How Does the Google Algorithm Work?
Last updated
Most discussions about “the algorithm” treat it as a single system that can be reverse-engineered or beaten. It is neither. Google runs a set of named ranking systems simultaneously, each addressing a different dimension of quality. Understanding what those systems are, and what they measure, is more useful than chasing any individual ranking factor.
What does “the algorithm” actually mean?
Google runs automated ranking systems that evaluate hundreds of billions of web pages and return results in a fraction of a second. The word “algorithm” is shorthand for this whole stack, but Google actually publishes a list of its named systems: the Ranking Systems Guide.1
Some of these systems assess language and intent. Others evaluate authority, content quality, freshness, or spam. Several run in parallel on every query; others apply only in specific contexts. The result is that no single change to a page adjusts one lever in isolation. Pages are evaluated across the full set of active systems simultaneously.
"200 ranking factors" was never an official list
You may still see the claim that "Google has 200 ranking factors". It traces back to early Google statements, going back around two decades, that described "over 200 signals" for ranking. That was a passing description of complexity, not a published list. There is no fixed, countable set of factors to optimise against: Google runs many named systems that are recalibrated constantly. Treat any strategy built around a definitive ranking-factor count as a sign the source is dated.
The ranking systems Google names
Google’s Ranking Systems Guide lists the following as active systems.1
Language and intent
BERT (Bidirectional Encoder Representations from Transformers) understands how combinations of words express meaning and intent, rather than treating each word in isolation. A query like “can you get medicine for someone pharmacy” has a different meaning depending on the prepositions involved. BERT handles that level of nuance.
RankBrain understands relationships between words and concepts, allowing Google to return relevant results even when the page does not contain the exact words used in the query. It was the first machine learning component applied to Google Search at scale.
Neural Matching maps concepts in queries to concepts on pages. Where BERT handles query phrasing, Neural Matching works at a higher level of abstraction, connecting ideas expressed in different wordings.
MUM (Multitask Unified Model) processes text and images together and can draw on information across languages. Google applies it to specific applications rather than running it on every query, so its live footprint in Search is narrow.
Authority
Link Analysis Systems and PageRank evaluate how pages link to each other to determine relevance and helpfulness. PageRank, Google’s original ranking signal, remains part of the active system. The number, quality, and relevance of links pointing to a page are significant authority signals.
Quality and relevance
Reliable Information Systems surface authoritative pages and demote low-quality results. For topics where accuracy matters most, these systems also display content advisories when confidence in available information is low.
Reviews System rewards high-quality reviews that demonstrate first-hand experience and original research, rather than thin summaries of information available elsewhere.
Original Content Systems prioritise original reporting over sites that aggregate or repeat it after the fact.
Passage Ranking identifies relevant sections within a page and can rank individual passages for queries that match a specific section, even if the page as a whole is not the most relevant for the broader topic.
Freshness Systems surface newer content for queries where recency matters: news, recent events, and topics that change over time. Freshness is not a relevant signal for stable, evergreen queries.
Site Diversity System generally limits any single domain to two listings in the top results, preventing one site from dominating the page for a given query.
Deduplication Systems remove near-duplicate results so that only the most relevant version of substantially similar content appears.
Spam
Spam Detection Systems, which include SpamBrain (Google’s AI-based detection system), identify and neutralise content that violates Google’s spam policies: cloaking, auto-generated content at scale, manipulative link schemes, and similar tactics.
Contextual systems
Crisis Information Systems provide hotlines and emergency information during personal crises, natural disasters, or widespread emergencies. These override normal ranking logic for affected queries.
Local News Systems identify and surface local news sources in features such as Top Stories.
Exact Match Domain System prevents domains designed to match specific search queries (such as best-seo-tools.com) from receiving excessive ranking credit on the basis of the domain name alone.
How does Google combine these ranking systems?
The named systems above do not all run at the same moment. Google’s ranking works in two broad stages: retrieval and initial scoring, then re-ranking.
Stage one: retrieval and initial scoring
In the first stage, a primary scoring system pulls candidate pages from the index and gives each an initial relevance score, producing an ordered list, often described as roughly the top thousand candidate results for the query. Testimony and exhibits from the 2023 US antitrust trial, alongside the 2024 leak of Google’s internal Content Warehouse API documentation, describe this initial score as a combination of a small number of top-level signals: topicality (how well the page matches the query, built from anchors, body content, and clicks), quality (a measure of trustworthiness that PageRank feeds into), and a handful of others.2
Stage two: re-ranking with twiddlers
Once the first stage has produced its ordered list of candidates, that list is adjusted before it reaches the page. Google’s internal name for the functions that make these adjustments is twiddlers. A twiddler does not score pages from scratch; it takes the ranked list and nudges it, promoting some results, demoting others, filtering a few out, and enforcing rules about how the final page is composed. Several of the systems in Google’s public guide are re-ranking constraints of this kind: the Site Diversity System capping any one domain at two listings, the Deduplication Systems removing near-identical pages, and the Freshness Systems elevating recent content are all adjustments applied to an already-ordered list, not part of the initial score.
