Skip to content
Faisal Mahmud TusherDigital marketing, SEO & analytics
with Faisal Tusher

Experiment Library

Internal Linking Test

An experiment design for testing whether added internal links lift target pages: random page groups, organic clicks as the metric, source-page guardrails.
Experimentadvanced7 min readUpdated

Hypothesis

If contextual internal links are added to a random half of a set of under-linked target pages, those pages will gain organic impressions and clicks relative to the control half, because links help discovery and signal each page's topic and importance. The decision is whether to invest in adding contextual internal links to under-linked pages at scale, or not.

The question is an investment question. Adding links by hand takes editorial time across hundreds of source pages, and the internal linking guide explains how to find under-linked pages and write anchors that describe the target; what it cannot say is how much traffic that work returns on a particular site. The test answers that. Write down the selection rule for target pages, the linking rule and the reading in the experiment hypothesis template before any link is added.

Metrics and guardrails

The primary metric is organic clicks per target page, read as the change in the variant group relative to the change in the control group over the same weeks. Read impressions alongside clicks: under-linked pages often have little traffic, and impressions respond first and in larger numbers when a page is discovered and ranked for more queries. Search Console's performance report gives both by page, and the Google Search Console guide explains how to export page-level data for two lists of pages over matching periods.

Two guardrails protect the rest of the site. Organic clicks to the linking source pages check that the added links did not damage the pages they were placed in, through cluttered copy, a changed topic focus or a link that pulls readers away before they act. The indexing status of target pages checks the mechanism: pages that were not indexed before the test and are indexed after it show that discovery, not just ranking, was the constraint. Google's link best practices explain what makes a link crawlable and why the Anchor Text should describe the target, so check that every added link meets them before launch.

Design

Choose the target pages by a written rule: pages that should rank and have fewer inbound internal links than pages that matter less, taken from a crawl and from the links report in Search Console. Exclude pages being rewritten, redirected or merged. Split the list at random into variant and control groups, within traffic bands if traffic is uneven, and run a pre-period to confirm the groups move together. Then add contextual links to the variant targets only, from source pages whose content discusses each target's topic, following one linking rule for every variant page.

Design elementChoiceReason
Unit of randomisationTarget pageSearch traffic cannot be randomised by visitor
GroupsVariant targets receive new links; control targets do notIsolates the added links
Source pagesRelevant pages only, chosen by topic, one rule for every targetKeeps the treatment consistent and natural
Held constantNavigation, breadcrumbs, related modules, content of targetsOnly the contextual links differ
Primary metricChange in clicks per target, variant relative to control, with impressionsLow-traffic pages respond in impressions first
GuardrailsSource-page clicks, indexing status of targetsProtects source pages, checks the mechanism

Topic structure shapes which source pages are relevant. On a site organised into clusters, as the topic clusters guide describes, the natural sources are the hub and sibling pages in the target's own cluster; on a site without clusters, the site architecture guide helps identify which sections share a topic. Links from sections with no topical connection test a different and weaker treatment. The academy checkpoint on measuring and testing SEO changes tests designing a linking test with source-page guardrails beside title tests, comparison pages and stating a mixed result honestly.

Sample size reasoning

The sample is the number of comparable target pages, and under-linked pages often have little traffic, so read impressions as well as clicks and expect only a large effect to be visible. Each page contributes a noisy weekly series, and the fewer pages there are, the larger a change must be to stand out from the wander between groups seen in the pre-period. The incrementality testing guide explains the few-units problem and how to set the smallest change worth detecting before the test. The SEO traffic estimator helps translate a plausible ranking gain into clicks, which is a reasonable way to decide whether the smallest worthwhile change is large enough to see. If the list is short, pool target pages from several sections under the same rule rather than running an underpowered test on one.

Duration

Run a pre-period, then several weeks after the linking pages have been recrawled; avoid site-wide navigation changes during the test. The treatment reaches search engines only when the source pages are recrawled, which happens at different speeds across a site, so start the post-period once most source pages show the new links in the index. Large sites with many low-value pages may find that recrawling takes longer, which is one way Crawl Budget shows up in practice. Fix the end date before launch.

