History of SEO: from PageRank to AI Overviews

Every big Google change since 1998 rewarded one thing and punished another. Read the pattern and AI search stops looking like a mystery.

Google sign on Charleston Road at the company campus in Mountain View, California
Photo: Dietmar Rabich, CC BY-SA 4.0, via Wikimedia Commons

The history of SEO is a 28-year contest between Google and people gaming it. PageRank (1998) made links the currency, Florida (2003), Panda (2011) and Penguin (2012) punished cheap links and thin pages, and the 2022 helpful content system and the 2024 AI Overviews moved the reward from ranking a page to being the source an answer quotes.

Key takeaways
  • PageRank (1998) made links the currency of search, with a damping factor of 0.85.
  • Panda hit almost 12% of results in February 2011; Penguin hit 3.1% of English queries in April 2012.
  • 1.3 to 52: J.C. Penney's average rank across 59 terms after Google caught its paid links in February 2011.
  • 45% less low-quality, unoriginal content after the March 2024 core update, by Google's count.
  • 8% vs 15%: click rate on Google pages with and without an AI summary (Pew, 2025).
  • Statistics, quotes and citations lift visibility in AI answers by about 30 to 40% (GEO study, 2024).

Each update had a clear target, a measured share of queries hit, and a lesson that still holds. I have laid the history out in dates, with what each change rewarded, because the same pattern now runs inside ChatGPT, Perplexity and Google's AI Mode.

If you understand why Panda hit content farms, you already understand most of what answer engine optimisation (AEO) and generative engine optimisation (GEO) ask of you.

24 millionpages in Google's 1998 prototype index
12%of queries hit by Panda, Feb 2011
3.1%of English queries hit by Penguin, Apr 2012
8% vs 15%click rate with vs without an AI summary (Pew, 2025)

How did SEO start, and what did PageRank change?

SEO started as keyword matching: early engines like AltaVista ranked pages largely on what the page said about itself, so repeating a word worked.

The 1998 paper

In 1998 Sergey Brin and Lawrence Page published the design of Google, a prototype with a full-text and link database of at least 24 million pages, and its key idea, PageRank, scored a page by the number and quality of other pages linking to it.

PageRank treated a link as a vote, weighted by the voter's own score, with a damping factor the paper sets at 0.85 (the chance a random surfer keeps clicking rather than jumping to a random page). That one change moved the game off the page.

Whoever controlled links controlled rankings, and an industry grew to feed it:

  • link directories;
  • reciprocal link swaps;
  • paid links placed on unrelated sites.

Almost every update for the next 15 years was Google closing a loophole in its own currency.

PageRank: a link is a weighted voteHigh-score pageHigh-score pageMid-score pagePage CPage Aranks highone weak link: ranks lowEach link passes 0.85 x (linker's score / its number of outlinks)Fake "votes" were theshortcut Penguin killed
The idea in Brin and Page (1998): a page matters if pages that matter link to it. Paid and farmed links were the shortcut every later link update targeted.
The William Gates Computer Science Building at Stanford University
Stanford's Gates Computer Science Building. Brin and Page built the Google prototype as Stanford graduate students, and the 1998 paper came out of this department.Photo: Frank Schulenburg, CC BY-SA 4.0, via Wikimedia Commons
The Anatomy of a Large-Scale Hypertextual Web Search EngineSergey Brin, Lawrence Page · Computer Networks, vol. 30 · 1998

The founding paper of Google describes a prototype with a full-text and hyperlink database of at least 24 million pages, and defines PageRank as a recursive score in which a page is important if important pages link to it. It also argued that anchor text, the words inside a link, describe the target page better than the page describes itself. The limit: it is a system design paper from a web of millions of pages, written before link spam existed at scale, so it says nothing about manipulation.

Read the paper (free)

Which Google updates changed SEO the most?

Six updates did most of the shaping: Florida (2003), Panda (2011), Penguin (2012), the 2015 mobile update, the 2022 helpful content system, and the March 2024 core update that folded it into the main algorithm. The table lists them with the date, the target and the share of queries Google or its reporting said was affected.

