The most common internal link on the internet says "read more." It is also the least useful thing you can publish, because it describes the act of clicking rather than the thing being linked to.
Internal linking has been solved advice for fifteen years: link related pages, use descriptive anchors, do not orphan anything. All still true. What has changed is why it is true, and the new reason has different implications for how you structure a site.
Traditional search treats a link as a vote and a discovery path. Retrieval-based AI systems do something different. They break content into passages, embed those passages, and retrieve individual chunks to support individual claims. In that architecture, the useful question is not how much authority flows through a link but what the link tells the system about how two pieces of content relate.
This guide covers the architecture that follows from that, the audit process, and a real rewrite we ran on this site.
Link graph — the complete network of internal links across a site, considered as a structure rather than as individual connections. Its shape communicates topical organisation: tight, bidirectional clusters read as coherent subject expertise, while scattered links across unrelated topics read as a site with no particular specialism.
Why LLMs Read Links Differently
| Dimension | Traditional search | AI retrieval |
|---|---|---|
| Unit of interest | The page | The passage |
| What a link does | Passes authority, enables discovery | Signals topical relationship and context |
| Anchor text role | Relevance hint for the target page | Entity and concept label |
| Cluster effect | Consolidates ranking strength | Defines a topical boundary the model can perceive |
| Orphan cost | Hard to discover, weak authority | Reads as peripheral to the site's subject |
The practical consequence
Because retrieval selects passages, a page's internal links contribute to how the system understands what that page is about and what neighbourhood it sits in. A page on ecommerce category mapping that links out to product taxonomy, structured attributes and AI consideration sets is legibly part of a coherent subject. The same page linking to your holiday shipping policy and your careers page is noisier.
You are not distributing authority. You are drawing a boundary around a topic and standing inside it.
All the traditional reasons still apply. Links still aid crawling, still pass ranking signal, still help human readers. Nothing in this guide asks you to trade one for the other, because the practices that serve retrieval also serve traditional search. The difference is emphasis: anchor text and cluster coherence matter more than raw link count ever did.
The Four Architectures Compared
A pillar page linking down to specific spokes, each linking back up. Produces a clear topical boundary that both crawlers and retrieval systems can perceive.
Strict separation with no cross-linking between sections. Clean in theory, but real topics overlap and enforcing separation breaks genuinely useful connections.
Everything links to everything. Feels thorough, communicates nothing. If every page relates to every page, no topical boundary exists to be read.
Links only from navigation and archives. Common on blogs that never retro-linked. Every page is equidistant from every other, which conveys no structure at all.
Why hub-and-spoke wins
It is the only one of the four that produces a readable answer to the question what is this site an authority on. A pillar with fifteen spokes linking bidirectionally creates a dense region in the link graph with a clear centre. That density is the signal.
The refinement worth adding: allow controlled cross-linking between clusters where the relationship is genuine. Pure silos break useful connections; the fix is to link across when a reader would actually benefit, and not otherwise. The test is whether you can articulate the relationship in the anchor text. If you cannot, the link probably should not exist.
Our guide to building AI-citable content clusters covers the content side of the same structure.
Anchor Text as an Entity Signal
The highest-leverage and most neglected element in this whole subject.
The hierarchy, worst to best
| Anchor | What it tells a model | Verdict |
|---|---|---|
| "read more", "click here" | Nothing | Wasted |
| "this guide", "our post" | That a document exists | Nearly wasted |
| "learn about pricing" | A vague topic | Weak |
| "Amazon FBA fee structure" | A specific concept | Good |
| "the 2026 Amazon FBA fee structure" | A specific, dated concept | Best |
The rules
- Name the entity or concept, not the format. "Our schema markup guide" is weaker than "schema markup for ecommerce."
- Vary naturally across instances. Fifty identical anchors to one page reads as manipulation to traditional search and adds no information for retrieval. Say the same thing different ways.
