Internal linking: real evidence for Google, none yet for AI citations

A 23-million-link dataset, and what it found

Cyrus Shepard published the largest public study of internal linking we could find, first in July 2022 and updated on 23 February 2026 as new data came in. The study, run through Shepard's agency Zyppy, analyzed 23 million internal links across 1,800 websites and roughly 520,000 individual URLs, matched against Google Search Console click data. Pages sitting in the 0 to 4 internal link range averaged 2 clicks. Pages in the 40 to 44 range averaged 8, close to a fourfold difference. The trend reversed past roughly 45 to 50 links, which the study attributes to sitewide navigation and footer links diluting the count rather than genuine contextual linking losing its effect.

The study is explicit about its own limits, describing itself as mostly a correlation study and recommending its conclusions be read as directionally useful rather than scientific proof. That caveat is honest, and it does not erase the finding. A page with almost no internal links pointing to it is a page few people, human or crawler, are likely to stumble onto, and 520,000 URLs is a large enough sample that the direction of the relationship is hard to write off as noise.

What Google says about it, directly

Google's own link best practices documentation, last updated 10 December 2025, states the requirement without hedging: every page you care about should have a link from at least one other page on your site. The same page explains the mechanism in plain terms. Google can only crawl a link if it is an <a> HTML element carrying an href attribute, and internal links act as a signal Google uses both to find new pages to crawl and to judge how those pages relate to each other.

On how many links is enough, Google declines to give a number. Its guidance reads: there's no magical ideal number of links a given page should contain, however, if you think it's too much, then it probably is. That is a deliberately unhelpful target to plan against on its own, but it lines up with what the Zyppy data shows independently. The useful range sits well above a handful of links and well below the kind of sitewide template that links to everything from everywhere with the same boilerplate anchor.

Anchor text moved the traffic more than the link count did

The more interesting number in Shepard's dataset sits inside the anchor text, not the raw link count. Pages carrying at least one exact-match anchor to their target keyword saw at least five times more traffic than pages with no exact match anywhere pointing to them. Naked URL anchors, where the visible link text is just the raw address, correlated with 50% more traffic than pages without one. Only 6% of the 23 million links carried no anchor text at all, meaning empty anchors were already rare in practice, and the study found no meaningful traffic difference where they did appear.

Google's own guidance backs the mechanism, not just the correlation. Good anchor text, its documentation says, is descriptive, reasonably concise, and relevant to the page that it's on and to the page it links to, and it specifically warns against cramming every related keyword into it. The two sources agree on the same practical point from different directions. A handful of well-placed, descriptively worded links does more than a template that links to everything from everywhere with the same generic anchor.

Where AI crawlers fit in, and where the evidence runs out

Whether any of this transfers to being cited by ChatGPT, Claude, Gemini or Perplexity is a different question, and it's the one with the least evidence behind it. The mechanism is straightforward. Our own guide to the AI crawlers covers how GPTBot, ClaudeBot, PerplexityBot and the rest are conventional HTTP crawlers that request URLs the same way Googlebot does, and our guide to whether AI crawlers render JavaScript shows they follow static href links and do not execute client-side navigation. A page with no path leading to it from anywhere else on the site has no more reason to be found by these crawlers than by Googlebot.

What we could not find is a study measuring the AI-citation side of that logic the way it has already been measured for backlinks. Our guide to whether backlinks work on AI search covers two large studies that disagree on how strongly link authority correlates with AI mentions, at 0.65 in one and 0.27 in the other, and neither study broke the number out by internal versus external links. Nobody has yet published the internal-linking equivalent, a dataset connecting a site's internal link structure or crawl depth to how often its pages get cited in an AI answer. Until someone does, the honest position is that the classic SEO case for internal linking is settled and the AI citation case is inferred, not measured.

What this means for the next hour of work

Start with the pages nothing links to. An orphan page, one with zero internal links pointing at it, is invisible to both systems for the same reason: nothing on the site tells a crawler, human or AI, that it exists. Closing that gap is usually a matter of adding two or three contextual links from pages that already rank or already get traffic, pointed at the page that has none.

After that, look at anchor text before adding more links. A page linked to ten times with the same generic "click here" or exact page title does less, on the Zyppy numbers, than a page linked to five times with descriptive, varied anchors that actually name what the destination covers. Do not chase the roughly 40-link mark as a literal target either. It is a correlation observed across 1,800 sites of unknown size and structure, not a number to hit on every page regardless of how much genuinely relevant content exists to link to. Add a link because it helps a reader find the next real thing to read, and the count tends to land somewhere reasonable on its own. An On-Page Optimization order specs exactly this, two to four contextual links in per page with the anchor written out, and its Deep tier adds a full internal link plan across up to 30 pages when the structural work is more than a page or two can absorb.

What we could not verify

Shepard's study is a correlation study built from one company's own client and tool data, not an independently replicated experiment, and it says so itself. We could not find a second, unrelated dataset that reproduces the roughly fourfold click difference at the 40-link mark, so treat the exact multiplier as directionally right rather than a number that will reproduce precisely on any given site. Google's own guidance on link volume is intentionally vague, offering a feeling rather than a number, so there is no official threshold to check a page against. And the entire AI-crawler section above is a mechanism argument built from how these crawlers are documented to behave, not a measured citation outcome. We looked for a study connecting internal link structure to AI citation rate specifically and did not find one, which we are flagging as a real gap rather than filling with an inference dressed up as data.

Sources

The internal link and anchor text findings come from Cyrus Shepard, 23 Million Internal Links, an SEO Case Study, Zyppy, published 12 July 2022 and updated 23 February 2026, covering 23 million internal links across 1,800 websites and roughly 520,000 URLs. Google's own guidance comes from SEO Link Best Practices for Google, Google Search Central documentation, last updated 10 December 2025. The backlink correlation figures cited above come from our own guide to backlinks and AI search, and the AI crawler mechanics come from our guides to the AI crawlers and whether they render JavaScript.

Common questions