Reading the results and the decision rule

Read once at the planned end. Compute each group's change in clicks and impressions per page from pre-period to post-period, the difference between those changes and an interval built from the pre-period wander. Then check source-page clicks and target indexing status. The rule, written before launch:

ResultDecision
Variant targets gain relative to control and source pages holdScale the linking programme across the site under the same rule
Impressions rise but clicks do not reliablyScale cautiously; pair linking with title and content work on the targets
Variant targets gain but source pages lose clicksRevisit the linking rule; place fewer, better-fitting links
No reliable differenceFocus links on fewer, more relevant targets rather than linking at scale

Pitfalls

Links from irrelevant sources are the first pitfall. Hitting a link quota by placing links in pages that do not discuss the target's topic tests a weaker treatment than the one the business would actually scale, and can clutter the source pages enough to cost them clicks. Hold to the relevance rule even when it means fewer links per target.

Navigation changes reaching both groups are the second. A new menu, footer or related-content module launched during the test adds links to variant and control targets alike, and the difference between the groups shrinks or disappears for reasons unrelated to the contextual links. Freeze structural link changes until the read. The third is reading low-traffic pages on clicks only. Many under-linked pages receive few clicks, and a real improvement in discovery and ranking can show up as a clear rise in impressions long before clicks separate from noise. Read both, and say in advance which one carries the decision. If links were added across a section before any test was planned, the tutorial on measuring an SEO change reads the effect afterwards against similar pages that gained no links, with the confounders checked.

Illustrative example — not a real result

A garden supplier tests contextual links

Trugspindle Garden Supply, an invented garden equipment retailer, finds 220 buying guides with three or fewer inbound links. It removes 20 scheduled for rewrite and splits the remaining 200 into two groups of 100. After a six-week pre-period in which the groups track each other, editors add four contextual links to each variant guide from relevant product and category pages. Eight weeks after recrawling, variant impressions per page are up 18% against 4% for control, and clicks per page up 11% against 3%, with an interval for the click difference of roughly 2% to 14%. Clicks to the source pages are unchanged, and 9 variant guides that were not indexed before the test are indexed after it. Trugspindle extends the linking rule to every under-linked guide.

Common mistakes

Linking tests fail when the treatment is inconsistent or the control is contaminated. The SEO title tag test uses the same page-group split on titles, and the content refresh test applies it to updating decayed articles, and the experiment library lists every design.

  • Linking without a control group. Weak pages move on their own; only a control separates the effect.
  • Linking from irrelevant pages. The test then measures a treatment nobody should scale.
  • Changing navigation during the test. Both groups receive links and the difference disappears.
  • Starting the clock at publication. Source pages must be recrawled first.
  • Reading only clicks on low-traffic pages. Impressions show the effect sooner.
  • Applying a different number of links to each page. A consistent rule makes the result repeatable.

Frequently asked questions

Why not just add links everywhere and watch traffic?

Because under-linked pages are chosen for being weak, and some of them will improve on their own while others decline for unrelated reasons. Without an unchanged control group, any movement after the linking work could be seasonality, an algorithm update or regression to the mean. The control group is what turns a before-and-after story into a measurement.

Should the control pages get no links at all?

They keep the links they already have and receive no new ones during the test. Existing navigation, breadcrumbs and related-content modules stay as they are for both groups, so the only difference is the contextual links added to the variant group.

How many links should each variant page receive?

Decide a rule before launch and apply it to every variant page, for example a set number of contextual links from the most relevant source pages, each with a descriptive anchor. A consistent rule is what lets the result be repeated; links scattered by instinct produce a result nobody can scale.

Terms used on this page

Tools to try

Frameworks

Where this page is used

Part of these learning paths

Also related

Sources

  1. Link best practices for Google — Google Search Central
  2. Performance report (Search results) — Google Search Console Help

Related