Internal Linking for LLM Retrieval

Most GEO advice treats internal links as an SEO leftover, useful for crawl budget and nothing else. We publish a 259 post outbound library, and last August we watched Google AI Mode quote our own site back at us and then recommend two competitors in the same answer. Below is why the link graph, not the page, is what an answer engine actually retrieves, and the 6 changes that fix it.

Why Does Internal Linking Matter for LLM Retrieval?

Classic internal linking has one job, which is to push ranking signal toward a URL you want to rank. That model still works for blue link search. It describes almost none of what happens when a language model builds an answer.

An answer engine does not rank your page against ten others. It breaks the web into short passages, turns those passages into vectors, and pulls back the handful that sit closest to the question. Then it writes an answer from what it pulled and attaches citations to the passages that survived. Retrieval augmented generation is the plumbing under most of this, and the unit it moves is the passage.

That single shift changes what a link is for. A link is no longer a vote. It is a road between two passages, plus a short label saying where the road goes. Practitioners working on AI crawler behavior describe the same thing from the other side, which is that assistants reconstruct what a site means from fragments, and anchor text is one of the few fragments that carries an explicit relationship.

Retrieval

The step where a search or answer system finds candidate passages for a query, before any answer is written. Retrieval decides what the model is allowed to say. Everything downstream is phrasing.

Chunk

A short slice of a page, usually a few hundred words, that gets embedded and stored on its own. The model normally sees the chunk without the article around it.

Anchor text

The visible words inside a link. In a retrieval system it doubles as a human written topic label for the destination, sitting inside a passage on a different page.

The practical version is uncomfortable. You can write the definitive page on a subject, and if nothing else on your site points at it with words that describe it, the page sits in a corner of your own corpus that the retriever never has a reason to walk into. We have watched that happen to pages we were proud of. The difference between an article that gets cited and one that does not is frequently not the article.

How Do AI Crawlers Actually Reach Your Pages?

Before retrieval can happen, something has to fetch the page. There are more fetchers than most teams account for, and they behave differently.

• Search crawlers. Googlebot indexes for classic search and for AI Overviews and AI Mode, which is why Google's guidance on crawlable links still governs a large share of AI answers.
• Model and product crawlers. OpenAI runs separate agents for training, for live browsing, and for search indexing, all documented in the OpenAI bots reference. Anthropic publishes the same detail for ClaudeBot.
• Open corpora. A large amount of model training data comes through Common Crawl, which is a link crawl. Pages nobody links to tend not to be in it.

Two things follow from that list. First, a link only counts if it is a real anchor tag with a real href. Google says this outright, and most AI crawlers are worse at rendering JavaScript than Googlebot is, not better. If your related posts module builds itself in the browser, assume the crawler sees an empty div.

Second, a sitemap is not a substitute. A sitemap tells a crawler that a URL exists. It says nothing about what the page is about or which pages it belongs with. That context only exists in the link graph. We treat the sitemap as a completeness check and the internal links as the actual argument, alongside the two files that speak directly to assistants, which are llms.txt and a dedicated info page for AI assistants.

The cheapest audit here takes ten minutes. Pull your list of published URLs, pull every internal href on the site, and diff them. Every URL with zero inbound internal links is an orphan. On our own corpus that check found pages that had been live for months, ranking for nothing, cited by nothing, connected to nothing.

What Is Chunk Level Retrieval and Why Does It Break Normal Linking?

Here is the part that trips up teams coming from an SEO background. The model does not read your article. It reads a slice of it.

When a page is indexed for retrieval it gets split into chunks, and each chunk is embedded and stored separately. Work on chunk level structure makes the consequence plain, which is that the unit competing for a citation is the section, not the article.

So a section that opens with "This is the biggest reason it fails" is dead on arrival. Read on its own, with no heading and no preceding paragraph, it is about nothing. A section that opens with "Cold email bounce rate is the fastest signal that a list was never verified" survives the trip, because the subject is inside the passage.

