Why LLMs Cite Some Pages and Ignore Others

Two pages can target the same keyword, rank on the same page of Google, and get treated completely differently by an LLM. One gets pulled into ChatGPT’s answer with a link. The other never shows up, no matter how many times you ask. The gap isn’t always about authority or backlinks. It’s usually about what happens in the moments before the model decides what to say.

This post breaks down the actual mechanics behind that decision, so you can look at your own content and understand why some of it gets cited and some of it doesn’t.

Retrieval happens before trust

Before an LLM can decide whether your page is a good source, it has to successfully pull the relevant information out of it. That’s a retrieval step, and it fails more often than marketers assume.

If a model can’t cleanly extract a claim from your page, that page is out of the running before credibility even gets evaluated. This is why a page with a good backlink profile and strong domain authority can still lose to a smaller, less authoritative page: the smaller page made its claim extractable, and the bigger page buried it in a wall of unstructured text.

Retrieval favors:

• A clear, direct answer near the top of the page or section
• One claim per paragraph, not several ideas blended together
• Headings that describe content instead of teasing it
• Lists and tables for anything comparative or sequential

None of this is new SEO advice dressed up. It’s a different mechanism. Google’s crawler indexes your whole page and ranks it against a query. An LLM’s retrieval step is closer to reading comprehension: it has to locate an answer inside your text on the fly, often with limited context window per source. Content that resists quick comprehension gets skipped even when it’s accurate and well-researched.

The claim has to exist before it can be cited

A surprising number of pages that “should” get cited never make a citable claim at all. They discuss a topic without ever stating a position.

Take a page about “choosing a project management tool for agencies.” If the page walks through five options with pros and cons but never says which one is best for which situation, there’s nothing for the model to cite as an answer. The model has to synthesize its own conclusion from ambiguous material, and it will usually prefer a source that already did that synthesis for it.

Pages that get cited tend to make a direct, falsifiable claim: “For agencies under 20 people, Tool X is the better fit because of Y.” That’s a sentence a model can lift and attribute. A page that just lists features isn’t a source, it’s raw material, and raw material gets used less often than finished conclusions.

Source type shapes trust, independent of content quality

This is the part that frustrates a lot of brands: the same claim, worded the same way, gets treated differently depending on who’s saying it.

Source type	How LLMs tend to treat it
Brand’s own website	Useful for facts about the brand (pricing, features) but discounted for comparative or evaluative claims
Independent review sites (G2, Capterra)	High trust for evaluative claims, seen as less biased
Reddit and forum threads	High trust for real-world experience and edge cases
Industry publications and press	High trust for market context and third-party validation
Wikipedia	Very high trust as a foundational, neutral reference

A claim like “this is the best tool for small agencies” carries more weight coming from a Reddit thread or a G2 review than from the vendor’s own comparison page. That doesn’t mean brand-owned content is worthless. It means brand-owned pages perform better making factual claims about themselves (pricing, integrations, specs) and third-party content is what carries evaluative and comparative claims. A content strategy built for GEO accounts for this split instead of trying to make every claim from the same channel.

Freshness and specificity compound each other

Vague content ages faster than specific content, and stale content gets cited less regardless of how vague or specific it is. These two factors interact more than people expect.

A page that says “AI search is becoming more important” is both vague and destined to look dated within months. A page that says “as of Q3 2026, three of the top five AI search tools support real-time retrieval from indexed pages” is specific enough to be useful today and specific enough that its shelf life is obvious, which makes it a candidate for a scheduled update rather than something you’d forget about for two years.

Specific content also tends to get updated more often, because specificity forces you to notice when a claim is outdated. Vague content can sit unchanged for years without anyone noticing it’s wrong, which means it slowly drifts out of citation rotation as models start preferring fresher sources on the same topic.

What this means for a content plan

None of these factors work in isolation. A page can have perfect structure and still lose out if it’s making a claim better suited to a third-party source. A page can be hosted on a high-trust domain and still get skipped if the claim is buried three paragraphs down.

The practical version of this, in order of what to fix first:

1. Check extractability. Can a human skim your page in ten seconds and state your main claim? If not, an LLM can’t either.
2. Check whether you’re making a claim at all. Comparison and roundup content especially tends to hedge. State the conclusion.
3. Check whether the claim belongs on your domain or a third-party one. Evaluative claims often perform better seeded through review platforms or Reddit than through your own blog.
4. Check freshness on anything with a shelf life. Numbers, rankings, and “best of” claims need a maintenance cycle.

This is the same logic behind structuring blog posts for citation: the mechanics of extraction, claim-making, and source trust aren’t separate concerns, they’re the same problem viewed from different angles. Fix the extraction and the claim clarity first. Everything else compounds from there.