Reference Freshness Decay

The mathematical backbone of freshness decay, where citation weight decreases exponentially over time. Half-life parameters are calibrated per domain—breaking news may have a 6-hour half-life, while legal precedents might use a 5-year curve. The decay constant λ determines the rate: weight(t) = e^(-λt). This prevents stale references from dominating retrieval results while allowing evergreen content to maintain residual authority.

Decay rates must be tuned to the information velocity of each knowledge domain:

• Fast-decay domains: News (hours), social media trends (days), stock prices (minutes)
• Medium-decay domains: Technology documentation (months), medical research (1-3 years)
• Slow-decay domains: Mathematical proofs (decades), historical analysis (centuries)

Misaligned calibration causes either premature obsolescence of valid references or persistence of outdated information.

Freshness decay engines must distinguish between original publication date and last-modified timestamp. A document published in 2019 but substantively revised in 2024 should receive a decay reset based on the update, not the origin. This requires parsing HTTP Last-Modified headers, sitemap lastmod fields, and in-page revision metadata. Partial updates may receive a partial decay reset proportional to the scope of revision.

A critical tension exists between temporal relevance and source authority. A newly published article from an unknown source should not automatically outrank a slightly older reference from an established authority. Sophisticated decay models apply a composite score that multiplies freshness weight by baseline authority: final_score = authority_score × freshness_weight. This prevents recency bias from undermining citation quality.

Decay functions activate selectively based on query temporal intent classification:

• Recency-sensitive queries: 'latest,' 'current,' 'today,' '2024' trigger aggressive decay
• Evergreen queries: 'how to,' 'definition of,' 'history of' suppress or disable decay
• Hybrid queries: 'best practices 2024' apply moderate decay

Misclassification leads to inappropriate temporal weighting and degraded answer quality.

Beyond gradual decay, systems implement hard staleness thresholds where citations crossing a maximum age boundary are fully deprecated. For example, medical treatment guidelines older than 10 years may be excluded entirely regardless of prior authority. These thresholds are often paired with deprecated knowledge markers—explicit signals flagging superseded content to prevent AI models from surfacing obsolete recommendations.

A ranking mechanism that assigns higher authority weights to recently published or updated content for time-sensitive queries. The algorithm applies a decay function—often exponential or logarithmic—to reduce the influence of older references.

• Freshness boost: Content published within the last 30 days receives maximum weight
• Half-life decay: Authority halves at a rate determined by query temporality (e.g., news = hours, research = years)
• Update detection: Republishing with new timestamps without substantive changes is penalized as freshness spoofing

The automated categorization of search queries by their time-sensitivity profile, which determines the aggressiveness of the freshness decay function applied to candidate citations.

• High temporality (Q1): Breaking news, stock prices, weather—decay measured in hours
• Medium temporality (Q2): Product reviews, regulatory updates—decay measured in months
• Low temporality (Q3): Historical facts, mathematical proofs—minimal or zero decay applied
• Mixed temporality: Queries requiring both evergreen foundations and recent developments

A preprocessing step that converts absolute publication dates into relative age scores for fair comparison across a candidate set. Normalization prevents newer but substantively weaker content from outranking older, authoritative sources in low-temporality contexts.

• Z-score normalization against the median age of top-ranked documents
• Bucketization: Grouping content into recency tiers (last 24h, last week, last month, last year)
• Recency floor: A minimum authority retention for seminal works regardless of age

A machine-readable indicator of a document's revision frequency and pattern, used to predict future freshness reliability. Consistent, substantive updates build a higher baseline authority that decays more slowly.

• Regular cadence: Weekly or monthly updates signal active maintenance
• Event-driven updates: Revisions triggered by external events (regulation changes, version releases)
• Staleness risk: Documents with erratic or abandoned update patterns face accelerated decay
• Change log transparency: Published diff histories increase trust in update authenticity

A content design strategy that minimizes freshness decay by structuring information to remain relevant and accurate over extended periods. Evergreen content resists temporal depreciation through foundational principles rather than time-bound specifics.

• Principle-based framing: Teaching concepts over listing current-state facts
• Modular design: Isolating time-sensitive sections for targeted updates without full rewrites
• Forward-compatibility: Avoiding references to specific versions, dates, or transient states
• Self-maintaining elements: Dynamic data pulls that auto-update statistics and figures

A manipulative practice where content is superficially refreshed—through minor edits, timestamp updates, or cosmetic changes—to artificially reset the freshness decay curve without adding substantive new information.

• Timestamp spoofing: Changing publication dates without meaningful content revision
• Content churn: Minor rewording or reordering to trigger recrawling
• Detection mechanisms: AI models compare document hashes and semantic similarity across versions
• Penalty: Sources flagged for recency manipulation face accelerated decay and reduced citation trust