Content Decay in the LLM Era: When Evergreen Stops Being Evergreen

By ScaleGrowth Digital Editorial

The evergreen content theory was simple. Write a definitive page on a topic, build links to it, refresh it occasionally, and watch it compound for years. The retrieval era has broken parts of that theory. A page that ranked stably for three years can drop out of LLM citation sets in a single quarter without losing its classical position. Content decay was always a real phenomenon. The new question is which decay patterns matter, on which surfaces, and what an editorial calendar should look like in response.

The Working Thesis

Three changes have compounded. Retrieval models cite recent content disproportionately on topics the model classifies as time-sensitive. Schema markup drift accumulates silently over time as CMS edits, plugin updates and template changes introduce small inconsistencies between schema and prose. Entity records on Wikidata, Crunchbase and other reference sources go stale and the brand entity record falls out of sync with the operational reality. Each of these is invisible to classical rank tracking. Each of these reduces AI mention rate measurably.

The Four Decay Patterns That Actually Matter

Pattern 1. Topical recency decay on fast-moving categories

A page on “best ATS for India tech hiring” published in 2023 carries a structural disadvantage in 2026. Not because the recommendations have necessarily changed. Because retrieval models treat the topic as time-sensitive and weight recency. Perplexity weights this particularly aggressively. ChatGPT Search weights it less aggressively but still measurably. The page may still rank in Google. It progressively stops appearing in agentic-origin citations.

The categories where this decay matters most: technology buyer-shortlist queries, regulatory and compliance content, salary and pricing benchmarks, market sizing, vendor comparisons, statutory rates and gazette-driven content. Categories where it matters less: definitional content, historical context, evergreen how-to where the underlying technique has not changed.

Pattern 2. Schema-prose drift accumulated over many edits

The day a page is published, schema and prose are usually aligned. Over 18 months of editorial updates, plugin version bumps, template changes and CMS edits, small disagreements accumulate. A revised price in the body that did not propagate to the Offer block. A new author byline in the prose that did not update the Person schema. A FAQ added in the visible page that did not get added to the FAQPage block. None of these are visible to a casual review. Each of them lowers extraction confidence on the page in retrieval models. A 25,000 page NBFC audit ScaleGrowth ran surfaced 224 invalid structured-data items across the indexed footprint, and the schema-prose drift class accounted for the largest share. The retrieval consequence was an 8 percent ChatGPT mention rate against an Authority Score of 64.

Pattern 3. Entity record staleness

A Series A SaaS that has since raised a Series C, hired 200 more employees, opened offices in two new cities, and added two new product lines. The corporate site reflects all of this. The Wikidata record still says Series A, 50 employees, one office. The Crunchbase entry is two years out of date. The LinkedIn company page is current. Retrieval models that resolve the brand entity across multiple sources detect the inconsistency and lower confidence on the brand entity. The brand starts losing citations on queries where entity verification matters (institutional buyers, regulated industries, sales-led B2B shortlists).

Pattern 4. Internal link rot inside the surviving content

The same NBFC audit surfaced 4,431 broken internal links across the 25,000 page footprint and a 78 percent hreflang error rate across 4,330 tagged links. Internal link rot operates as a slow signal degrader. Each broken link is a small subtraction from the topical authority graph the site is building. The aggregate effect over 18 months is measurable on classical rank for cluster-dependent queries and on AI citation density across the site.

What Doesn’t Decay the Way Conventional Wisdom Says

Two patterns are commonly assumed to drive decay and on close measurement do not.

The first is publish date alone. A 2021 page with current schema, current entity references, recent dateModified, and an internal link graph still pointing to live targets, does not decay materially on classical search and decays slowly on retrieval citation. The publish date matters as one signal among many. It is not the dominant variable.

The second is total word count. The “thin content” pattern that classical SEO worried about does not map cleanly to retrieval citation. A 600-word page with one factual claim per sentence and clean schema can outcite a 3,000 word page padded with anecdote, marketing copy and AI-tells. The retrieval model rewards extractable factual structure, not raw length.

A Decay-Audit Frame

What a Refresh Calendar Actually Looks Like

Editorial calendars in the retrieval era look different from editorial calendars built for classical SEO. The shift is from publishing cadence as the primary metric to refresh cadence as the parallel primary metric. The cadence depends on topic category.

High time-sensitivity categories (vendor shortlists, salary benchmarks, regulatory rates, pricing comparisons, market sizing): refresh on a quarterly schedule. Update the visible body. Update dateModified in schema. Re-validate sameAs targets. Re-check internal links. Adjust the FAQPage block if the FAQs have changed.