- Keep it inside a sentence. An anchor embedded in prose carries surrounding context; a bare link on its own line carries none.
- Match the target's actual subject. An anchor promising one thing pointing at another damages both pages.
- Front-load the distinctive word. Anchors get truncated in some contexts and scanned quickly in all of them.
Every anchor is a label you are attaching to another page. "Read more" labels it nothing. Do that six hundred times and you have published a site with no stated subject.
Building a Cluster That Reads as Authority
Defining the pillar
The pillar answers the broadest question in the topic and links to every spoke. It should be genuinely comprehensive rather than a table of contents with paragraphs — a pillar that is only a hub with no substance of its own gets linked to and never cited.
One pillar per topic you actually want to own. Two pillars competing for the same subject split the cluster and confuse both the boundary and your own ranking.
Spoke coverage mapping
Spokes answer specific questions inside the pillar's territory. Map them by listing every question a buyer asks in this subject, then checking which have a dedicated page. The gaps are your content plan.
- Definitional — what is X.
- Comparative — X versus Y.
- Procedural — how to do X.
- Evaluative — is X worth it, what does X cost.
- Diagnostic — why is X not working.
A cluster covering all five intent types across ten to twenty spokes reads as far more authoritative than thirty pages all answering variants of one question.
The size that works
Roughly eight to twenty spokes per pillar. Below eight the cluster is too sparse to form a perceptible region in the graph. Above twenty, sub-clusters usually want to become their own pillars, and forcing them under one hub creates a structure nobody can navigate.
The long-term version of this is covered in building a compounding content moat.
Bidirectional Linking Rules
The rule that most sites get half right: pillars link down, spokes forget to link back.
The required links
- Pillar to every spoke. Non-negotiable. A spoke the pillar does not reference is not in the cluster.
- Every spoke back to the pillar. Ideally early in the body, in context, with an anchor naming the pillar's subject.
- Spoke to two or three sibling spokes, where the relationship is genuine. This creates density without becoming a mesh.
- Nothing forced. A link you cannot justify in the anchor text is noise.
Why the upward link matters most
The spoke-to-pillar link is the one that establishes hierarchy. Without it a model sees a set of related pages with no centre; with it, one page is repeatedly identified as the subject's anchor. That repeated identification, in varied anchor text, from many pages, is as close to declaring "this is what we are an authority on" as internal linking gets.
Some people avoid bidirectional internal links believing reciprocity is penalised. That concern applies to manipulative external link exchanges, not to your own site architecture. Bidirectional internal linking within a topical cluster is standard, sensible practice and nothing about it is risky.
Link Density and Placement
How many is right
There is no magic number, but useful working bounds exist. On a 2,000-word article, roughly three to eight contextual internal links is a sensible range. Below three the page is under-connected; above ten or so on a piece that length, links start competing with the content and each one carries less weight.
The better test is not count but justification: can you say, for each link, what a reader gains by following it? If not, remove it.
Where they belong
- In body prose, inside sentences. This is where surrounding context makes the relationship legible.
- Early for the pillar link. Establishing the cluster relationship near the top helps.
- At the point of genuine relevance, not clustered at the end. A link placed where the reader is thinking about that subject is doing real work.
- Not exclusively in a related-posts block. Template-generated related links are weaker signals than in-prose links because they carry no sentence context and appear identically across many pages.
The passage-level implication
Since retrieval works on chunks, a link's context is the passage it sits in rather than the whole page. That means a link inside a well-written paragraph about the linked subject is worth considerably more than the same link in a generic footer list — the passage carries the relationship, and the passage is what gets retrieved.
Auditing Your Existing Link Graph
The process, in order.
- Crawl the site and export all internal links with source URL, target URL and anchor text. Any standard crawler does this.
- Count inbound internal links per page. Sort ascending. Everything with zero is an orphan; everything with one or two is nearly one.
- Group anchors by target. For each important page, read every anchor pointing at it. This is the single most revealing view in the audit.