Named examples from the disclosures include NavBoost, which re-ranks using aggregated click behaviour over a rolling thirteen-month window, a quality boost, and a freshness twiddler.2 Google has confirmed the leaked documents are genuine but cautioned that they lack the context needed to infer exact weightings, and that some fields may be unused or historical. Treat the architecture as real and the precise mechanics as unknown.
Initial relevance is what gets a page into the candidate set, but engagement, quality, and freshness signals determine whether it is boosted or demoted within that set. There is no way to optimise for a single twiddler directly. The signals they act on are the same ones the rest of this page describes: satisfied clicks, genuine quality, and appropriate recency.
What is Query Deserves Freshness (QDF)?
Query Deserves Freshness, or QDF, is the idea that Google raises the weight it gives to recency for queries that show signs of being time-sensitive, and keeps it low for queries that do not. It is the mechanism behind the Freshness Systems listed above.
The term is not one Google maintains. It comes from a 2007 interview with Amit Singhal, then head of Google search, who described a model that tries to work out when a topic is “hot” and, if it is, favours fresh pages.3 Google’s current documentation calls this “Freshness Systems” rather than QDF, so treat QDF as useful shorthand for confirmed behaviour, not an official product name.
The signal Google uses to decide a query “deserves” freshness is a spike in interest: a sudden rise in how often people search the term, together with a burst of new publishing and news coverage on the topic. When both appear, recently published pages are temporarily elevated, sometimes ones published within hours. When the spike passes, the weighting relaxes and established, authoritative pages return to the top.
This applies to a minority of queries. Breaking news, recurring events (an annual awards show, a quarterly earnings date), and topics where the current answer changes (best cameras, prices, software releases) can trigger it. Stable, evergreen queries do not: adding “2026” to a guide on a subject that has not changed will not earn a freshness boost, because the query itself carries no freshness demand.
Two caveats matter. First, freshness never overrides quality and authority wholesale; it is a temporary adjustment on top of them, not a replacement. Second, do not manufacture freshness. Re-dating an unchanged page, or updating timestamps without materially revising the content, is a pattern Google identifies and can suppress. Freshness is earned by genuinely covering what changed.
Which systems are no longer separate?
Several systems that once ran as distinct, periodic filters have been absorbed into Google’s core ranking. They still influence rankings, but continuously, not in named update cycles.
Panda (2011, absorbed 2015) assessed site-wide content quality, penalising sites with a high proportion of thin or low-quality pages. Its logic now runs as part of core, meaning quality signals are assessed on every crawl rather than in periodic sweeps.
Penguin (2012, absorbed 2016) targeted manipulative link profiles. Since its integration into core, spammy links are devalued in real time as Google recrawls them, rather than requiring a Penguin update cycle to take effect.
Hummingbird (2013) was a near-complete rewrite of the core search algorithm to handle conversational and intent-based queries. It was not a penalty system but an architectural change that became the foundation for subsequent language understanding improvements, including RankBrain and BERT.
Helpful Content System (2022, absorbed March 2024) added a site-wide classifier that identified sites producing a high proportion of content written for search engines rather than people. It was integrated into Google’s core ranking systems in March 2024, meaning its signals now inform every core update rather than running as a separate layer.
Understanding this history explains why broad core updates now produce the kinds of ranking shifts that named updates once did. The systems that previously ran as distinct, occasional interventions are now continuous, and core updates adjust the weighting across all of them at once.
What this means for SEO
Individual ranking systems cannot be optimised in isolation. A page’s ranking reflects simultaneous evaluation across language, authority, quality, and spam systems. Improving one dimension while neglecting another produces limited results.
The consistent thread across all systems: they are designed to identify and reward genuine usefulness. BERT rewards clear, intent-matching writing. PageRank rewards content worth citing. The Reviews System rewards first-hand experience. The Reliable Information Systems reward accuracy and authoritativeness.
That consistency is why the most durable SEO approach does not change with algorithm updates. Pages built around genuine depth, clear authorship, earned links, and accurate information hold up across changes because they reflect what all of these systems, individually and collectively, are designed to find.
For a full timeline of how these systems evolved through named updates, see the Google algorithm updates timeline. For how core updates work and what to do when one affects your rankings, see what are Google core updates.
