Short answer: an internal linking experiment for publishers should test graph and navigation outcomes, not a topical-authority score. Primary metrics are orphan reduction, wrong-target reduction, contextual relevance and pathway completeness. Google recommends crawlable HTML links and contextual anchor text, but does not publish a universal topical authority score.

Hypothesis

For evergreen articles and near-orphan explainers, adding contextual links based on editorial task will reduce orphan rate and improve pathway completeness compared to a comparable cohort.

Population

Choose similar pages such as age, section, content type and traffic band. Do not put the most important evergreen pages only in the treatment.

Baselines

Save:

  • source URL;
  • target URL;
  • link type;
  • anchor;
  • section/topic;
  • lifecycle;
  • canonical;
  • status;
  • incoming contextual links;
  • related-module links;
  • publish/update data.

The intervention group

Apply bounded rules:

  1. add contextual links to relevant owners and explainers;
  2. eliminate wrong-target/redirect debt;
  3. keep correction/retraction state;
  4. do not change the headline, body copy and global navigation at the same time.

The control group

Keep pages comparable to the existing graph, unless they have bad material links that need to be fixed immediately.

Change log

For each link save source, old/new target, anchor, reason, owner and timestamp.

Observation window

Graph metrics can be evaluated immediately. Behavioral metrics need comparable windows. Search outcomes are tracked separately.

Metric 1: orphan rate

Eligible pages with no inbound contextual links out of total eligible pages.

Metric 2: near-orphan rate

Pages with a single incoming contextual link or just global navigation.

Metric 3: wrong-target rate

Links to wrong version, wrong locale, retired/retracted target or canonical duplicate.

Link that pass through the redirect even though the final URL is known.

Metric 5: contextual relevance

On a sample, classify supports task',adjacent but useful', redundant',unrelated'.

Metric 6: pathway completeness

Define tasks such as breaking story → explainer → background or review → buying guide.

Confounders

  • breaking-news demand;
  • homepage placement;
  • newsletter/social traffic;
  • content rewrites;
  • external backlinks;
  • recommendation-engine updates;
  • Search changes;
  • major news events.

Stop criteria

Stop if:

  • the control receives the same template links;
  • section taxonomy changes materially;
  • a major news event affects only one cohort;
  • content lifecycle changes massively;
  • sample size becomes insufficient.

How do you deal with breaking news

Define the grace period for new pages. Do not consider an article orphan immediately after publication if the editorial workflow links it later.

How do you handle corrections

A wrong-target to retired version must be fixed even if it appears in control. Mark contamination and maintain ethics/editorial integrity before experiment purity.

Separate module links from body links. If modules change globally, note intervention contamination.

How do you deal with syndication

Keep canonical owner and don't add alternate links between duplicate copies as if they were separate support pages.

How do you treat article age

Old evergreen pages and new breaking pages have different linking dynamics. Match cohorts by age/lifecycle.

How do you treat anchors

Does not standardize exact-match. The anchor must naturally describe the next step.

How do you deal with lift traffic?

If traffic increases, report association and confounders. Do not use it as primary proof of linking intervention.

How do you treat AI citations

Citation changes are external outcomes and do not prove that internal linking caused source selection.

Negative control

It includes already well-connected pages where the intervention layer should not make big changes.

Blind rating

For contextual relevance, hide treatment/control on a sample and apply the same rubric.

How do you interpret a positive result?

If orphans and wrong targets decrease, and pathways improve, the intervention layer has direct value.

How do you interpret null result

Even without a ranking shift, the graph improvement remains a valid outcome.

Replication

Repeat in another section/topic family before editorial standardization.

Acceptance criteria

The experiment is valid when:

  1. the hypothesis is predefined;
  2. the cohorts are comparable;
  3. graph baseline is versioned;
  4. intervention layer is delimited;
  5. change log is complete;
  6. the denominators are explained;
  7. observation window is fixed;
  8. stop criteria exists;
  9. confounders are logged;
  10. external outcomes are separate.

How do you treat homepage and section-front exposure

An article can receive traffic and temporary links from the homepage without the intervention layer having changed the contextual graph. Mark front page modules separately and exclude them from the primary contextual-link metric if the purpose of the test is body linking.

How do you treat recirculation widgets

Recommendation engines can automatically modify targets based on engagement. Save rule/version and separate algorithmically generated links from editorial edits. If the engine changes during the experiment, it marks contamination.

How do you handle story updates and URL stability

Some publisher workflows create a new URL for the update, others update the existing page. Cohort matching must take into account the editorial model, otherwise redirect and canonical behavior can dominate the result.

How do you deal with paywalled explainers

A link may be editorially relevant, but the user path differs for subscribers and anonymous visitors. Keep access state when measuring path completion and don't confuse gating with wrong-target.

Claim ledger

  • FACT/EVIDENCE: Google recommends crawlable HTML links and contextual anchor text.
  • PRACTITIONER GUIDANCE: publisher linking experiments must measure graph health and editorial pathways.
  • INFERENCE: reproducible contextual links can reduce orphaning and wrong-target debt.
  • NOT PROVEN: a universal topical-authority score or direct effect on AI ranking/citations.

Conclusion

The good internal linking experiment for publishers measures the graph you can control. If important pages become better connected and wrong targets decrease, you have direct evidence. Traffic and AI visibility can be tracked, but remain external outcomes.

Sources reviewed

Related insights