- Flag non-descriptive anchors — read more, click here, here, this, this guide, learn more, our post.
- Check every link resolves with a 200 status and zero redirect hops. Redirect chains waste crawl budget and are trivially fixable.
- Map clusters. For each pillar, confirm it links to all its spokes and that all spokes link back.
- Look for cross-topic noise — links between pages in genuinely unrelated clusters that serve no reader purpose.
The view that tells you most
Step three. Grouping all anchors by target page shows you, in one screen, what your site is telling models about that page. If your best guide receives forty inbound links and thirty of them say "read more," you have found a large, cheap improvement.
Want your link graph audited?
We will crawl your site, map your clusters, find your orphans, and hand you the anchor rewrite queue sorted by impact.
Book a Strategy Call →The Ecom Profit Box
Eleven playbooks on listings, conversion, images, and email. Built for operators, no fluff, no email sequence.
Grab It Free →Orphan Pages: The Silent Problem
A page with no inbound internal links is telling every system that reads your site that it is not part of what you do.
How they happen
- Published and forgotten. The most common cause by a wide margin — nobody went back to link to it.
- Site migrations that dropped contextual links while preserving URLs.
- Template dependence. Pages reachable only through pagination or a category archive.
- Redesigns that removed a navigation section.
- Seasonal content linked from a homepage module that has since rotated.
Why it matters more now
Under traditional search, an orphan with strong external links could still rank. For retrieval-based systems, isolation carries an additional cost: the page sits outside every topical cluster on your site, so it contributes nothing to your entity's subject definition and receives no contextual reinforcement from it either.
The pages most often orphaned are, frustratingly, the ones with the most impressions — older guides that ranked well, were never retro-linked, and now sit disconnected from the clusters built around them afterwards.
Two or three contextual inbound links from genuinely related pages resolves an orphan entirely. This is an afternoon of work with a permanent effect, and it is almost always the highest-return item that comes out of a link audit.
A Real Rewrite: 614 Links, 89 Posts
We ran this process on evolveamz.com. The numbers and the findings, since process detail is more useful than a claimed outcome.
| Metric | Figure |
|---|---|
| Internal links reviewed and rewritten | 614 |
| Posts touched | 89 |
| Largest problem category | Non-descriptive anchor text |
| Second largest | Links to pages that had moved or 404'd |
| Third | Orphaned older guides with high impressions |
What we actually found
The expected finding was broken links, and there were some. The dominant problem was different: anchor text that described the act of reading rather than the subject being linked to. Hundreds of instances of "read more", "this guide", "our post on this" and similar. Every one of those was a page describing another page as nothing in particular.
The second finding was structural. Older high-impression guides — the ones with the most accumulated visibility — were the least connected, because clusters had been built around them later and nobody went back to link upward and inward.
What we changed
- Every non-descriptive anchor rewritten to name the target's actual subject, varied naturally rather than repeated identically.
- Every link verified to resolve at 200 with zero redirect hops, since redirect chains had accumulated through slug changes.
- Orphans connected with two to three contextual inbound links from genuinely related pages.
- Bidirectional cluster links completed, particularly the spoke-to-pillar direction which was missing far more often than the reverse.
We are not claiming a specific traffic outcome from this work, because it ran alongside other changes and we cannot isolate it cleanly. What we can report is the process, the volume and what the audit surfaced. Treat anyone publishing a precise percentage lift from an internal linking pass with appropriate scepticism, since almost nobody runs it as a controlled test.
The Technical Checks That Matter
- Every internal link returns 200 with zero redirect hops. Slug changes accumulate chains silently. Chains waste crawl budget and dilute the signal.
- Links exist in the server-rendered HTML. Links injected only by client-side JavaScript may not be seen by every crawler, and some AI crawlers do not execute JavaScript at all.
- Use absolute URLs in content where your platform allows it, since relative links can break when content is syndicated or parsed out of context.
