PageRank: Citation-Based Authority Ranking
Rank the web the way academia ranks research: by who cites you, and how authoritative they are.
- Difficulty
- Advanced
- Time to result
- ~months to results
- Steps
- 7
- Confidence
- 90%
PageRank is the mental model Larry Page and Sergey Brin used to solve web search's core problem: how do you separate authoritative content from noise at scale? Instead of counting keyword density like every other search engine of the era, they borrowed the academic citation model — a paper's importance is judged by how many important papers cite it. They mapped that onto the web: a hyperlink is a citation, and the anchor text around it is free descriptive metadata about the destination page. Because the web only exposes outbound links, they had to crawl and store a copy of the entire internet to reverse-engineer who links to whom. The result was a ranking system that surfaced the most authoritative pages first, and it worked almost immediately once implemented.
Origin
Began as a Stanford PhD dissertation idea to let users annotate web pages; Larry and Sergey realized they needed a way to rank whose annotations were trustworthy, which led them to citation analysis and, eventually, to ranking the pages themselves.
Core principles
- 01A link is a vote of confidence, not just a pointer
- 02Not all votes are equal — weight by the authority of the voter
- 03Metadata around a link (anchor text) often describes a page better than the page describes itself
- 04You can't ask the web who links to you — you have to go build that index yourself
How to run it
- 1
Reframe the problem
Larry and Sergey started by trying to rank human annotations on web pages, then realized the real unsteady variable was which sources were authoritative enough to trust — the same challenge every citation system faces.
Pro tip When a system starts to require its own ranking system to work, that's a sign you've found the real problem to solve.
- 2
Borrow a working model from an adjacent field
Academic citation analysis already had decades of research on how to weight the importance of a paper by who cites it and how important those citing papers are.
- 3
Map the model's atomic unit onto your domain
A hyperlink is functionally identical to an academic citation — an explicit, public statement that one piece of content is relevant to another.
- 4
Mine the metadata around the unit
Anchor text — the clickable words in a link — frequently describes the destination page better than the page describes itself, giving a free relevance signal.
- 5
Solve the missing-data problem
Web pages only display outbound links, not who links to them, so the team had to crawl and store a copy of the entire web to reverse-compute the backlink graph.
Watch out This only worked because the web was still small enough in 1996-97 to crawl on a research budget; a year or two later it would have cost tens of millions of dollars.
- 6
Weight links by authority, not just count
Pages linked to by other authoritative pages rank higher than pages with many low-quality inbound links, preventing simple gaming by raw link volume.
- 7
Validate against real queries before scaling
They tested the ranked output against existing engines' keyword-matching results and confirmed it surfaced authoritative pages first and spam last.
Pro tip Run a blind side-by-side comparison against the incumbent's results — it's the fastest way to prove a new ranking model to a skeptical audience.
In the wild
Larry and Sergey ran a live test comparing Excite's existing search algorithm to Back Rub's page-rank-based results in front of Excite's CEO as the final step of a licensing deal.
→ Back Rub's results were so obviously more relevant that the CEO killed the deal outright, fearing it would reduce page views and hurt Excite's ad-driven business model.
Once the ranking system was run against a modest crawled index at Stanford, it produced a sorted list of websites with the most authoritative pages consistently at the top and low-quality pages at the bottom.
→ This proof-of-concept convinced Larry's PhD advisor to greenlight the project and set off the chain of events leading to Google's founding.
Common mistakes
Ranking the wrong unit
The original plan was to rank user annotations/comments rather than the web pages themselves, which added unnecessary complexity before the real insight (rank the pages) emerged.
Assuming you can just query who links to you
The web's architecture only shows outbound links; teams that don't realize this underestimate the infrastructure required to build a real backlink-based ranking system.
Is it for you?
Not ideal for
Domains without a linkable, citable graph structure (e.g., closed private data sets with no cross-referencing behavior) where there is no natural 'citation' signal to mine.
From the transcript
“It wasn't that we intended to build a search engine. We built a ranking system to deal with annotations. We wanted to annotate the web.”
“When someone is linking to you, they often do a better job of describing your website than you do on the page yourself.”
From the episode
Google Part I: Origins of Search