How LLMs Actually Choose Their Sources

So traditional SEO is not dead: it’s the entry ticket. If you don’t rank, you’re far less likely to be in the candidate pool at all. But notice the flip side: roughly half of citations come from outside the top 10. Ranking gets you considered; it doesn’t guarantee you’re quoted.

Something else decides that.

Rankings get your page into the room. Relevance and trust decide whether the model actually quotes you once you’re there.

Step two: brand presence is the multiplier

This is where AI source selection diverges most sharply from classic SEO. Ahrefs analyzed 75,000 brands to see which factors correlate with showing up in AI Overviews. The strongest signals weren’t links. They were how widely a brand is talked about across the web.

Brand web mentions (every time your name appears across the web, linked or not) correlated at 0.664, about three times stronger than backlinks at 0.218. The effect compounds, too: Ahrefs found brands in the top quartile of web mentions averaged roughly 169 AI Overview mentions, more than 10× the next quartile down, while 26% of brands had zero AI Overview presence at all. LLMs learn what a brand is, and whether it’s an authority, from how often and in what context it’s discussed in their training and retrieval data. Seer Interactive’s research on ChatGPT found the same shape: presence in Google’s organic keyword results correlated at 0.65, while backlinks managed just 0.10.

Step three: write so a model can quote you

Once your page is retrieved, the model decides what to actually lift into its answer, and structure matters here as much as substance. The foundational research on this is the Princeton-led “GEO: Generative Engine Optimization” paper (accepted to KDD 2024), which tested optimization tactics across thousands of queries and found they could boost a source’s visibility in generative answers by up to 40%.

The tactics that mattered most were not keyword tricks. They were the things that make a passage easy to trust and extract: citing your own sources, adding direct quotations from credible authorities, and backing claims with specific statistics. In other words, content that already reads like a well-sourced answer is the content models prefer to quote. Practically, that means leading each section with a direct, self-contained answer, then expanding: exactly the principle behind citable content.

Every engine chooses differently

There is no single algorithm to optimize for. Each engine retrieves from a different index and weights trust differently, so the same brand can be cited heavily in one and absent in another. Ahrefs’ look at the most-cited domains in ChatGPT found Reddit, Wikipedia, Amazon, Forbes, and Business Insider at the top: a heavy lean on community and reference sources. Google’s AI Overviews, by contrast, draw far more on conventional organic results, and Perplexity weights real-time retrieval and citations differently again.

The takeaway isn’t to chase each engine’s quirks. It’s that durable AI visibility comes from signals every engine respects (a clear, authoritative, widely-referenced presence) rather than a trick tuned to one platform that disappears at the next model update.

Do backlinks still matter, then?

Yes: just not in the way they used to dominate. Backlinks still feed the ranking systems that decide which pages get retrieved in the first place, so they retain indirect value. What the data challenges is the assumption that link volume is the lever for AI citations. Across both the Ahrefs and Seer studies, link metrics consistently trailed brand and relevance signals: backlinks correlated at 0.218 in Ahrefs’ AI Overview analysis and just 0.10 in Seer’s ChatGPT analysis, while brand mentions and organic keyword presence sat two to six times higher.

The practical reframe is this: a backlink is now most valuable for the brand mention wrapped around it, not the link equity it passes. A paragraph on a respected industry site that names you, explains what you do, and links to you does double duty. It nudges rankings and adds to the web-wide brand footprint that LLMs read as authority. Chasing raw link counts from low-context directories does neither.

What this means for your content

Put the three steps together and the playbook is clear:

• Be retrievable. Rank for the query, keep the page crawlable and indexed, and use clean structure and schema so engines can parse it. If you’re not in the index, nothing else matters — the start of any AI visibility audit.
• Build broad brand presence. Earn mentions across the web — PR, expert commentary, partnerships, being named in other people’s content. Unlinked mentions count here in a way they never did for classic SEO. This is the heart of AI search and LLM visibility work.
• Write to be quoted. Lead with the answer, support claims with real data and citations, and keep passages self-contained so a model can lift them cleanly.
• Don’t abandon SEO. Rankings remain the entry ticket to the retrieval pool; treat them as necessary, not sufficient.