Internal linking: one method that works for Google and for AI engines

Ask an SEO team how many internal links point at their most strategic page. Almost nobody can answer. Internal linking gets decided by feel, article after article, with a vague sense of doing the right thing.

We wanted to know how ours held up. Rather than publish a theoretical method, we mapped the internal linking of vydera.com, then cross-checked the graph against twelve months of Search Console. The results are not flattering, which is exactly what makes them useful.

What we measured, and how

A crawl of the 84 French pages listed in the sitemap, extracting internal <a> links. One methodological precaution, and it changes everything:

We only count content links. Navigation and footer appear on every page of the site: including them levels everyone out and erases the signal. Only links placed in the editorial body count, because they are the only ones carrying intent.

Raw result: 825 content links, a maximum depth of 3 clicks from the homepage, and a graph we can finally look at squarely.

Linking does not follow performance

First question, the obvious one: do we link most to the articles that perform best?

The rank correlation is +0.05. On a scale where 0 means "no relationship" and 1 means "perfectly aligned", that is noise.

The detail says more than the coefficient. Our three most internally-linked articles, 88 links between them, total 6 clicks over twelve months. Meanwhile our article on Claude prompts pulls 399 clicks on its own, and ranks fourteenth out of fifteen on internal links, with 5 pointing at it.

Put plainly: we spent a year reinforcing the pages that do not convert, and leaving alone the one carrying two thirds of our organic traffic.

This is neither bad luck nor negligence. It is the mechanical result of linking as you write: you link what comes to mind while drafting, which means foundational articles and definitions, not the ones that perform.

The pages that sell are the poorest on the site

Second question, the one that costs money: where does the link equity go?

Our two service pages, the ones describing the offer and carrying contact requests, receive a single content link each. They send out 13 and 14. They are the two most depleted pages on the site.

The six pages under "solutions" do barely better at 3 inbound links on average. The contact page, by contrast, gets 75: it is the end-of-article CTA, placed everywhere, and it works very well.

The imbalance is stark. The editorial corpus pushes toward the final conversion but skips the middle step: the pages that explain what we do and why. A prospect reading an article reaches contact without ever passing through the offer.

39 of 84 pages are editorially unreachable

Third measurement: starting from the homepage and following only content links, how many pages do you reach?

45 out of 84. The other 39 exist only through navigation. Among them, 37 of the 62 glossary entries.

Not a disaster in itself: the glossary stays accessible, indexable, and it does its job of entity recognition. But a glossary sending out 11.8 links on average while receiving 4.9 is a one-way pump. It distributes authority it never gets back.

The about page and the legal notice are outright orphans, with zero inbound content links.

The method in five rules

What we fixed, and what we now apply to every publication.

1. Link from the content, never from an automated block. A CMS-generated "related articles" module carries no intent and weighs little. A link placed inside a sentence, with anchor text describing the destination, carries one.
2. 3 to 6 outbound links per page, no more. Beyond that each link is worth less and the reader stops clicking. Our pages with 14 outbound links serve nobody.
3. Check the balance, not the count. A useful page receives at least as much as it gives. The table above can be rebuilt with one command, and it is the only indicator that exposes the pages you are draining.
4. Link what performs, not what reassures. Cross the graph against Search Console once a quarter. If the correlation sits near zero, your linking is decorative.
5. No strategic page more than 2 clicks from the homepage, counting content links only. Navigation does not count: parts of the engine ecosystem ignore it, and it says nothing about hierarchy.

None of these rules needs a paid tool. The graph is built by a script that reads the sitemap, fetches each page and extracts links from the body. About a hundred lines.

What this changes for AI engines

A generative engine does not browse a site the way a person does. It fetches pages, extracts passages, and decides which to cite. The link structure tells it what is central and what is peripheral on your site.

Linking heavily to your definitions while ignoring your offer pages sends a clear message: this site talks about concepts, not services. Which is exactly what we were saying about ourselves without meaning to.

The retrieval mechanism is covered in our article on query fan-out, and making structure legible in the one on llms.txt.

Running the measurement yourself

The approach works on any site:

1. Pull the sitemap URLs, filtered to a single locale.
2. For each page, isolate the content (the main element, or the body stripped of nav, header and footer) and extract internal links.
3. Count inbound and outbound links per page, compute depth from the homepage breadth-first.
4. Export clicks and impressions from Search Console over the same period.
5. Compare the two rankings. If the ranks disagree, your linking is not working for you.

Budget an hour for a site under 500 pages. It is the best effort-to-insight ratio in all of technical SEO, and almost nobody does it.

For the definitions: internal linking and link juice.