SEO Content Refresh Priorities: Source Failures, Query Drift, and a lastmod Evidence Ledger

Once a content library reaches hundreds or thousands of articles, "rewrite everything once a year" both wastes editorial capacity and creates a false sense of freshness. A stable protocol tutorial may need no changes for years, while pricing, APIs, and security guidance may become outdated within weeks. The RAG Document Freshness SLO constrains the delay before internal knowledge enters the index; this article manages the facts, search demand, and publication evidence for public pages.

Make Each Claim an Auditable Asset

Generate contentEvidence.v1 for every page: content ID, release, canonical, primary intent, key claims, source URLs and access times, source type, reviewAfter, owner, and body summary. Set shorter review cycles for prices, version support, regulations, and performance figures; explanations of principles can use longer cycles. A cycle only determines when an item enters the review queue; reaching its deadline does not mean it is wrong.

Source probes store the HTTP status, final URL, content digest, and inspection time. A 404/410, a redirect to an unrelated homepage, a material title change, or an official deprecation notice all generate events; temporary 429/5xx responses enter the retry queue instead of being marked invalid immediately. For security or compliance claims, lower the page's trust status and require manual review when a source becomes invalid. Do not let a model automatically replace primary evidence with a blog that merely appears equivalent.

W3C PROV derivation relationships can connect the previous release, source snapshot, editing activity, and new release. This lets reviewers answer why a conclusion changed instead of seeing only a Git diff. The evidence-span principle from RAG Citation Provenance also applies to public articles: a source should support a specific claim, not serve as decoration at the end of the article.

Search Signals Only Determine Priority

Use fixed windows to retrieve clicks, impressions, CTR, and average position by page, query, and country/device from the Search Analytics API. Each snapshot must preserve the complete request: type, dimensions, aggregationType, date/time zone, filters, dataState, rowLimit, startRow, and property. Paginate with startRow until an empty page is returned, recording the row count for every page and the API completion watermark. The API returns top rows, and data such as anonymized queries may be missing. Even completed pagination does not represent the full set of queries, so "query drift" describes only changes in the visible set under the same request definition. Do not directly compare the last seven days with a previous month that includes holidays, and do not attribute an entire CTR decline to the snippet when position has changed substantially. For low-sample pages, report the raw values instead of using percentages to manufacture dramatic swings.

Refresh candidates combine multiple kinds of evidence: high business importance with an invalid source; a new intent appearing in target queries but not covered by the body; stable impressions while clicks consistently deviate from the baseline for similar positions; a page that is still crawled but cites an old version; or multiple pages beginning to compete for the same intent. You can define a transparent priority formula such as risk impact × source confidence + query opportunity + page coverage - editorial cost, but you must retain the component scores, raw evidence, and reasons for manual overrides.

AI Search Relevance Tuning addresses on-site retrieval, while the AI Answer Citation Visibility Experiment addresses citations by external AI systems. Signals from both can enter the queue, but they cannot replace fact-checking. A new page with no impressions may simply not have been discovered yet, so do not repeatedly rewrite it for that reason; first use Search Crawler Observability to verify its delivery and crawl paths.

A Refresh Must State What Changed

An editing task selects an action type: repair an invalid source, update a version matrix, cover a missing intent, merge duplicate pages, remove an incorrect claim, or only improve the wording. When the task is complete, generate a new release and record the changed claims, old and new sources, verification steps, and whether the user's conclusion has changed. Fixing only a typo or styling should not change the semantic lastmod; update the sitemap lastmod only when the main body, code, structured data, or page links have changed materially.

RFC 9110 defines Last-Modified and validators for HTTP cache negotiation, while sitemap lastmod tells crawlers when a page was last modified. Both can come from the same release time, but their semantics and consumers differ. Do not write the current time across the entire site on every deployment, and do not let a CDN serve an old body while returning a new timestamp. A synthetic probe should verify the body digest, HTTP validator, visible update time on the page, and sitemap value together.

When merging pages, choose one canonical to retain, redirect each old URL to it in a single hop, and update internal links. Return 404/410 for deleted pages that have no replacement. Changing a URL is not the default refresh action because it disrupts existing citations. Programmatic SEO Template Release Gates prevent similar duplicate pages from continuing to appear.

Validate with Cohorts and Allow Rollbacks

Each week, release only a group of pages that has passed manual technical review. Retain an unrefreshed cohort on the same topic and observe source errors, crawling, valid indexation, query coverage, and user completion behavior over a predefined window. Search systems and demand can change at the same time, so report correlation and uncertainty rather than crediting a one-time increase to a changed date. Quality guardrails include factual errors, broken internal links, body similarity, page performance, and user complaints.

Fault injection should cover source timeouts, a probe receiving a 200 response from a login page, missing days in Search Analytics, an incorrect canonical merge, a sitemap timestamp earlier than the release, and a CDN that continues serving the new body after a rollback. On failure, freeze automated publishing and point the content alias back to the previous verified release. A version that genuinely corrected dangerous facts must not be mechanically rolled back because traffic declined; the owner should decide and record the reason.

The output of a refresh system is not "100 articles updated today," but a reproducible queue: which evidence changed, why it received priority, which claims were modified, and how the result was validated. Only when lastmod agrees with this evidence chain does the update time become a signal rather than decoration.

Further Reading

• Programmatic SEO Template Release Gates
• Search Crawler Observability
• AI Answer Citation Visibility Experiment
• RAG Document Freshness SLO