The link prediction model: how internal links steer your AI visibility

You write a carefully crafted paragraph with an internal link to your product page or a related article. An AI system extracts that paragraph, uses it in an answer, and the link? Gone. Another time the link does survive and shows up as a clickable path to your site, right in the middle of the generated answer. That difference is not random. It follows patterns you can understand and steer.

The link prediction model is exactly that: a way to predict where links naturally belong, so your internal links survive extraction by AI systems. In this article I explain in my own words how AI judges links, and what that means for the way you link internally. This is part of our broader view on generative engine optimisation (GEO), which I work out in full in the ultimate GEO guide.

Check your own page: get a GEO score with concrete tips through our free GEO check.

Why links in AI answers matter

In classic search results, every result is a link. In AI answers, that is not the case. The synthesis often answers the question without the user clicking anything. A link that survives extraction is therefore one of the scarce moments when someone actually lands on your site.

But it is about more than clicks:

• Direct traffic: a surviving link is a direct path to your page, in an environment where that traffic has become scarce.
• Brand visibility: even without a click, your domain shown as a clickable source signals authority and relevance.
• Attribution: the link credits you as the source. Without a link, your contribution becomes invisible.
• Compounding effects: visible domains build familiarity, and link patterns can influence future selection.

Because direct traffic from AI searches is limited, every surviving link becomes disproportionately valuable.

How AI systems handle links

When an AI extracts content, it encounters links as HTML elements. The system then has to make a choice: keep the link, convert it to plain text, or do something else with it. The factors that weigh in look a lot like editorial common sense:

• Relevance to the answer: does the link fit what the AI is currently explaining?
• Natural placement: does the link sit where a link would normally sit, or does it feel forced?
• Authority of the source: if the system trusts the source, it will probably trust its links too.
• Value to the user: does keeping this link add anything, or is it pure self-promotion?

What happens to links that do not survive

A link does not always disappear entirely. Often it gets transformed:

Knowing these transformations helps you set more realistic expectations and target your optimisation more precisely.

The natural link pattern

Here is the core insight: AI systems learn link patterns from their training data. That data contains millions of web pages full of links. From it, the model learns where links naturally appear: in which contexts, at which sentence positions, with which anchor text. When it decides during extraction whether a link stays, it is implicitly asking: “Does this link sit where links normally sit in quality content?”

Links that match the natural pattern survive more often. Links that feel forced get stripped. Analysis of editorial content from quality publications shows a few recurring characteristics:

• Contextual relevance: the link sits exactly where a reader would want more detail, not as a detached call to action.
• Specific anchor text: descriptive and concrete (“real-time streaming architecture”) instead of vague (“click here”).
• First-reference linking: only the first mention of a concept gets a link, later mentions stay plain text.
• Information-dense context: links belong in paragraphs where the reader is actively learning something, not in an introduction or conclusion.

This links directly to how you structure your pages for extraction, which I explore in more depth in content architecture for AI extraction.

How a link prediction model works

These patterns are not just observations, they can be formalised into a prediction model. A link prediction model takes text as input and predicts where links would naturally appear. The training logic looks roughly like this:

1. Collect content with natural editorial links (news sites, quality publications).
2. Record the context for each link: surrounding text, sentence position, anchor text.
3. Train a model to predict link presence based on those features.
4. Have the trained model score new content.

The model picks up patterns such as: links often follow phrases like “read more about” or “according to”, links rarely sit in the first or last sentence of a paragraph, anchor text is usually 2 to 5 words long, and links mostly accompany the first mention of an entity.

Reading the prediction scores

For every possible link position, the model gives a probability score:

• High probability (>0.7): this position strongly matches natural patterns.
• Medium probability (0.4 to 0.7): plausible, but not optimal.
• Low probability (<0.4): this position matches no natural pattern.

Those same thresholds can be drawn as an ascending scale: the higher the score, the more strongly the position follows the natural pattern and the greater the chance your link survives extraction.

The optimisation is simple in principle: move links from low to high probability. You can also use the model to test content before publication, have optimal positions suggested, and analyse competitors.

Five principles to optimise your link placement

In practice it comes down to a handful of concrete rules.

1. Link at natural reference points. Place links where your content mentions something that invites elaboration: the first mention of an entity, the introduction of a concept, a source or figure, a recommendation. Not as a detached “read more”.

2. Use descriptive anchor text. Replace “click here for pricing” with something like “see the pricing and plans”. The anchor text describes the destination and helps the AI assess relevance.

3. Link only the first mention. If an entity appears several times, link only the first substantial mention. Repeatedly linking the same destination sends a spam signal.

4. Embed links in substantial content. A link in an information-dense paragraph is more natural than a link in thin text like “Want to know more? Check out our resources.” Deliver value around the link.

5. Respect editorial density. Quality publications typically link 1 to 3 times per 500 words. One well-placed link survives more readily than seven competing links in the same paragraph.

The context around the link counts

The words right before and after a link influence its odds of survival. Strong pre-link signals such as “read more about”, “as described in” or “according to” create the expectation of a reference. Weak signals like “click now” or “buy at” betray promotional intent and lower the naturalness. After the link, the sentence should flow on rather than collapse into a list of extra links. And links rarely sit right at the start or the end of a sentence: mid-sentence matches the editorial norm. This principle of smooth, contextual embedding overlaps heavily with the way you work on grounding snippets that AI is happy to pick up.

Internal versus external links

AI systems do not necessarily treat the two equally.

• Internal links (to your own domain) are sometimes seen as navigation rather than citation, and carry a lower authority signal because you are essentially pointing at yourself. They stay valuable if the linked page genuinely adds something. Keep the anchor text descriptive and avoid over-linking that resembles a navigation menu.
• External links (to other domains) carry a higher authority signal, because they read as an editorial reference to a third party. The risk is sending users away, so balance them with internal links.

One powerful pattern combines the two: an external citation of an authoritative source, followed by an internal link to your own relevant page. That mirrors editorial practice and can improve the survival odds of both.

The link-content relationship

The most important thing to remember: links do not exist separately from the content they sit in. If your content is strong enough to be selected as a primary source, your links come along into the selection. If your content is not selected, all link optimisation is wasted effort. So the order is clear: fix the quality and the selection odds of your content first, optimise the links only after that. Improving those selection odds is a discipline in itself, which I work out under Selection Rate Optimization.

And never sacrifice readability for a linking opportunity. The best link optimisation is invisible: the link sits where it belongs and adds value, instead of disrupting the flow.

Frequently asked questions

What exactly is a link prediction model?

It is a model that takes text as input and predicts where links would naturally appear. It is trained on large volumes of quality content with editorial links and learns from that which contexts, sentence positions and anchor texts correlate with links. You use it to test before publication whether your link placement feels natural, and to move links to positions with a higher probability score.

Why do some of my internal links disappear in AI answers?

Because during extraction the AI judges whether a link sits in a natural position. Links that feel forced, promotional or navigational are often stripped or converted to plain text or a standalone source reference. Links that match the editorial patterns, with descriptive anchor text in an information-dense spot, survive considerably more often.

How many internal links may a page have?

Quality publications typically link 1 to 3 times per 500 words. Too many links dilute the value of each individual link and send a downward quality signal. One well-placed link in substantial context survives extraction more readily than a handful of competing links in the same paragraph.

Should I optimise my content or my links first?

Content first. If your page is not selected as a source, your links get no exposure anyway, however well placed they are. Strong content that is chosen as a primary source carries the accompanying links into the selection with it. Link optimisation is the refinement, content quality is the foundation.

Need help?

Want to translate this into execution? See how we approach this with AI findability.