NLP analysis (more current data, newer sources)
• New information gain relative to competing documents

Preserve the signals unless the intent moved

A signal-preserving refresh keeps the URL, holds the query coverage, and adds information gain. Do not cut a section that satisfies a ranked query cluster. Update the structured data dates alongside the content. A signal-resetting refresh changes the URL with a 301, rewrites more than 60%, or consolidates several decayed pages. Reset only when the search intent moved and the page no longer matches it. For programmatic pages with daily data refreshes, Google retains only approximately 20 days of version history due to the 20-version storage limit.

Content decay presents in Google Search Console through a four-stage diagnostic sequence. Each stage is a lagging indicator of the one above it, which means the earlier you detect the signal, the less effort the intervention requires.

Animalz built a detection method into their Revive tool. It analyzes 12 months of organic search traffic and filters out seasonal and algorithm-update noise. It then flags a page after 3 or more consecutive months of decline. It ranks the flagged pages by their share of total organic traffic, at 1% or above.

AI search platforms prefer newer content more aggressively than traditional Google. A Seer Interactive study (5,000+ URLs, 2025-2026) measured the gap directly. 65% of AI bot hits target content from the past year, and 79% from the last two years. The average age of AI-cited URLs is approximately 2.9 years versus 3.9 years for traditional organic results. AI citations are 25.7% fresher overall.

The Ahrefs freshness study (July 2025, 16.975M cited URLs) provides per-platform granularity. Average days since publication for cited content: ChatGPT 958 days, Copilot 1,056, Gemini 1,118, Perplexity 1,166, organic SERP 1,416, and Google AIO 1,432. ChatGPT has the strongest fresh-content preference. Google AIO is the exception, citing slightly older content than organic results. From ChatGPT's top-1,000 cited pages, 76.4% had an update within 30 days, and 89.7% had one during 2025.

Metehan Yesilyurt identified a URL_freshness_score in ChatGPT's retrieval system. Artificial date refreshing (updating a page's visible date without substantive content changes) can shift citation positions by up to 95 ranks. Perplexity and ChatGPT both order in-text references newest to oldest, making publication date a visible sorting criterion beyond its effect on retrieval scoring.

Academic research confirms the pattern. Fang and Tao tested 7 LLM models in arXiv 2509.11353, presented at SIGIR Asia Pacific 2025. Every model promoted the passages marked as fresh. The top-10 mean publication year moved forward by up to 4.78 years. Individual items moved by as many as 95 ranks in listwise reranking. Pairwise preference reversed by up to 25% after the researchers injected a date.

The operational implication: a page maintaining rankings for 12+ months in traditional Google may lose AI citations within 3-6 months. Refresh cadence for AI visibility must be more aggressive than for traditional organic. On any page where AI visibility carries value, the content engineering practices that make a refresh cheap stop being optional. Modular sections, data-driven paragraphs, and a structured update workflow are the ones that pay.

AIO citations turn over 70% in 2-3 months (Authoritas). A page cited in AI Overviews today holds roughly a 30% chance of a citation 2-3 months from now. AirOps (2026) reports similar numbers: only 15% of retrieved pages earn citations, and retention sits at 30% per AI answer over time. The full funnel narrows sharply: retrieval, then 15% citation selection, then 32% text extraction. Roughly 5% of retrieved content reaches the user.

Before AI search, content had one decay curve: organic rankings degraded on a single trajectory you could monitor with position tracking. Now there are two. A page can hold position #3 for its target keyword in organic results while being entirely absent from AI Overviews for the same query. The reverse also happens. A page loses organic rankings and still appears in ChatGPT and Perplexity, because those systems hold an older and more citable version of it. Different signals drive the two channels, and the two clocks run apart.

The four freshness layers above drive organic decay: document fingerprinting, date triangulation, link velocity, and behavioral signals. Three different mechanisms drive AI citation decay. Re-crawl frequency, citation source rotation inside the model, and whether newer competing content clears the information gain threshold for the query.

