What Backlinks Are Actually For

01The link was a citation, not a vote

Every account of what backlinks are for starts in the same place: a link is a vote. It is the founding metaphor of the whole discipline, and it is not in the founding documents. Three of them were searched in full text: the 1998 Anatomy paper by Sergey Brin and Lawrence Page, the PageRank Citation Ranking technical report by Page, Brin, Rajeev Motwani and Terry Winograd, and US6285999B1, which names Lawrence Page as its sole inventor and assigns the invention to the Board of Trustees of the Leland Stanford Junior University, not to Google. The name in the algorithm and the sole inventor on the patent are the same person; the argument behind it is four people's work, and the property was a university's. The word "vote" appears in none of the three. What does appear is a pair of models that are not votes either: a citation whose weight depends on who is doing the citing, and a reader clicking links at random, whose visit probability is the score itself — "the probability that the random surfer visits a page is its PageRank". Section four returns to what Google later did to the second model.

What appears instead is "citation". Page and Brin were not inventing a democracy; they were carrying academic bibliometrics onto the web, where a link plays the part a footnote plays in a journal. The distinction matters because the two metaphors license different behaviour. A vote is something you cast, and more of them are better. A citation is something you owe, and it is only worth anything if it is honest.

The vote language does exist — in Google's own product marketing. The 2004 Google Technology page states it plainly, and that page, not the research, is where the industry got it. Reading it today, the sentence is a simplification of a citation model for a lay audience, and it was taken as a specification.

02What a link does: it says what a page is about

In February 2025, Google engineer Hyung-Jin Kim described the ranking stack to the Department of Justice. The resulting exhibit names three fundamental signals and calls them ABC. Anchors are the A, and the exhibit defines them in five words: a source page pointing to a target page.

The consequential sentence is the next one. The three signals are not described as measures of quality. They are "the key components of topicality" — Google's determination of how relevant a document is to a query. Links sit in the relevance machinery, next to the terms in the document and the clicks on the result.

Read that way, a link is a statement about subject matter before it is a statement about worth. The linking page asserts that the target belongs to a topic; the anchor text and its surroundings say which one. That is a much smaller claim than a vote, and a much more checkable one — which is presumably why it survived twenty-five years of people trying to forge it.

One caveat belongs here, and it comes from Google's own side. Three weeks before Kim, Pandu Nayak described the same signal differently: the anchor signal, he said, identified a document's quality in part by how many other documents linked to it. Quality, by count — the exact reading this piece argues against. Both statements are sworn material from the same proceeding.

Nayak is speaking in the past tense, about the era of content farms, and his next sentence supplies the ending: the abuse of that proxy is what forced Google to develop separate quality signals. Read together, the two engineers describe a shift rather than a disagreement — what was once a quality count became a topical statement once counting it stopped working. That reading is ours, not theirs, and the contradiction is left standing rather than explained away.

03Authority runs on a different axis

If links carry relevance, what carries authority? The same exhibit answers that too, and the answer is a separate signal on a separate axis. PageRank is described in one sentence as a measure of distance from a known good source, feeding the Quality score rather than the topicality score.

Q* is glossed there as page quality in the sense of trustworthiness, and it is described as largely static across queries and largely a property of the site rather than the page. That is the thing the industry calls authority. It is fed by PageRank, but it is not PageRank, and it does not live where the anchors live.

One consequence follows immediately and is worth stating because it cuts against a comfortable assumption. Distance from a known good source is measured along paths that run toward a page. Linking out does not move you closer to anything. Whatever outbound links do — and the next section argues they do something — they do not do it here.

04The other direction: what linking out is worth

The hypothesis this piece started from runs the other way. If I choose the sources I build an argument on, and I show them, I have demonstrated that I read them. That is a claim about expertise, made by the act of linking rather than by the receipt of a link. Does anything support it?

Something does, from an unexpectedly old and unexpectedly direct source. In 2009 Matt Cutts, then head of Google's webspam team, answered the question on his own blog. The answer is careful in a way the industry has mostly discarded: linking out does not help PageRank, and other parts of the system encourage it anyway. Two different mechanisms, and only one of them is the one everybody optimises for.

Google's current guidance points the same way without naming links. The helpful-content self-assessment asks whether content presents information in a way that makes you want to trust it, "such as clear sourcing". The Quality Rater Guidelines go further on the concept and stop short on the mechanism: their worked examples say citations support a page's E-E-A-T, and a full-text search of the current edition finds no passage that treats an outbound hyperlink as a quality signal. The idea is endorsed; the implementation is never named.

The one place the idea has been formalised is Hilltop, which defines an expert document as a page linking to many non-affiliated pages on a topic — non-affiliated meaning, concretely, different IP neighbourhoods and different hostnames. That is this hypothesis, stated as an algorithm in 2000. Two caveats belong with it: Hilltop uses expert documents to rank their targets rather than to reward the expert, and despite the common claim, Google never owned the patent (US7346604B1 ran DEC → Compaq → HP → HPE → Valtrus).

05What this leaves a new site

Put the axes together and the advice that follows is unfamiliar in its emphasis rather than in its content. Inbound links carry topical relevance and, through PageRank, feed a quality score that is largely site-level and largely static. Outbound links carry no PageRank of their own and are nonetheless encouraged. Neither is a lever you pull; both are consequences of being the kind of page other pages have reason to cite.

Two corrections to the volume instinct sit in the documented material. The first is the Reasonable Surfer patent, and it is worth naming whose assumption it abandons: PageRank's own. Alongside the citation analogy, the 1998 paper justifies the score with a second model — a surfer clicking links at random, where "the probability that the random surfer visits a page is its PageRank" — and Page's patent formalises that as a transition probability matrix over the graph. Every outgoing link is equally likely to be taken. The reasonable surfer is the same imaginary reader with the coin flip removed: the model "reflects the fact that not all of the links associated with a document are equally likely to be followed", so each link gets its own probability, computed from features of the link itself — font size of the anchor text, position on the page, commerciality of the anchor, whether the target sits on the same host — and calibrated against user behaviour data on the navigational actions actually taken at that link. The dating deserves care: the application was filed 17 June 2004 and granted in 2010 as US7716225B1; US8117209B1, cited here, is its continuation, granted 2012, and the family runs on to US10152520B1 in 2018. Six years, then, separate the uniform model from the weighted one — and both are Google's. Worth noting in passing that the name outlasted its definition twice: in 1998 PageRank was a visit probability, which is a popularity measure, while the 2025 exhibit calls it a distance from a known good source, which is not the same quantity. The second correction is the leaked ranking corpus, which contains three link-related fields against thirty-eight documented in total — all three of them on the penalty side, measuring anchor spam rather than anchor value.

The honest closing note is about the limits of all of this. The same Google engineer who described ABC to the DOJ also described what the leaked documents do not contain, and it is the best-sourced caution in the whole piece: the documents name components, not curves and thresholds. Everything above establishes what the parts are for. None of it establishes what any of them weighs.

06References