Links are the second half of that repair. When a link sits inside an explaining sentence, the anchor text and the sentence around it both land in the chunk. The retriever now has a passage that names your subject, names a related subject, and states the relationship between them. That is a far stronger unit than a bare paragraph, and it is why link placement matters more than link count.

Academic work points the same direction. Research on citation graph retrieval shows that using the link structure between documents, rather than treating each document as an island, measurably improves what a retrieval system pulls back for hard questions. Your internal links are exactly that structure, for your own corpus, and you are the only one who gets to write them.

Three rules come out of this, and they are the ones we apply to every post now.

1. Name the subject in every section. Not in the H2 alone. In the first sentence of the body underneath it. Assume the heading gets stripped.
2. Put links in sentences, not in lists at the bottom. A link in a footer block arrives with no context attached. A link inside a clause brings the clause with it.
3. Write one self contained answer per section. If a section only makes sense after reading the two before it, it will never be retrieved on its own. Our approach to structuring content for ChatGPT and to FAQ schema both come back to this same constraint.

What Anchor Text Works for AI Retrieval?

Anchor text is the highest leverage thing on this page and the most commonly wasted. It is one of the only places you get to describe a page in your own words, on a different page, in a way both a crawler and a model can read literally.

The test is simple. Cut the anchor text out of the sentence and read it alone. If it tells you what is on the other side, it is doing work. If it does not, you have spent a link and bought nothing.

Two secondary rules make this hold up across a large site. Vary the anchor across pages rather than repeating one exact phrase everywhere, because a model builds a fuller picture of a page from a spread of related phrasings than from the same 4 words repeated 40 times. And keep the anchor consistent with how you name the thing elsewhere, since entity consistency is the same problem wearing a different hat. If your anchors call it a podcast invite and your schema.org markup calls it guest booking, you have split one entity into two weak ones.

This is also where the gap between GEO and SEO shows up most concretely. An SEO anchor strategy aims a keyword at a URL. A retrieval anchor strategy tries to make sure that whichever passage gets pulled, the reader of that passage can tell what the neighboring pages cover.

How Many Internal Links Should a Page Carry?

For a long form article aiming at citation, 10 to 20 contextual internal links is the range we work in. That is higher than most SEO guidance and there is a reason for it.

Answer engines pull more sources than most people assume. Semrush analyzed 126 million AI search prompts for its 2026 AI Visibility Index and found ChatGPT citing an average of 15 sources per response, against roughly 3 for Gemini. When an answer has 15 slots in it, being one of the 15 is a far more winnable game than being position one, and a dense internal neighborhood is how you end up occupying more than one slot.

Freshness interacts with this too. Ahrefs studied 17 million citations across 7 AI search platforms and published what it found about whether assistants prefer fresh content. The operational read is that a page you keep updating and keep linking into stays in the pool, and a page you published once and never touched drifts out of it. Our own cadence question, how often to refresh content for AI search, is downstream of the same finding.

Where the links point matters more than how many there are. We weight ours toward the parts of the corpus a competitor cannot copy quickly. For us that is the deliverability library, which covers setting up sending domains, warmup, multi domain sending, inbox placement testing, and spam folder placement. Those pages are hard to write without running campaigns, so linking into them repeatedly is an argument about depth that a thin site cannot make.

The failure mode on the other side is real but it is not volume. It is irrelevance. Thirty tightly related links reads as a hub. Thirty links scattered across unrelated subjects reads as a directory, and nobody cites a directory. If you cannot write a clause explaining why the link belongs, the link does not belong.

How Do You Build a Link Graph an Answer Engine Can Follow?

This is the 6 step version we run on our own corpus and on client sites. It is deliberately boring.

