Internal Linking Experiment: What Changed

Internal linking is the ranking lever everyone endorses and almost nobody measures — so we measured it: one content site (ours to experiment on), ~180 published posts, no new content and no new backlinks for the test window, one intervention: a systematic internal-link pass on twenty target pages. Rankings and clicks tracked for twelve weeks against a matched control group of pages left untouched. Here's the design, the numbers, and what moved most.

The experiment design

Targets: twenty pages stuck at positions 6–15 on their primary keywords (the page-two layer — chosen precisely because they had proven relevance and lacked push), each averaging just 2.1 internal inlinks. Controls: twenty comparable pages (similar positions, traffic, age, topic spread) left alone. The intervention, per the linking guide: five to nine new contextual inlinks per target — mid-paragraph, descriptive anchors, placed in genuinely relevant older posts (crawl-identified via inlink counts and topic match) — plus fixing three orphaned pages discovered in passing. No hub pages built, no nav changes: pure contextual-link addition, ~14 hours total. Freeze discipline: no other edits to targets or controls, publishing continued elsewhere on the site (a live site can't be a lab; the control group is the honesty mechanism).

The results at twelve weeks

1. Targets: average position improved from 9.8 → 6.4 (median gain 3.1 positions); controls drifted 9.6 → 9.2 — ordinary churn. The delta is the finding.
2. Fourteen of twenty targets improved, four flat, two declined slightly (both in a subtopic where a competitor shipped major content mid-test — noted, not excluded; real-world noise cuts both ways).
3. Clicks to targets: +38% vs +4% for controls — amplified beyond position math by three targets crossing the page-one threshold, where CTR reality lives.
4. Timeline shape: nothing for ~2 weeks (recrawl lag, per the pipeline), most movement in weeks 3–7, stable after — faster than external-link lag, consistent with internal changes being discovered on the site's own crawl rhythm.
5. The orphan fixes were the outliers: the three formerly-orphaned pages posted the largest individual gains (one from position 14 to 5) — near-zero inlinks to a normal count is the steepest part of whatever curve governs this.

What we think actually moved

Three mechanisms, unbundled as far as observation allows: equity distribution (the targets went from link-starved to normally-fed — the flow model doing what it claims); anchor-context relevance (descriptive anchors from topical paragraphs are relevance statements — targets gained most on the query variants their new anchors described, a detail worth designing for); and crawl attention (recrawl frequency on targets visibly increased in the logs — more paths in, faster refresh). Which dominated? Unknowable from outside; the practical point is they arrive bundled and cheap.

The operating conclusions

Internal linking is the highest-certainty, fastest-feedback lever in the page-two playbook — ~14 hours moved twenty pages' worth of positions no external campaign could touch that cheaply, per the series' cheapest-traffic pattern. It's maintenance, not a project: the gains came from correcting accumulated neglect (every new post since links backward and receives links per the publish checklist — the audit shouldn't need repeating at this scale). And it has a ceiling: the four flat targets sat in the site's most competitive subtopic, where the gap to page one is external authority internal plumbing can't supply — internal links spend the equity you have; they don't mint it.

Frequently asked questions

How many internal links per page is optimal?

Wrong frame — the experiment's read is deficit correction: pages far below their site's normal inlink count gain most from normalisation; stacking links beyond relevance showed nothing in our data and degrades the linking pages. Link where a reader is genuinely served, audit for starvation quarterly, per the guide.

Did the pages you added links FROM lose anything?

No measurable movement — consistent with the flow model's dilution being marginal at ordinary link counts. The one legible cost was editorial: forced-feeling links would degrade the host page for readers, which is why placement stayed contextual.

Can I replicate this without a crawler and rank tracker?

Minimum kit: Search Console (find the page-two layer, watch positions) and Screaming Frog free (inlink counts, orphan discovery) — the zero-dollar version loses precision, not the effect. Then the ceiling question is the usual one, answered the usual way: external authority (our lever).