Short answer: internal linking can be tested by interventions that change the graph in a measurable way. The most useful experiments in eCommerce check orphans, wrong-target/lifecycle guardrails, market/locale routing, related-module relevance and redirect debt after migration. Google recommends crawlable links and contextual anchor text, but does not publish a topical-authority score.

Common prerequisites

Save graph snapshot before, define cohort and separate body links from global navigation and automated modules. Don't change copy, taxonomy and navigation at the same time if you want to isolate the effect.

Test 1: orphan discovery

Hypothesis

Adding contextual links to priority near-orphan pages will improve path completion without increasing redundancy.

Control

A category comparable to the existing graph and without critical defects.

Intervention

Add links only from pages where the next step is natural.

Outcome

Orphan rate, near-orphan rate and path completion.

Stop criteria

Stops if links become repetitive or the control receives the same global rule.

Test 2: lifecycle guardrails

Hypothesis

An `active/retired' guardrail in related modules will reduce referrals to retired products.

Control

Comparable modules without the new guardrail, if there are no P0 defects.

Intervention

Filter retired products and keep a log of the reason for the exclusion.

Outcome

Lifecycle-error rate and module relevance.

Stop criteria

If the guardrail hides temporarily out-of-stock products that should be kept, stop and redefine the lifecycle policy.

Test 3: market and local routing

Hypothesis

Adding market/locale context to recommendation logic will reduce wrong-region targets.

Control

A region with stable routing.

Intervention

Run target selection with market and language constraints.

Outcome

Wrong-region rate, wrong-currency paths and user-path completion.

Confounders

Campaigns and regional inventory can change click behavior.

Hypothesis

A module based on product family + buyer task will produce more relevant links than one based only on lexical similarity.

Control

Existing mode.

Intervention

Add category, family, lifecycle and task constraints.

Outcome

Human relevance rubric, redundant-link rate and wrong-product rate.

Stop criteria

It stops if the coverage drops so much that the modules become almost empty.

Test 5: redirect debt after migration

Hypothesis

Updating internal links to final destinations will reduce redirect-link rates and simplify the graph.

Control

A subset without migration or with already clean links.

Intervention

Replace targets that go through redirects with the final verified URL.

Outcome

Redirect-link rate, broken-link rate and path latency.

Stop criteria

Stop if the redirect represents an intentional transition and the final target changes meaning.

Observation window

Graph metrics can be evaluated immediately after release. Click behavior needs comparable volume and windows. Search and AI outcomes should be tracked separately.

Common confounders

  • catalog changes;
  • pricing;
  • inventory;
  • promotions;
  • navigation redesign;
  • product launches;
  • Search updates;
  • AI platform changes.

How do you choose the cohorts

Stratify by category size, traffic and product maturity. Don't put best sellers only in treatment.

How to avoid contamination

Template rules can affect both cohorts. Keep the release log and checks that global modules have not changed control.

How do you treat the positive result

Promote the rule only if primary graph metrics improve and editorial review confirms relevance.

How do you handle the null result

Keep it. Don't add more links just for effect.

How do you treat Search lift

Report association and confounders. The ranking is not the primary proof of the experiment.

Acceptance criteria

Each test is valid when:

  1. the hypothesis is predefined;
  2. the control is comparable;
  3. intervention layer is delimited;
  4. graph snapshot is preserved;
  5. the denominators are explained;
  6. stop criteria are defined;
  7. confounders are logged;
  8. rollback is possible;
  9. raw evidence is re-auditable;
  10. external outcomes are separate.

Additional contamination test

Before analysis, check if navigation, merchandising or template updates have changed both cohorts. If a global rule engine has added links to the control as well, it marks the experiment as contaminated and does not assign the difference to the original intervention layer.

How do you treat low volume pages

Graph metrics remain valid even when click data is insufficient. Keep insufficient behavioral data' separate fromno effect'. Don't remove niche pages from analysis just because they don't produce volumes comparable to best sellers.

How do you deal with temporarily unavailable products

Out of stock' is not synonymous withretired'. Lifecycle policy must say when a page and its links remain useful. An overly aggressive guardrail can remove soon-to-be-returned products and damage the user path.

How do you handle canonical after migration

A link to the final URL is only useful if canonical and page identity match. After merges, check that the destination keeps the same product or category, not just a 200 status.

Replication and promotion

Repeat tests on categories with different structure before standardizing global rules. Only promote interventions that reduce errors in multiple contexts and do not disproportionately increase the cost of editorial review.

Reproducibility note

Keep exactly the cohorts, graph snapshot and rule versions used. If the recommendation algorithm or taxonomy changes, mark a new version of the test before comparison.

Claim ledger

  • FACT/EVIDENCE: Google recommends crawlable HTML links and contextual anchor text.
  • PRACTITIONER GUIDANCE: internal-link experiments must measure graph health and task paths.
  • INFERENCE: lifecycle and market guardrails can reduce wrong-target errors.
  • NOT PROVEN: a universal topical-authority score or direct effect on AI citations.

Conclusion

The five tests separate verifiable interventions from authority stories. In eCommerce, a good graph is seen in correct destinations, respected lifecycle and useful paths. Search and AI can be noticed afterwards, not used to mask a broken graph.

Sources reviewed

Related insights