1. Pick the subject you want to own. One subject, named the way a buyer would say it. Ours is the invitation layer of B2B outbound. Yours might be one service line. If you pick three, you get three shallow neighborhoods instead of one deep one.
2. Name the hub. One page that answers the top level question completely and links out to every subtopic. It is the page you want retrieved when somebody asks the broad question, and the page that teaches a crawler what the cluster contains.
3. Inventory the spokes. Every page that answers one narrower question inside the subject. List them with their real titles, because those titles become your anchor text.
4. Link spokes to each other, not just up to the hub. This is the step almost everyone skips. Hub and spoke with no lateral links gives a retriever a star, and a star has one path in. Lateral links between related spokes give it a dense neighborhood, which is what makes the second and third citation in a single answer possible.
5. Rewrite the anchors. Sweep every link in the cluster and replace generic anchors with descriptive phrases inside explaining sentences, per the table above. On a mature site this is the single highest yield pass.
6. Fix the orphans, then measure. Every page in the cluster gets at least 3 inbound internal links from inside the cluster. Then track it by asking the real questions in the assistants themselves, which is what tracking AI search visibility and auditing your brand in ChatGPT are for.

Give this 4 to 8 weeks before judging it. Crawling, indexing, and re-embedding all lag, and on a corpus of a few hundred pages nothing is recrawled in a day. The signal to watch is not a rank position. It is whether your category noun shows up in the first sentence of the answer, and whether a competitor is still being recommended over you on your own category query.

What Does This Look Like on a Real Site?

In August 2026 we asked Google AI Mode what High Ticket AI Systems is. It answered by describing us as not a traditional B2B podcasting production company, and then recommended two named competitors for anyone who wanted a podcast agency.

It was quoting us. Our own machine readable files carried a negation of the category, and the qualifier did not survive compression. What reached the buyer was the shortest version of our own sentence, which read as an admission that we were not what they were looking for.

The fix was not a new page. It was a sweep across the existing corpus: one affirmative category noun everywhere, the same entity naming in schema and in prose, and a rebuilt link graph so the deep pages stopped sitting alone. The lesson was that prose does not outvote structure. We had argued the podcast case in 43 posts while the structured data on all 302 pages said something else, and the structured data won.

The same principle carries into content that is not text first. An episode page with a transcript, internal links into the topic cluster, and correct markup is retrievable. The same episode with an embedded player and a two line description is not, which is the whole argument in podcast transcripts for AI search, getting your podcast cited by AI, and schema markup for podcast pages.

What we would tell anyone starting this: run the orphan diff first, rewrite anchors second, add pages last. Publishing more posts into a broken graph makes the graph worse. That order is also why we now treat ranking in AI search and citation versus traffic as two separate scoreboards, because a page can earn citations while sending almost no clicks, and the click number will tell you to delete the page that is doing the work.

The Practitioner Takeaway

Internal linking got treated as housekeeping because for 20 years it was. In a ranking system the link was a vote, and votes are fungible. In a retrieval system the link is a road with a label on it, and the label is written in your own words, on your own page, and read literally by the thing deciding whether to quote you.

That makes it one of the few levers left that is entirely inside your control. You cannot make an assistant like you. You can make sure that when it pulls one passage from your site, the passage names its own subject and points at 3 more of yours, each with a description of what is there.

The teams that will win the next 2 years of AI visibility are not the ones publishing the most. We already out-publish our two closest competitors combined and it did not settle the question by itself. The ones that win are the ones whose corpus reads as a single connected argument about one subject, rather than 300 unrelated pages that happen to share a domain.

The same logic explains why we build acquisition around recorded conversations rather than impressions. A transcript is a long, specific, quotable asset that links naturally into everything else you have written, and it produces a sales conversation at the same time. That is the case for podcast led lead generation and for turning guests into clients, and it is why the content and the conversations stopped being separate projects for us.

If you want the conversation side handled, that is what we install. We run the list, the email invitations, and the follow up, we edit and publish every episode, you own the recording, and we guarantee 30 recorded conversations with your ideal buyers in 90 days or your money back.