- Canonical consistency. Link to the canonical version. Linking to a non-canonical variant splits the signal you were trying to consolidate.
- No links to noindexed pages from within clusters. You are pointing at something you have told search engines to ignore.
- Trailing slash consistency. Mixed forms create unnecessary redirects at scale.
- Check the mobile template. Some themes hide in-content links or entire modules on mobile, which removes them from the rendered page.
The one that catches most sites
Redirect hops. A site that has changed slugs a few times accumulates internal links pointing at old URLs that redirect to current ones. Everything works for readers, so nobody notices, while every one of those links is a small unnecessary cost. Auditing for zero-hop resolution is quick and the fix is mechanical.
The wider technical picture is covered in our schema markup stack guide and the 30-signal citation audit.
Common Mistakes
- Optimising link count instead of link quality. Adding links to hit a number produces noise. Every link should be justifiable in one sentence.
- Identical anchor text repeated at scale. Fifty links all saying "Amazon PPC guide" looks manipulative and adds no incremental information after the first few.
- Relying entirely on related-posts modules. Template links lack sentence context, which is exactly what makes an in-prose link valuable for passage retrieval.
- Linking from every page to the homepage repeatedly. Navigation already does this. Contextual body links should go to specific relevant pages.
- Cross-linking unrelated clusters to spread authority. It blurs the topical boundary you were trying to draw.
- Never retro-linking. New content should link to old, and old content should be updated to link to new. Most sites only ever do the first.
- Treating it as a one-time project. The graph degrades continuously as content is added, moved and retired.
- Ignoring the pillar's own quality. A hub page that exists only to link out gets linked to and never cited, because there is no passage worth retrieving.
The Maintenance Protocol
| Cadence | Task | Time |
|---|---|---|
| Per new post | Link to pillar and 2–3 siblings; add inbound links from 2–3 existing pages | 15 minutes |
| Monthly | Check for new orphans; verify recent links resolve at 200 | 30 minutes |
| Quarterly | Full crawl; anchor text review by target; redirect hop check | 2 hours |
| Annually | Cluster structure review; split or merge pillars as topics evolve | Half a day |
| After any migration | Full re-audit before anything else | Half a day |
The rule that prevents most of the work
The inbound half is not optional. Publishing a post and linking outward from it is half the job; the other half is going back to two or three existing pages and linking in. Most sites do the first and skip the second, which is exactly how a site accumulates orphans while feeling well-linked.
Make it part of the publishing checklist rather than a separate project. Fifteen minutes per post prevents the two-hour quarterly cleanup and the eventual 614-link rewrite.
The batch exception
When several posts publish close together, hold the cross-links between them until all are live. Linking to a page that has not published yet creates a 404 that may be crawled before you fix it. Run a single retro-link pass once the batch is complete — which is precisely the approach we are using for this series.
The Short Version
- Retrieval systems work on passages rather than pages, so internal links function as topical context signals rather than primarily as authority pipes.
- Hub-and-spoke with controlled cross-linking is the right architecture. Silos are too rigid, mesh is too noisy, and flat structures communicate nothing.
- Anchor text is the highest-leverage element. Name the entity or concept, vary it naturally, and keep it inside a sentence where surrounding context reinforces it.
- The spoke-to-pillar link matters most and is missing most often. It is what establishes which page is the centre of a topic.
- Orphan pages sit outside every cluster and contribute nothing to your subject definition. Two or three contextual inbound links fixes one permanently.
- On our own site, rewriting 614 links across 89 posts found that the dominant problem was not broken links but anchors describing the act of reading rather than the subject.
- Do the inbound half. Publishing a post and linking outward is half the job; adding inbound links from existing pages is the half most sites skip.
External Sources Cited in This Article
- Google Search Central — Make your links crawlable
- Google Search Central — AI features and your website
- Perplexity — Crawler documentation
- arXiv — Generative Engine Optimization
- evolveamz.com internal link audit — first-party process data reported in section nine