Internal-link health audit: from rendered anchor to completed task

Audit charter: protect important reader transitions

Define the audit around a bounded set such as the top learning routes, recently consolidated pages, or commercially important guides. Record crawl date, rendered coverage, included templates, excluded areas, and the reader tasks the routes should support. The evidence pack combines link inventory, HTTP and canonical checks, link-purpose review, keyboard inspection, and task notes. PASS means an essential transition is exposed, understandable, healthy at the destination, and justified by the source context. REVIEW marks uncertainty or maintenance debt with an owner. STOP marks dead routes, inaccessible controls, misleading labels, or links to content that no longer owns a useful answer. Google’s crawlable-link documentation and W3C’s link-purpose guidance ground different gates; passing one cannot compensate for failing the other.

Evidence: Google Search Central; World Wide Web Consortium

Define what the crawl sample can and cannot see

Do not set an internal-link count target. The charter protects paths, not graph density, and it can approve a page whose job legitimately ends without another click.

Evidence: Google Search Central; Google Search Central

Rendered edge checks

For every sampled edge, inspect the production DOM for an anchor and href, activate it with keyboard, and record the visible text plus surrounding context. Follow the destination to its final URL and note redirects, status, canonical target, robots state, and any interstitial. PASS requires a direct or justified stable route to the intended answer. REVIEW covers a necessary legacy redirect with documented ownership. STOP covers loops, broken responses, onclick-only essential navigation, or a final target whose job has changed. Google’s link guidance supports the anchor and destination checks; a crawler export alone may miss interactive states. Reproduce at least one mobile and one desktop navigation state when templates differ.

Attach source and destination evidence to the audit row. A green status from last month cannot verify today’s rendered component.

Evidence: Google Search Central; GOV.UK Service Manual

Purpose and context checks

Read each anchor as a person encountering it out of sequence. Can the destination purpose be inferred from the text and programmatically associated context? Does repeated wording lead consistently, or do several ‘learn more’ links require guessing? Confirm that the surrounding paragraph creates a real evidence need, prerequisite, or continuation. PASS means the prediction and destination answer match. REVIEW applies when local context rescues a vague label but reuse is risky. STOP applies when keyword-heavy wording promises a different job or when identical labels lead to incompatible destinations. W3C’s guidance supplies the accessibility standard; direct reader observation can reveal confusing but technically conforming language.

Preserve the predicted job in the row before following the link. Retrospective explanations often make a weak label look clearer than it was.

Evidence: Google Search Central; World Wide Web Consortium; GOV.UK Service Manual

Graph and ownership checks

Review valid edges together. Identify important answer owners with no incoming justified path, clusters that require repeated hub backtracking, cycles that never advance a task, and components that send every page to the same destinations. Confirm who maintains each destination and which source contexts must change after consolidation. PASS means essential routes have accountable owners and no sampled page is linked merely to satisfy a score. REVIEW covers a useful page whose route depends on one fragile template. STOP covers an orphan with no distinct job, an abandoned destination, or a sitewide link that misstates relevance across most sources. A link audit should be allowed to remove edges and retire content; success is not a monotonically larger graph.

Keep canonical intent and content ownership in the same row so technical consolidation does not leave editorial links pointing at obsolete variants.

Evidence: Google Search Central; World Wide Web Consortium

Representative path checks

Choose a few realistic tasks and begin from the places readers actually enter: homepage, category, deep article, and search landing page. Record first edge, wrong turns, expectation before activation, answer completion, and whether a recovery route was needed. GOV.UK’s usability benchmarking approach supports consistent prompts and observable results. Do not coach participants through the taxonomy or report only successful sessions. PASS means the intended route supports the task even if the exact path differs from the diagram. REVIEW means completion relies on an unexpected route that may be worth legitimizing. STOP means the right destination appears only after explanation by the facilitator. Small samples cannot estimate all behavior, so preserve qualitative reasons and contradictory cases.

A quick route with an unexpected answer is a failure; a longer route with necessary prerequisites may be valid. Keep task quality separate from click depth.

Evidence: Google Search Central; Google Search Central

Disposition, repair owner, and next sample

Resolve every STOP item before closing the audited scope. Each REVIEW item needs a named owner, decision date, and trigger that will supply missing evidence. Retain the prechange inventory and rollback for template or redirect repairs. The immediate next action is to audit ten essential edges with the evidence pack and fix the first failed layer, rather than launching a sitewide link campaign. Recrawl and manually retest changed paths after deployment. Limits remain: tools differ in JavaScript coverage, platform reports lag, observed clicks cannot establish comprehension, and search engines decide how they crawl and index. Schedule the next sample after a navigation release, consolidation, accessibility issue, or destination ownership change. The audit is complete when routes and dispositions are reproducible, not when a vendor score reaches a preferred number.