Medium time-sensitivity categories (how-to guides on stable techniques, methodology explainers, framework descriptions): refresh on a semi-annual schedule. The check is lighter. Focus on internal link integrity and schema-prose match. Body content usually does not need substantive change.

Low time-sensitivity categories (definitional content, historical context, evergreen reference): refresh on an annual schedule or on a trigger basis when the underlying topic shifts. Annual is sufficient for dateModified and internal link health.

The realistic constraint on this is editorial team capacity. A 200-page commercial site can sustain quarterly refresh on the top 30 commercial pages, semi-annual on the next 60, annual on the remainder. The 794-brief content engine ScaleGrowth ran for a major NBFC ships briefs on the publish cadence the editorial team can absorb. The same engine pattern, applied to refresh, sustains the cadence above. Detail on the engine sits on our programmatic SEO page.

What the Engagement Data Says

The clearest decay measurement reference in the source pool is an industrial-materials manufacturer engagement where category-level traffic patterns showed sharp divergence. One product category was converting at 21.1 percent but losing traffic at 17.8 percent quarter on quarter. Another category had 278 sessions and zero conversions. A third (the highest-converting at 74.7 percent) was operating quietly below the radar because its content was older and was not being included in the refresh rotation. The intervention sequence reordered the refresh priority by conversion-weighted decay risk, and the higher-converting categories got the editorial calendar’s first quarter.

A separate measurement reference: the same manufacturer’s GA4 tracking work caught a silent tracking break that dropped daily pageviews from 527 to 24 on Dec 11 with engagement rate falsely depressed from 72 percent to 52 percent. Decay analysis that runs against a broken baseline returns nonsense. Tracking integrity is a precondition for any decay audit that intends to be operationally useful. The integrity check sits inside the diagnostic phase on our technical SEO audit page.

Five Actions a Practitioner Can Run Monday

1. Tag your top 100 commercial pages by topic time-sensitivity. High, medium, low. The tagging takes a day. The refresh calendar falls out of it.
2. Run a schema-prose diff on the high time-sensitivity pages. Flag and reconcile every drifted field. Republish with current dateModified.
3. Audit your Wikidata and Crunchbase records. If either is more than 12 months behind current corporate state, update them. This single move can shift AI mention rate on entity-sensitive queries within two weeks.
4. Run a broken internal link crawl against the full site. Reconcile against the current sitemap. Replace broken internal links with surviving targets.
5. Set up an AI mention rate baseline if one does not exist. Without it, decay is invisible until the pipeline number moves. The baseline methodology sits on our AI visibility audit page.

FAQ

Does refreshing a page’s dateModified without changing content help with AI citation rate?

Marginally and only when paired with a body update that actually changes something material. Retrieval models check the consistency between dateModified and the underlying content delta. A dateModified bump with no body change is detected and may be discounted. The reliable pattern is a real refresh of body content, even modest, paired with the dateModified update.

How does internal link rot affect AI mention rate specifically?

Internal link rot lowers the topical authority graph the brand site builds for itself. Retrieval models do not parse internal links the way Google’s link graph does, but the consequence shows up indirectly. Pages that depend on cluster authority for entity confidence lose citation probability when the authority graph degrades. The 4,431 broken internal links surfaced on the NBFC audit operated as one input into the 8 percent ChatGPT mention rate finding.

Is content decay worse on YMYL topics than on non-YMYL?

Materially worse on YMYL. Healthcare, financial advice, legal interpretation, statutory compliance. The retrieval models apply stricter recency and authority filtering on YMYL queries. A 2022 lending product page that does not reflect current interest rate ranges loses citation probability on retrieval models faster than an equivalent 2022 page on a non-YMYL topic. The refresh cadence on YMYL content should default to quarterly.

How long does an entity record update take to show in retrieval citations?

Wikidata updates propagate to retrieval models within days to weeks. Crunchbase updates propagate slower. LinkedIn company page updates appear in retrieval citations on a faster timescale because the page is recrawled frequently. The compound effect of synchronising all three reference sources is typically observable in mention rate measurement within one to two months on entity-sensitive query cohorts.

What is the realistic editorial capacity required to sustain a retrieval-era refresh calendar?

For a 200 commercial-page site with quarterly refresh on the top 30, semi-annual on the next 60 and annual on the rest, the editorial capacity required is roughly 0.5 to 1.0 full-time-equivalent depending on the depth of refresh. The cadence is achievable inside a normal content team budget when the brief library, schema validators and entity sync are all engineered to support the cadence rather than fight it.

Commission the Decay Audit

A brand that has never measured the schema-prose drift, entity record staleness, internal link rot and topical recency decay across its commercial pages is carrying an invisible tax on AI mention rate. The decay audit is the first artefact to commission. Request the decay audit.