YearUpdateWhat it targetedWhat it rewarded
1998PageRank (launch)Pages that only described themselvesLinks from important pages
Nov 2003FloridaKeyword stuffing and 1990s on-page tricksPages relevant beyond exact-match words
Jun 2010CaffeineSlow, batch indexingFresh content, indexed in near real time
Feb 2011Panda (12% of queries)Content farms, scrapers, thin and ad-heavy pagesOriginal, substantial pages
Apr 2012Penguin (3.1% of English queries)Bought links, link farms, private blog networksEarned links
Sep 2013HummingbirdLiteral word matchingMeaning of the whole query
Apr 2015Mobile updatePages unusable on phonesMobile-friendly pages
Oct 2015RankBrainUnseen queries (about 15% of searches)Pages that satisfy intent
Oct 2019BERTMisread prepositions and contextNatural-language answers
Aug 2022Helpful contentContent written mainly to rankPeople-first content
Mar 2024Core update + spam policiesScaled content, expired domains, site reputation abuse45% less low-quality, unoriginal content
May 2024AI Overviews (US)The ten blue links as the default answerPages an AI summary chooses to cite
The story

In February 2011, weeks before Panda, The New York Times asked search consultant Doug Pierce of Blue Fountain Media why J.C. Penney sat at or near the top of Google for "dresses", "bedding", "area rugs", "skinny jeans" and even "Samsonite carry on luggage", above Samsonite's own site. Pierce found 2,015 pages linking to Penney's dresses page with phrases like "little black dress", on unrelated sites, one of them about nuclear engineering. The rankings had held through the holiday shopping season. Google's webspam head Matt Cutts confirmed the links broke its guidelines and started a manual action. For "living room furniture", Penney fell from first at 7 p.m. to 68th by 9 p.m. Across 59 terms its average position went from 1.3 on 1 February to 52 by 10 February. Penney said it had not authorised the links, and fired its search firm, SearchDex.

The New York Times (republished by NBC News): The dirty little secrets of search, 13 Feb 2011

What did Florida, Panda and Penguin punish?

Florida, 2003

Florida, in November 2003, was the first update that SEO practitioners felt as a shock: it devalued keyword-stuffed pages, and commentators at the time said it ended the value of 1990s SEO tactics. The lesson was that on-page repetition would no longer buy relevance.

Panda, 2011

Panda, launched on 23 to 24 February 2011 and named after engineer Navneet Panda, hit almost 12% of search results. It went after content farms, scrapers and pages thin on substance or heavy on ads, and it could demote a whole site, not just a page.

For two years it re-ran roughly monthly; in March 2013 Google said it would be built into the algorithm, and it became part of the core ranking system by 2015. The modern idea of site quality starts here.

Penguin, 2012

Penguin, launched on 24 April 2012, affected about 3.1% of English queries and went after manipulated links:

  • bought links;
  • link farms and automated links;
  • private blog networks.

Penguin 4.0, announced on 23 September 2016, made it real-time and changed the penalty: bad links were now discounted rather than sinking the whole site.

Between them, Panda and Penguin forced SEO to earn both of its inputs, content and links, instead of manufacturing them.
UpdateShortcut it killedPenalty style
Florida (2003)Keyword stuffingDevalued pages
Panda (2011)Thin and farmed contentWhole-site demotion
Penguin (2012)Bought and farmed linksSite demotion, then (2016) link discounting
1998PageRank2003Florida2011Panda2012Penguin2015Mobile, RankBrain2019BERT2022Helpful content2024AI OverviewsLinks era: who links to youMeaning era: what the query meansQuality era: is the page worth itAnswer era: will a model quote you
Four eras of search optimisation. Each began with an update that devalued the previous era's shortcut.

How did Google move from keywords to meaning?

From 2013 Google stopped matching words and started modelling what a query means. Three systems did it:

  • Hummingbird (rolled out around August 2013, announced 26 September 2013) rewrote how queries were interpreted;
  • RankBrain (announced 26 October 2015) used machine learning on the roughly 15% of queries Google had never seen before;
  • BERT (October 2019, extended to 70 languages by December) read words in the context of the words around them.

The practical result for writers was that synonyms, entities and full questions began to matter more than an exact keyword in the title. A page that answered "can I ride to Leh in October" well could rank for dozens of phrasings it never used. This is also the bridge to AI answers: the same language models that interpret queries now write the response.

The mobile update

Mobile changed the other half of the equation. Google warned site owners on 19 January 2015 and began demoting mobile-unfriendly pages from 21 April 2015. In India, where most people meet the internet on a phone, that update decided which publishers survived the next decade.

Google data centre buildings at The Dalles, Oregon
Google's data centre at The Dalles, Oregon. Caffeine (2010) and later language models changed what this hardware computes for every query.Photo: Visitor7, CC BY-SA 3.0, via Wikimedia Commons

What did the helpful content system and the 2024 core update do?

