Short answer: an internal linking experiment needs to change a clear graph rule, not the entire content. Google recommends crawlable links and contextual anchor text, but does not publish a universal topical authority score. In B2B SaaS, primary outcomes can be orphan reduction, path completion and wrong-target rate; ranking and AI citations remain secondary outcomes.

Hypothesis

For comparable clusters, adding contextual links to owner pages and removing wrong targets will improve path completion and discovery without increasing redundancy.

Population

Choose two clusters similar in size and type of pages: for example, two integrations or two families of feature content. Don't compare docs to general blog without layering.

Baselines

Save:

  • URL;
  • types;
  • owner;
  • incoming/outgoing contextual links;
  • click depth;
  • wrong targets;
  • redirect links;
  • path completion;
  • Search observations;
  • AI source observations.

The intervention group

It applies a precise rule: articles and overview pages are sent to the actual owner and the relevant next step. Update links through redirects to the final destination.

Do not rewrite all pages at once.

The control group

Keep the existing graph temporarily if it does not contain critical defects. 404, security issues or wrong-product links must be fixed immediately and the control marked as contaminated.

Intervention log

For each link added or removed, it saves the source URL, target, anchor, reason and date.

Metric 1: orphan/near-orphan rate

Measure eligible pages without incoming contextual links or with insufficient paths.

Metric 2: path completion

Define tasks: explainer → feature → docs; comparison → product detail; integration overview → setup.

It measures whether the route can be traveled without a dead end.

Metric 3: wrong-target rate

Includes 404, noindex, wrong language, wrong product and deprecated targets.

Use rubric: relevant, redundant, misleading. Don't just use count.

Metric 5: click behavior

If there is analytics, compare clicks on next-step links. A higher CTR is not automatically better if the user subsequently abandons.

Observation window

Graph metrics can be evaluated immediately. User behavior needs volume. Search/AI requires longer windows.

Set the period before.

Confounders

  • content rewrites;
  • product launches;
  • document migration;
  • navigation changes;
  • backlinks;
  • campaigns;
  • Search updates;
  • AI model changes.

Stop criteria

Stop if:

  • the control receives the same global rule;
  • CMS changes automatic modules;
  • the clusters are restructured;
  • the population changes materially;
  • a critical defect requires immediate intervention.

How to avoid contamination

A template rule can modify both cohorts. It monitors the deployment log and separates manual links from automated modules.

How do you interpret the positive result

If path completion increases and wrong-target rate decreases more in treatment, the rule is operationally useful.

Search lift can only be reported with confounders.

How do you interpret the null result

If graph quality improves but Search doesn't, don't automatically consider it a failure. User and maintenance can benefit directly.

How do you interpret the negative result

If the links increase redundancy or distract the user, rollback to the previous manifest and keep the finding in the playbook.

Acceptance criteria

The experiment is valid when:

  1. the hypothesis is predefined;
  2. the clusters are comparable;
  3. intervention layer is delimited;
  4. the link manifest is saved;
  5. the denominators are clear;
  6. observation window is fixed;
  7. stop criteria are respected;
  8. confounders are logged;
  9. raw graph data is available;
  10. external outcomes are separate.

Don't confuse them with contextual links. Header and footer can make a URL easy to discover, but they do not demonstrate the semantic relationship between two topics. In the experiment, keep global, module and body-contextual links separately.

How do you handle related-content modules

If an automatic module is active, it freezes the rule or logs changes. Otherwise, both cohorts can receive new links independently of the intervention layer. For a subset, manually check relevance, product, language and lifecycle status.

How do you treat click depth

Don't use a universal three-click rule. Define priority pages and user task. A rare troubleshooting page can be more in-depth; a setup page for an important feature may require direct routing.

How do you handle reinforcements during the test

If two pages are merged, the link manifest and denominator change. Mark the event and exclude the direct comparison for that pair or version the cohort. Do not hide intervention in a redirect.

Rework metric

It measures how many links need to be patched again after a release. A rule that looks good editorially, but produces a lot of rework with each product change, may require automation or a different owner.

Expansion criterion

Extend the rule to other clusters only if the path completion improves, the wrong-target rate does not increase and the editorial review confirms that the links remain natural.

How do you keep the rollback

Save the links manifest before intervention. If path completion drops or redundant links appear, you can return exactly to the previous graph without rewriting other unrelated editorial changes.

Replication criterion

After a positive result, repeat the rule in a different cluster with the same rubric. A single cluster is not enough to declare the universal method.

Claim ledger

  • FACT/EVIDENCE: Google recommends crawlable HTML links and contextual anchor text.
  • PRACTITIONER GUIDANCE: internal-link experiments must measure graph and task paths.
  • INFERENCE: better contextual links can help discovery and navigation.
  • NOT PROVEN: a universal topical-authority score or direct effect on AI citations.

Conclusion

Editorial A/B for internal linking is useful if you're testing a rule, not an impression. In B2B SaaS, path completion and wrong-target rate provide more direct evidence than an abstract authority score.

Sources reviewed

Related insights