When someone says Google changed the algorithm, they mean one of the systems on this page. Google maintains an official guide to its ranking systems, and this pillar walks through all of them: what each one does, which ones deserve your attention, which are dead, and the layer Google prefers to keep out of its documentation. Where a claim rests on testimony or leaked material instead of official docs, it’s labelled that way.
Systems, signals and updates: 30 seconds of orientation
A system is standing machinery: it runs continuously, doing one job in the ranking pipeline described in how Google Search works. A signal is an input those systems read, like page speed or anchor text. An update is Google changing the machinery, which is when rankings lurch and the update history gets a new row. Most confusion in SEO commentary comes from mixing the three.
Every documented system, in one table
| System | What it does | Worth knowing |
|---|---|---|
| Core ranking systems | The main relevance and quality machinery | Refreshed by core updates; absorbed the helpful content system in March 2024 |
| RankBrain | Relates words to concepts for unfamiliar queries | First ML system Google confirmed (2015) |
| BERT | Reads word combinations in context | Query understanding, live since 2019 |
| MUM | Multimodal language understanding | Used in specific features, per Google, rather than general ranking |
| Neural matching | Matches queries to pages by concept rather than keywords | Retrieval-side machinery |
| Passage ranking system | Scores individual sections of a page | Long pages can rank on one good passage |
| Link analysis systems and PageRank | Weighs links between pages to judge importance | Diminished since 1998, still alive |
| Spam detection systems | SpamBrain and friends, catching policy violations | Refreshed via spam updates |
| Reviews system | Rewards first-hand, evidence-based review content | Runs continuously; updates unannounced since November 2023 |
| Freshness systems | Prefers recent content where the query deserves it | Query-dependent, so evergreen pages are safe elsewhere |
| Original content systems | Ranks original reporting ahead of those citing it | Mostly felt in news |
| Deduplication systems | Collapses near-identical results | Why syndicated copies vanish |
| Exact match domain system | Stops keyword domains ranking on the name alone | Killed a 2000s tactic |
| Site diversity system | Caps most domains at 2 top results | Why doubling up on one SERP is rare |
| Crisis information systems | Serves vetted information in emergencies | SOS alerts, personal crisis queries |
| Local news systems | Surfaces local outlets in news features | Top stories, local packs |
| Reliable information systems | Lifts authoritative sources, demotes low-quality ones | The docs’ broadest quality claim |
| Removal-based demotion systems | Demotes sites generating volumes of valid removal requests | Legal and personal-info removals |
The systems that decide most outcomes
Core ranking systems
The centre of gravity. Everything Google says about them, and the strategy for living with them, is in what is a core update and core update recovery. Since the helpful content system merged in, sitewide quality assessment lives here too, which raised the stakes of every core rollout.
The language stack: RankBrain, BERT, MUM, neural matching
Four systems doing related work: turning queries and pages into meaning rather than keyword strings. Practical consequence: exact-phrase matching keeps losing value, and covering a topic in natural language keeps gaining it. The community widely believes Google’s MUVERA multi-vector retrieval research entered production around the June 2025 core update; Google hasn’t confirmed it, and that claim stays labelled as informed speculation here.
Link analysis and PageRank
Still running, still named in the official guide, and much diluted from the days when it was the whole story. Links remain the strongest external vote a page can get; internal links are the version you control, and the internal linking guide covers extracting full value from them.
Spam detection: SpamBrain
The enforcement arm. It neutralises link schemes, scaled content abuse and the rest of the spam policies, and its refreshes appear in the update history as spam updates. Modern SpamBrain tends to devalue rather than demote: bought links quietly stop counting, which is why link sellers can show you the link and never the effect.
The reviews system
Rewards reviews showing evidence of actually using the thing: photos, measurements, comparisons with alternatives. Since November 2023 it updates continuously without announcements, so review-site volatility no longer maps to a named update. If you publish reviews, the experience letter of EEAT is the whole game.
Retired systems, and the zombie advice they left behind
| System | Fate |
|---|---|
| Panda | Absorbed into core ranking systems, 2015 |
| Penguin | Absorbed into core ranking systems, 2016 |
| Hummingbird | 2013 engine rewrite, long since superseded |
| Helpful content system | Merged into core ranking systems, March 2024 |
Google also reworked its documentation in late 2023 to stop describing page experience as a distinct system: Core Web Vitals and friends are signals the other systems read, useful and modest in weight. The practical point of this table is defensive. When an audit tells you that you’ve been hit by Panda, you’re reading a document written by someone a decade behind the documentation.
The undocumented layer: Navboost and the click signals
Google’s official guide is honest and incomplete. Two events forced the gap into public view: the US antitrust trial and the May 2024 leak of internal Content Warehouse API documentation. This section grades its own evidence, because the sourcing quality varies.
- Sworn testimony (strong evidence): Google’s Pandu Nayak confirmed under oath that Navboost is one of Google’s important ranking signals: a system that re-ranks results using aggregated user click behaviour over a rolling 13-month window. Trial exhibits included a Google engineer describing it as more powerful on click metrics than much of the rest of ranking combined.
- Leaked documentation (real, but read with care): the 2024 Content Warehouse leak exposed thousands of internal attributes, referencing Navboost extensively: good clicks, bad clicks, long clicks, and a siteAuthority attribute. Google confirmed the documents were genuine while warning against out-of-context conclusions, and that warning is fair: the leak shows what Google can store, with no weights attached, and an internal attribute existing tells you nothing certain about how, or whether, it’s used in live ranking today.
- Correlation studies (weak evidence): plenty of published work shows engagement metrics correlating with rankings. Correlation studies can’t separate cause from effect, and never have. Treat any confident causal claim built on them accordingly.
What should you do about Navboost? Roughly nothing you weren’t already doing. If clicks and post-click satisfaction feed rankings, the winning move is a result people choose and stay on, which is the same thing Google’s people-first guidance asks for and the same thing that survives core updates. Attempts to fake click signals run straight into the fraud filtering the leaked docs also describe. The lesson of the undocumented layer is that Google measures satisfaction harder than its documentation admits, and satisfaction was already the job.
Where to go deeper
Each spoke of this page has its own guide: how Search works, crawling and indexing, EEAT, YMYL, AI Overviews, internal linking and crawl budget. For what changed and when, the update history is the companion pillar to this one.
Sources
Google’s guide to ranking systems (all current and retired system descriptions), the spam policies, DOJ trial testimony and exhibits as reported in the trial analyses, and the Content Warehouse leak documentation. Evidence grades and the advice at the end are mine. Each source above has its own entry in the research library, with the method, sample and limitations set out in full.