Backlinks are the strongest independent correlate of AI recommendations in OppAlerts' 403,000-prompt study: ahead of Reddit, ahead of news coverage, ahead of Wikipedia and Wikidata, ahead of everything else measured.The full leaderboard, the partial-correlation test, and the caveats are in What Actually Correlates with AI Search Visibility; method and tables in The Research Behind This Guide, part of the AI Search Visibility research. The oldest signal in search marketing predicts the newest channel. This chapter explains why that is not a coincidence, and what it changes about link work.
The index inheritance effect
Links reach an AI answer through both of the sources described in Model Memory and Live Retrieval, and through a different mechanism in each.
On the retrieval side, the inheritance is direct. When an AI product searches, it runs queries against Google, Bing, or Brave, and those indexes still rank with link-based authority at their core. Google's founding paper treated the web's link graph as a citation graph and built PageRank on it, and link authority has been a load-bearing part of web ranking ever since.Brin and Page, The Anatomy of a Large-Scale Hypertextual Web Search Engine (1998), which introduced PageRank as "citation (link) graph" analysis. The modern systems are far more complicated, but links never left. A retrieval pipeline that reads the top-ranked pages inherits every authority judgment those rankings encode. The model never sees your backlink profile; it sees the pages your backlink profile pushed into the results.
On the memory side, the inheritance is statistical. A brand that has earned links from thousands of independent sites is, almost by definition, a brand that thousands of independent sites wrote about. Those pages are the training corpus. Heavy linking and heavy mention are the same underlying fact about a brand, so the link graph predicts what the model absorbed even though training never parses a link graph as such.
One signal, two routes in. That is the assessment for why backlinks top the correlation table on both memory answers and searched answers.
What the data says
Measured one signal at a time, backlink authority ties Reddit discussion volume at the top of the leaderboard (Spearman rho 0.23). The stronger claim comes from the partial-correlation test, which holds the other signals fixed: backlinks keep a real independent relationship with AI recommendation strength (partial correlation 0.077) while Wikipedia and Wikidata drop to roughly zero. Brands in Wikipedia carry about 1.26x the recommendation strength of brands that are not, but compare two brands with the same backlink authority and that lift falls to 1.00 to 1.03. The encyclopedia page was never the lever; the authority that earned the page was. Backlinks also rank as the strongest single signal inside all 13 memory models tested, and win the industry-by-industry count 75 of 100.
All of this is correlational, and controlled experiments on authority signals largely do not exist. The honest position is the one the evidence chapter states: strongly consistent with a lever, unproven as one. The mechanism above is why I treat it as a lever anyway.
Linked citations versus unlinked mentions
SEO forums spent a decade arguing whether unlinked brand mentions matter. The two-source split settles the argument: both matter, through different machinery.
Retrieval weighs the link. Rankings move with link authority, so a linked citation from an authoritative page does what it always did: it helps the linked page, and every page on the domain, rank in the indexes AI products search.
Memory weighs the mention. Training does not transfer authority along hyperlinks; it learns statistical associations from text. A mention of your brand next to your category's vocabulary teaches the model what you are whether or not it carries an href. An unlinked mention in a widely read publication does close to nothing for rankings and real work for the model's memory of you.
So the old dismissal ("no link, no value") and the old consolation ("mentions are just as good") are both wrong, each about one mechanism. When you pitch coverage, take the link when you can get it, and stop treating a linkless placement as a failure. Digital PR, News, and the Gatekeeper Publishers builds the earning side of this.
Co-occurrence: being in the right sentences
The memory mechanism is more specific than "get mentioned a lot". Models represent meaning as embeddings, vectors where distance tracks how often things appear in similar contexts, and Embeddings: How Machines Represent Meaning covers the mechanics. What moves your position in that space is which words your brand keeps company with across the training corpus.
A thousand mentions of your brand in sentences about your category ("accounting software for freelancers, alongside FreshBooks and Wave") place you inside the category's neighborhood. A thousand mentions in funding announcements and hiring news place you in a different neighborhood, one that recommendation prompts never query. The practical test for any placement: does the surrounding sentence say what you are and who you are for, or does it only say that you exist. Coverage that states facts in category language is the kind a model can absorb into an answer.
Listicles, roundups, and directories
"Best X for Y" pages deserve their own line item, because they sit at the intersection of both mechanisms and of the prompt classes that decide revenue. When a recommendation prompt triggers a search, the queries the system generates look like listicle titles, so listicles and comparison roundups dominate what gets retrieved. Being on the page that ranks is close to being in the candidate set for the answer. The same pages get crawled into training corpora, where they are dense, clean co-occurrence: your brand, your competitors, and the category vocabulary in one document, often with rankings and reasons attached.
Treat the roundups that rank for your category's prompt-shaped queries as a target list, and treat absence from them as a measurable gap. The commercial version of this work, including comparison and alternatives pages you build yourself, is in Commercial Visibility: Products, Comparisons, and Agentic Buying.
What survives from link building, and what changes
The veteran's instincts mostly transfer. Editorial links from real publications, earned with something worth covering, remain the asset; Evidence, Expertise, and Original Research covers the strongest earning strategy. Link schemes remain the liability, and the risk calculus is covered in Manipulation, Spam, and Risk.
What changes is the target list and the reasons. Anchor text optimization matters less than the sentence around the mention. Domain-authority spreadsheets matter less than two questions about each prospect: does this page rank in the indexes AI products actually search, and will this page enter the corpora models train on. A placement that does neither is decoration. A placement that does both is working on this year's answers and next year's model at the same time.