Writesonic's GPT-5.4 study measured the instability. Source overlap between GPT model versions on identical prompts is only 7%, and 22 of 50 prompts produced zero overlap. The signals between the channels do not align. A page with strong backlinks and stable CTR can hold its organic position while AI systems rotate to fresher sources. A page with no external links and highly structured data can persist in AI citations long after its organic rankings fade.

Two curves need two monitoring cadences

Traditional rank tracking catches organic decay. AI citation monitoring, checking whether your pages appear in AIO, ChatGPT, or Perplexity responses for target queries, catches the second curve. Neither one is a proxy for the other.

Run two cadences. Standard position tracking, weekly or daily for high-priority pages. Periodic AI query sampling, 20-50 target queries a month across the platforms, to measure citation presence and turnover. Authoritas, Profound, and a manual spot-check all work. Track and triage the two signals separately.

When the curves diverge, the intervention differs. Organic decay with stable AI citations suggests the page's content is strong but its traditional ranking signals (links, CTR, technical performance) are weakening. AI citation loss with stable organic rankings points the other way. The content is authoritative and no longer novel, and a competitor published something fresher. The diagnosis decides between a technical fix, a link campaign, and a content expansion.

Pruning (removing, redirecting, or consolidating underperforming pages) concentrates crawl budget, link equity, and quality signals on pages that matter. The mechanism aligns with the API leak's site-level quality scoring: CompressedQualitySignals, siteAuthority, and the Panda site quality ratio (patent US9031929).

The mechanism is likely indirect. Low-quality pages dilute site-level quality ratios, consume crawl budget, and fragment link equity. Pruning removes the dilution, not the age. Google's Danny Sullivan cautioned against pruning for freshness reasons specifically, but the traffic data from multiple independent cases suggests that removing pages which drag down site quality scores produces measurable gains.

For topical authority, pruning has a secondary effect: removing off-topic or thin content tightens the site's siteFocusScore and reduces siteRadius, reinforcing the domain's semantic coherence in Google's topic embeddings.

HubSpot reported a 106% increase in organic search views to historically optimized posts and double the monthly leads. By 2018, they dedicated a full-time staff member to monitoring and optimizing 10+ years of archives. AirOps (2026) estimates that content refreshes deliver 3-5x higher ROI than new content, because updated content retains existing authority signals and backlink profiles. Contentoo/HubSpot data shows refresh lowers content production costs by over 80% compared to net-new production while delivering comparable or better traffic results.

Demand Metric reports that evergreen content delivers 4x ROI compared to time-sensitive or seasonal content over its lifecycle, which makes the maintenance math even more favorable: the content most worth refreshing is also the content with the highest baseline return. Budget allocation guidance from multiple sources converges on 20-30% of content investment allocated to refresh and maintenance. Most organizations spend close to 0%, which means they are building an asset that depreciates faster than they are adding to it.

Cadence follows the tier of the page

For AI search visibility, review a high-priority page every 90 days, whatever its traditional rankings do, because of the 70% citation turnover rate. That cadence is the largest operational difference between the two channels. The maintenance load roughly doubles.

The two findings in the headline answer two different questions. The 31% threshold sets how large a refresh has to be before Google registers it. The dual decay curve sets how often the refresh has to happen, on two clocks that do not agree.

Both answers point the same way. A team that edits a paragraph every quarter pays the full cost of the work and clears neither bar. The RepublishAI data puts a moderate update below the do-nothing control. The 70% AI citation turnover erases the gain from a quarterly cycle before the next one starts. Light maintenance is the most expensive cadence available.

A page that carries revenue needs a substantive rewrite of a third of its body, on a cycle near 90 days. Check the result on both channels. That bill is heavy on hand-edited content. The same bill is light on a page with modular sections and numbers that come from a data asset. There the refresh runs as a pipeline instead of an assignment.

Content is a depreciating asset, and the depreciation rate is rising. The cheapest answer is not more new content. The cheapest answer is the maintenance infrastructure that holds the existing pages above the threshold. Monitoring on both curves, a refresh workflow, and an editorial process built around substantive updates rather than one-time publication.