There is no single algorithm to update
The phrase "the algorithm" implies one program with one set of rules, periodically rewritten. That has not been true of any large search engine for a very long time, if it ever was. What exists is a stack of independent systems each contributing to a final ordering: text matching, link analysis, neural systems that map queries and documents into a shared representation, freshness scoring, locality, deduplication, site diversity limits, and a large body of spam detection running alongside all of it.
An "update" is therefore a change to one of those systems, or a new one, or the retirement of an old one. Google's Search Status Dashboard reflects this, classifying changes into recurring types — core updates, spam updates, reviews updates and Discover updates — each with a start date, an end date and a duration in days.
That is a recent way of talking about it. For most of search's history, updates were neither announced nor named by the search engine at all. They were noticed from outside, by people watching rankings move, and named by those observers. Understanding the history means understanding that shift from folklore to changelog.
Before names: the Google Dance
In Google's early years the index was rebuilt in batches. A crawl would run, PageRank would be recomputed over the resulting link graph, and the new index would be pushed out to data centres over several days. Because the rollout was not instantaneous, the same query asked twice in one afternoon could return different results depending on which machine answered. Observers called it the Google Dance, and it happened roughly monthly.
Two things follow. Early updates were events with edges — a before and an after that everyone could see. And a page discovered by the crawler could wait weeks to appear in results, because it had to wait for the next batch.
That ended with Caffeine, an indexing infrastructure rather than a ranking change, announced in August 2009 and live in June 2010. Caffeine replaced batch rebuilds with continuous incremental indexing: pages could be crawled, processed and made searchable individually. Google described the result as an index around 50% fresher.
Caffeine is the hinge in this story. Once indexing was continuous, ranking changes could roll out gradually and partially rather than all at once — which is why modern updates take a fortnight rather than a night, and why they became harder to observe from outside at precisely the point Google started announcing them.
Florida, and the era of nicknames
The first Google update to become famous was Florida, in November 2003. It was aimed at ranking manipulation, moved a very large number of results, and landed weeks before Christmas, which is why it is remembered so vividly. It was named not by Google but by the webmaster forums watching it happen, in a convention borrowed from conference locations.
Similarly nicknamed changes followed — Austin in January 2004, Brandy in February 2004 with a large index expansion, Big Daddy from December 2005 into March 2006 as an infrastructure change. Universal Search, on 16 May 2007, was a different kind of event: rather than reordering web links, it merged images, video, news, books and local results into a single ranked list — the direct ancestor of the modern results page.
The pattern recurs and is worth naming. Something changed; outsiders detected it, quantified it badly, and named it; Google confirmed little or nothing; the nickname became the historical record. Much of what is written about pre-2011 updates is reconstruction rather than documentation.
Panda and Penguin: content and links
Two updates from the early 2010s changed how search engines are talked about, partly because they were the first Google both confirmed and quantified.
Panda launched in February 2011, rolling out globally that April, and demoted sites with thin, duplicated or low-value content. Google said it affected around 12% of all search results — an unusually specific public figure. It is named after Navneet Panda, the engineer whose work made it feasible. It was notable for being site-wide: rather than scoring pages in isolation, it computed a modification factor for a whole site. Panda entered Google's core ranking systems in 2015.
Penguin launched on 24 April 2012, aimed at link spam — bought links, link farms, automated link networks. Google said it affected around 3.1% of English-language queries, with higher figures in more heavily spammed languages. Early versions refreshed only periodically, so a site affected by one refresh stayed affected until the next, sometimes for months. On 23 September 2016 Google announced that Penguin had become part of the core algorithm, updated in real time, and now devalued the offending links rather than demoting whole sites.
Both are consequences of PageRank's founding assumption: that a link is an editorial endorsement. Once that assumption was widely known to drive rankings, it stopped being reliably true, and much of the following decade went on building systems to detect where it had broken.
From strings to meanings: Hummingbird, RankBrain, BERT
The second strand of update history is not about quality at all. It is about the engine understanding what a query means rather than which documents contain its words.
Hummingbird, August 2013, is described in Google's own ranking systems documentation as "a major improvement to our overall ranking systems" — a rewrite of how queries were interpreted rather than a filter bolted onto the existing system. It arrived alongside the growth of spoken and conversational queries, which are longer, phrased as questions, and full of words that carry grammar rather than meaning.
RankBrain, announced in October 2015, was Google's first widely publicised use of machine learning in ranking: a system that could return relevant results for queries it had never seen by relating them to concepts it had, rather than by matching keywords. Google later added neural matching, described as understanding representations of concepts in queries and pages and matching them to each other.
BERT is the clearest case of research becoming infrastructure. The paper, "BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding" by Jacob Devlin, Ming-Wei Chang, Kenton Lee and Kristina Toutanova, was published in October 2018. Google began using BERT in Search on 25 October 2019, initially for English queries in the United States; by December 2019 it was applied in more than 70 languages, and by October 2020 Google stated almost every English query was processed by a BERT model. Its practical effect was on small words — prepositions and negations that change a sentence's meaning and that keyword matching discards.
MUM, the Multitask Unified Model, followed in 2021, but Google's documentation describes it as used for specific applications rather than as a general ranking system.
Device, experience and the shift to continuous updating
A separate line of updates concerned the conditions a page is delivered under rather than its content. The mobile-friendly update of 21 April 2015, nicknamed Mobilegeddon in advance by an industry expecting more upheaval than it got, made mobile usability a ranking input on mobile searches. Page loading and stability metrics followed under the page experience heading.
These were inputs among many, not overrides, and the gap between how they were reported and what they did was wide. Mobilegeddon is a case study in how update coverage works: the name was coined before the event, the anticipation exceeded the effect, and the name outlived both.
