The Complete History of Google's Search Algorithm: From PageRank to AI Overviews

1998–2002: Links as votes

Larry Page and Sergey Brin's PageRank, published in 1998 while both were at Stanford, treated the hyperlink structure of the web as a citation network rather than a pile of text to keyword-match.

The patent behind it, US 6,285,999, was filed on January 9, 1998, names Page as inventor, and was assigned to Stanford. Its "random surfer" model imagines someone clicking links at random; a page's score is the probability that surfer lands on it.

A link from a page with few outbound links and high authority itself passed far more weight than a link from a low-authority page linking out to hundreds of others.

For most of this era, Google refreshed its index in roughly monthly batches, an event webmasters nicknamed the "Google Dance" because rankings visibly swung for days as the new data reached each data center. PageRank on its own counted link weight without checking topical relevance, so link farms and reciprocal-linking rings soon followed. The first update that webmasters documented came in September 2002. Google never said what it changed.

2003–2009: The first spam-fighting decade

Six named changes arrived in 2003 before November: Boston, Cassandra, Dominic, Esmeralda, Fritz, and the Supplemental Index. Esmeralda, in June, turned out to be the last of the regular monthly updates, and with Fritz in July Google announced a change to its update policy.

The most consequential update of the year was Florida (November 2003). Google has never disclosed what Florida changed. Many SEOs at the time credited Hilltop, a 1999 algorithm for finding "expert" pages linked to by independent authorities, while Search Engine Journal's retrospective argues for statistical link analysis.

Either way, many commercial and affiliate sites lost rankings, and commentators still describe Florida as the update that changed how SEO was done.

Brandy (February 2004) brought a large index expansion. It is often described as introducing Latent Semantic Indexing, meaning the matching of synonyms and related concepts, but that label comes from SEO commentary rather than from Google, and Google's John Mueller said in 2019 that "there's no such thing as LSI keywords." In January 2005, Google, Yahoo, and Microsoft jointly introduced rel="nofollow", an attribute that tells search engines not to pass credit through a link.

It gave blogs a way to take the reward out of comment spam, and it later became the standard marking for paid links. With 2009's Vince update, big brands rose for broad commercial queries.

Google's Matt Cutts called it a minor change and said Google thinks in terms of trust, authority, and reputation rather than brands.

2010–2013: Panda, Penguin, and Hummingbird

Before Google could run sophisticated quality filters, it needed infrastructure that could apply them in near real time. Caffeine (June 2010) rebuilt the crawling and indexing pipeline from a batch process into a continuous one. The old index was built in layers, and the main layer was refreshed every couple of weeks. Google said Caffeine delivered 50 percent fresher results than the index it replaced.

With that infrastructure in place, three landmark classifiers shipped in under three years:

Panda assessed quality across a whole site: a domain with a high share of thin or duplicated pages could be demoted as a whole, dragging down its good content too. Penguin, announced on April 24, 2012 as a webspam change, lowered rankings for sites violating Google's quality guidelines through tactics "from keyword stuffing to link schemes." Google said it affected about 3.1% of English queries to a degree a regular user might notice.

In practice, SEOs found it keyed on patterns across a backlink profile, such as heavy exact-match anchor text and link networks, rather than on any single link. That's the same footprint logic behind why a direct, reciprocal link exchange still carries risk today.

Google now lists both as retired systems: Panda became part of the core ranking systems in 2015 and Penguin in 2016. Hummingbird, launched in August 2013 and announced that September, was an overhaul of the core ranking engine.

Google described it at launch as the biggest rewrite since 2001, built to handle conversational queries by reading the meaning of the whole query instead of matching individual words.

2014–2017: Mobile-first and the first AI ranker

Smartphone adoption forced a structural rebuild. The Mobile-Friendly Update of April 21, 2015, nicknamed "Mobilegeddon," made mobile-friendliness a ranking signal for mobile searches in all languages worldwide. It preceded a multi-year shift to Mobile-First Indexing, announced in 2016 and rolled out from 2018, where Google indexes a site's mobile version, not its desktop version, as the primary one.

In October 2015, Google confirmed it was using RankBrain, a machine-learning system that helps it understand how words relate to concepts. It is not a spam filter. It converts queries into vectors so that a novel or ambiguous search can still match pages that never use the literal search terms.

Google said at the time that RankBrain was its third most important ranking signal, after links and content. Locally, Pigeon (July 2014) tied local results more closely to the signals used for ordinary web ranking. Possum (September 2016) and Fred (March 2017) are names the SEO industry gave to changes Google did not detail.

Local SEOs reported that Possum gave more weight to the searcher's location and filtered out similar listings at the same address, and that Fred hit ad-heavy, thin informational sites hardest.

2018–2020: Trust becomes a formal framework

On August 1, 2018, Google released a broad core update that the SEO industry nicknamed Medic, because health and medical sites saw the largest swings. Google said it targeted no particular vertical. The industry connected it to E-A-T (Expertise, Authoritativeness, Trustworthiness), the framework in Google's Search Quality Evaluator Guidelines, and to Your Money or Your Life topics (health, finance, safety), where those guidelines hold pages to a higher standard.

Google did not confirm that link, but from this point E-A-T became the lens most SEOs used to judge content on sensitive topics: who wrote it, what their credentials are, and what it cites.

BERT (October 2019) gave Google bidirectional language understanding, reading a word's meaning from the words on both sides of it. That helped with prepositions such as "for" and "to" in long, conversational queries.

