How Google Search ranking works

Stop picturing one algorithm

The most useful mental model I can give you after two decades of doing this: Google is not a formula, it is an assembly line. A query enters, passes through distinct stages, and a results page comes out the other end. Each stage has its own systems, its own failure modes, and its own optimization levers. When rankings move, the diagnostic question is not "what did the algorithm do," it is "which stage changed its mind about me."

Stages 1 and 2: crawling and indexing

Before ranking exists, Googlebot has to fetch your page and Google has to decide it deserves a slot in the index. Indexing includes rendering (Googlebot runs an evergreen Chromium, so JavaScript content does get seen, eventually), canonicalization (picking one URL among duplicates), and storing signals about the page and its site. Plenty of "ranking problems" I get called in for are actually indexing problems: the page is canonicalized away, stuck in Discovered - currently not crawled, or serving different content to the renderer. Check Search Console's URL Inspection before theorizing about anything fancier.

Stage 3: understanding the query

Google rewrites and interprets queries before matching them to documents. Its documented systems here include RankBrain (launched 2015, its first deep-learning ranking system, helping with queries it has never seen), neural matching (relating query concepts to page concepts), and BERT (rolled out in 2019 to model how words change meaning in context). This is why exact-match keyword stuffing keeps losing value: the system matches meanings, not strings.

Stage 4: retrieval and core scoring

From the index, Google pulls a candidate set and scores it. The signals Google itself documents: topicality (does the content actually address the query), links and PageRank (still in Google's own ranking systems documentation, diminished from 2010 but very much alive), freshness where the query deserves it, page experience, and site-level quality signals. E-E-A-T deserves a clarification because it gets mangled constantly: it is not a score. It is the rubric human quality raters use, and Google says its systems use measurable proxies to approximate it. You cannot "add E-E-A-T"; you can build the evidence its proxies detect.

Stage 5: reranking, Navboost, and twiddlers

This is the stage the industry underestimated for years, and the one the US v. Google antitrust trial dragged into daylight. In 2023 testimony, Google's Pandu Nayak confirmed Navboost, a system that aggregates 13 months of user interaction data on queries and results and uses it to adjust rankings. Trial exhibits also referenced Glue, a sibling system covering non-web SERP features. So yes: aggregated user behavior influences rankings. That is not the same as "CTR is a ranking factor you can game with click bots"; Google has systems specifically for detecting that, and the data is aggregated and slow-moving.

The May 2024 leak of over 2,500 internal Content API Warehouse documents added texture: attributes like siteAuthority, hostAge, and Chrome-derived view data, plus "twiddlers," reranking functions that boost or demote after initial scoring. Google confirmed the documents were real but warned that a field existing in storage does not prove it is used in ranking. My read: treat the leak as a parts catalog, not a wiring diagram. It tells you what Google can measure, not what it currently weighs.

Also living at this stage: the documented boost-and-demote systems. SpamBrain handles link spam and hacked content. The helpful content system, which flattened a lot of thin affiliate sites in 2022 and 2023, was folded into the core ranking systems in the March 2024 core update. The reviews system rewards first-hand product testing. These site-level classifiers are why one bad content strategy can drag down your good pages too.

What this means for how you actually work

First, satisfied clicks are the endgame. If Navboost is watching 13 months of interactions, the page that answers the query and gets remembered beats the page that merely matches keywords. Second, quality is scored at the site level, so prune or fix the dead weight instead of hoping your good pages carry it. Third, links still matter, but as one stage-4 signal among several, not the whole game. Fourth, stop chasing single factors. A pipeline with reranking passes means a page can win stage 4 and get quietly demoted in stage 5, which is exactly what a helpful-content demotion looks like in the wild.

And diagnose in pipeline order: indexed correctly? Understood for the right queries? Competitive on core signals? Only then start asking whether a classifier is sitting on you.

FAQ