From 2018 Google began announcing broad ranking changes itself, under the plain description "core update". Nicknames continued from outside — the August 2018 core update is widely called "Medic" because health and medical sites featured prominently among those affected — but Google's own naming has been resolutely boring ever since, deliberately so. A month and a type is not a story.
The modern rhythm: dated core updates
Google now publishes each ranking update on its Search Status Dashboard with a start time, an end time and a duration. The recent record reads like a maintenance log, which is the point.
- March 2024 core update — 5 March to 19 April 2024, forty-five days, the longest core update Google has documented. It ran alongside a spam update of 5–20 March, and with it Google folded the helpful content system, announced separately in 2022, into its core ranking systems rather than running it as a distinct filter.
- 2024 — August core update, 15 August to 3 September; November core update, 11 November to 4 December; December core update, 12 to 18 December, followed by a spam update from 19 to 26 December.
- 2025 — March core update, 13 to 27 March; June core update, 30 June to 17 July; August spam update, 26 August to 22 September, the longest spam update on the dashboard; December core update, 11 to 29 December.
- 2026 — February Discover update, 5 to 27 February; March spam update, 24 to 25 March; March core update, 27 March to 8 April; May core update, 21 May to 2 June; June spam update, 24 to 26 June. An August 2026 spam update began on 18 August and was in progress as this page was written.
Two observations. Core updates now run for roughly two weeks, so anyone dating one to a single day is describing something other than what Google published. And the pace is regular: broadly two to four core updates a year, punctuated by spam updates.
Retired names, and who else updates anything
Google's ranking systems documentation keeps a list of retired systems, which is the most honest statement available about what the famous names mean today. Panda "became part of our core ranking systems in 2015". Penguin "was integrated into our core ranking systems in 2016". Hummingbird is listed as a 2013 improvement superseded by continued evolution. The helpful content system became part of core ranking in March 2024. These are not switched-off ideas but ideas absorbed into a larger system that no longer has a separate name for them. Writing that a site has been "hit by Panda" in 2026 describes something that stopped existing as a distinct entity eleven years ago.
It is also worth noticing how one-sided this history is. Almost every named update belongs to Google. Microsoft changes Bing continually and writes about some of it on its webmaster blog, but publishes no changelog of named ranking releases, and no folklore of Bing nicknames ever grew up because the audience watching was too small to sustain one. The history of search algorithm updates is largely the history of Google's updates — a market-share fact rather than an engineering one.
The consequence for everyone else is structural. Engines that do not crawl inherit changes they did not make: Yahoo has served Bing's results since the search alliance completed, DuckDuckGo blends Bing's index with its own sources, Startpage resells Google. When Bing changes its ranking, several visibly independent engines change with it, on a date none of them announced. Only engines running their own crawlers — Google, Bing, Yandex, Baidu, Mojeek and a handful of others — are updating anything at all.
Frequently asked questions
What is a Google core update?
A core update is a broad change to Google's main ranking systems, announced on its Search Status Dashboard with a start date, an end date and a duration. Recent ones have taken between six days and forty-five days to roll out. Google distinguishes them from spam updates, which change the systems that detect policy violations, and from narrower updates to specific surfaces such as Discover.
What did the Google Panda update do?
Panda, launched in February 2011 and rolled out globally that April, demoted sites carrying thin, duplicated or low-value content. Google said it affected around 12% of all search results. It applied a site-wide factor rather than judging pages in isolation, and it is named after the engineer Navneet Panda. It was folded into Google's core ranking systems in 2015 and no longer exists separately.
What did the Penguin update target?
Link spam. Penguin launched on 24 April 2012 and affected around 3.1% of English-language queries, targeting bought links, link farms and automated link networks. Early versions refreshed only periodically, so effects persisted between refreshes. On 23 September 2016 Google announced Penguin had become part of the core algorithm, updating in real time and devaluing individual links rather than demoting entire sites.
What was the Florida update?
Florida was a Google ranking change in November 2003, aimed at manipulation, that moved a very large number of results shortly before the Christmas trading season. It is remembered as the first update to become a public event. Google neither named nor described it at the time — the name came from webmaster forums, following a convention of naming updates after conference locations.
When did Google start using BERT in search?
On 25 October 2019, initially for English-language queries in the United States. The underlying research paper was published in October 2018 by Google researchers Jacob Devlin, Ming-Wei Chang, Kenton Lee and Kristina Toutanova. By December 2019 BERT was in use across more than 70 languages, and by October 2020 Google stated that almost every English query was processed by a BERT model.
Why do algorithm updates take weeks to roll out now?
Because indexing became continuous. Until the Caffeine infrastructure went live in June 2010, Google rebuilt its index in batches and pushed the result out at once — the effect observers called the Google Dance. Continuous incremental indexing lets changes be applied gradually and partially instead, which is why Google's dashboard records core updates lasting from about a week to forty-five days.
Does Bing have named algorithm updates?
Not in the way Google does. Microsoft changes Bing's ranking continually and writes about some changes on its webmaster blog, but it publishes no dated changelog of named ranking releases, and no folklore of nicknames grew up around it. Named updates are largely a Google phenomenon, which reflects the size of the audience watching rather than the pace of engineering.
Do algorithm updates affect search engines that resell results?
Yes, and usually without announcement. Yahoo has served Bing's web results since the Microsoft search alliance completed, DuckDuckGo blends Bing's index with other sources, and Startpage resells Google. When the underlying index changes its ranking, every engine built on it changes too, on a date none of them chose. Only engines running their own crawlers update their own results.