Google said at launch that BERT would apply to one in 10 searches in US English. Passage Ranking, announced in October 2020 and live for US English queries from February 10, 2021, lets Google score a specific passage deep inside a long page when judging relevance.

Google still indexes whole pages, not passages, but a single strong section can now earn a ranking for a narrow query.

2021–2023: Multimodal AI and real experience

MUM (announced May 2021) is a model Google described as 1,000 times more powerful than BERT. It is trained across 75 languages and understands text and images, with video and audio planned. Google pitched it at complex, multi-step questions, though its ranking systems guide says MUM is not currently used for general ranking, only for specific tasks.

In June 2021 the page experience update began rolling out, making the Core Web Vitals (loading, interactivity, and layout stability) a ranking signal. Behind the scenes, SpamBrain, the AI-based spam-prevention system Google launched in 2018, became central to its spam fighting.

In its 2022 webspam report, Google said SpamBrain detected 50 times more link spam sites than in the previous link spam update.

The Helpful Content System (August 2022) introduced a site-wide signal aimed at what Google called a "search engine-first approach": content produced to rank rather than to help a reader. Its last standalone update rolled out from September 14 to 28, 2023.

In December 2022, E-A-T grew an extra E for Experience: does the content show that the author actually used the product or visited the place? This is a change to the guidelines human quality raters use, not a ranking switch, but it tells you what Google wants its systems to reward: first-hand detail over secondhand summary.

2024–2026: Generative answers and a harder line on spam

The March 2024 Core Update began on March 5 and took 45 days to roll out. It changed multiple core systems at once, and Google now lists the helpful content system as retired because it became part of core ranking with this update.

Google later reported 45% less low-quality, unoriginal content in results. Three new spam policies arrived alongside it: scaled content abuse (many pages generated mainly to manipulate rankings, whether by automation, people, or both), expired domain abuse (buying an expired domain and repurposing it mainly to manipulate rankings), and site reputation abuse (third-party pages published on a host site with little or no first-party oversight, often called "parasite SEO"). The third took effect on May 5, 2024.

AI Overviews began rolling out to everyone in the US on May 14, 2024, announced at Google I/O. They use a Gemini model customized for Search to write a summary with links to supporting pages.

Those links come largely from pages that already rank: an seoClarity analysis of 36,000 keywords, reported by Search Engine Land, found that 99.5% of AI Overviews cited at least one page from the top 10 organic results. That is a measure of "at least one match," and studies that count every citation find much lower overlap.

This is also the period when ChatGPT search and Perplexity built their own retrieval pipelines on top of the open web, a mechanism our AI-era ranking framework breaks down pipeline by pipeline.

Core updates continued after that. Google's Search Status Dashboard lists three more in 2024 (August, November, and December) and three in 2025: March (from March 13), June (from June 30), and the December 2025 Core Update, which ran from December 11 to December 29.

Google gave no specific targets for any of them. SEO analysts tracking the December rollout reported volatility for forums that appear in the "Discussions and forums" feature and traffic losses at Quora, but those are third-party observations, not Google statements.

2026 opened with the February 2026 Discover core update, announced on February 5 as a broad update to the systems that surface articles in Discover, the content feed in the Google app. Google said it aims to show more locally relevant content from sites based in the reader's country, reduce sensational and clickbait content, and surface more in-depth, original, and timely content from sites with expertise in a topic.

It started with English-language users in the US. The March 2026 Core Update followed, running from March 27 to April 8 with no companion blog post, and a May 2026 core update began on May 21.

For all of these, Google's standing guidance applies: core updates are broad, they don't target specific sites or pages, and a drop doesn't necessarily mean a page did something wrong.

To act on this era, use our AEO checklist for getting cited by AI.

Inside the black box

Two events opened up mechanisms Google had kept private for years: sworn testimony from Google search executives in the US Department of Justice antitrust trial, which began in 2023, and a May 2024 leak of internal Content Warehouse API documentation. iPullRank's analysis counted 2,596 modules and 14,014 attributes in the leaked documents.

Google confirmed the documents were authentic but cautioned against "making inaccurate assumptions about Search based on out-of-context, outdated, or incomplete information." That caution matters: the documents name attributes, but they don't show how any of them is weighted, or whether each one is used in ranking at all. Read anything below that rests on the leak as informed interpretation, not as a confirmed ranking factor.

The modern link-building playbook

Nearly every classic SEO tactic maps to a specific update that made it stop working, and to what replaced it. This is the direct lineage:

The throughline across 28 years is consistency, not caution: links, reviews, and mentions that would exist whether or not Google was watching survive every update. Links built solely to move a ranking signal are exactly the pattern each new classifier is trained to find.

For ways to earn links that don't depend on pitching, see how to get backlinks without outreach.

Every major dated update referenced above, in order. Names marked "industry name" were coined by SEOs, not by Google.

Where this leaves link building

The mechanism never changed. A link is still one site vouching for another. What changed is how precisely Google can tell a genuine vouch from a manufactured one.

Every era in this history ends the same way: a signal gets exploited, a classifier learns the exploit's footprint, and the sites still standing are the ones that were doing the real thing all along: publishing useful content, earning citations from relevant, independent sources, and building a brand a search engine (and a person) can verify.

That is the same standard non-reciprocal link exchange is built around: verified metrics instead of self-reported screenshots, topically relevant partners instead of anyone who'll swap, and no reciprocal footprint for a future classifier to find.