The helpful content update, which rolled out from 25 August to 9 September 2022, targeted content created mainly to rank rather than to help a reader. It was Google's direct answer to the wave of templated "best X" pages and, soon after, AI-written filler.

The March 2024 core update, announced on 5 March 2024, merged that system into core ranking and added three spam policies:

  1. Scaled content abuse: mass-produced pages, by humans or AI;
  2. Expired domain abuse: buying old domains to borrow their authority;
  3. Site reputation abuse: third-party pages riding a strong host site.

Google forecast a 40% cut in low-quality, unoriginal results; when the rollout finished on 19 April 2024 it reported 45%. Note what the policy did not ban: AI-written content as such. It banned scale without value.

How have AI Overviews changed clicks from Google?

AI Overviews cut clicks to ordinary results sharply. Google launched them for all US users on 14 May 2024 and said it expected to reach over a billion people by the end of that year.

A Pew Research Center study of real browsing found users clicked a traditional result on 8% of visits to pages with an AI summary, against 15% without one.

Ahrefs measured the same effect from the publisher side. Its first study (April 2025) found AI Overviews cut the click-through rate of the top-ranked page by about 34.5%; its February 2026 update, comparing December 2023 with December 2025 across 300,000 keywords, put the drop at 58%.

Keyword typePosition-one CTR, Dec 2023Position-one CTR, Dec 2025
With an AI Overview0.0730.016
Informational, no Overview0.0760.039

Position one on an AI Overview keyword fell from a click-through rate of 0.073 to 0.016, while comparable informational keywords without an Overview fell from 0.076 to 0.039.

Share of Google visits that led to a click, with and without an AI summary
See the data as a table
Page with AI summary (%)Page without AI summary (%)
Clicked a traditional result815
Clicked a link inside the summary1
Ended the browsing session2616

Source: Pew Research Center, browsing data of 900 US adults, March 2025

Google users are less likely to click on links when an AI summary appears in the resultsPew Research Center, short read · July 2025

Pew tracked the actual browsing of 900 US adults in March 2025, covering 68,879 unique Google searches, and re-ran the searches to capture the result pages. 18% of searches produced an AI summary, and 58% of people saw at least one. Clicks on traditional links fell from 15% of visits to 8% when a summary appeared, and only 1% of visits clicked a link inside the summary. Summaries appeared on 8% of one- or two-word searches but 53% of searches with ten or more words. Wikipedia, YouTube and Reddit were the most cited sources. The limit: US users only, one month, and Google only.

Read the report (free)

What does SEO history predict for AEO and GEO?

It predicts that answer engines will reward the same things Panda and Penguin did, verifiable and original substance, and will eventually punish whatever is cheap to fake. The best evidence so far comes from a 2024 study that measured which edits actually make a page more visible inside AI answers.

GEO: Generative Engine OptimizationPranjal Aggarwal et al. · KDD 2024 (arXiv 2311.09735) · 2024

The authors built GEO-bench, 10,000 queries across domains, and tested nine ways of rewriting a source page to see how much of the generated answer it captured. Adding quotations, adding statistics and citing sources raised visibility by roughly 30 to 40% on their position-adjusted word count measure; keyword stuffing did nothing or made it worse. On the live Perplexity engine, quotation addition improved visibility about 22% and keyword stuffing did about 10% worse than baseline. The biggest gains went to lower-ranked pages: citing sources lifted a fifth-ranked source by 115%. The limit: these are controlled rewrites of single pages, and engines change their retrieval often.

Read the paper (free)

Reading GEO against the history

Read that result against the history:

  • Keyword stuffing failed in AI answers the way it failed after Florida.
  • Statistics, quotes and citations work because they are what a model can lift and attribute.
  • Lower-ranked pages gained most, which is the opportunity: an AI answer reads several sources, so it is not winner-takes-all the way position one was.

The Reddit thread in the forum list above asks who has actually got results from AEO. The honest pattern in that discussion and in the data is that brands which already earned mentions on Wikipedia, Reddit, YouTube and review sites show up, and brands that only optimised their own pages mostly do not.

I break down which domains get cited in the domains AI search engines cite most, how the engines pick sources in how AI Overviews and ChatGPT choose sources, and what to do about it in how to get cited by ChatGPT and Perplexity.

Outdoor recreation area at the Google campus on Charleston Road, Mountain View
Google's Mountain View campus in 2022. AI Overviews launched from here to all US users in May 2024.Photo: Dietmar Rabich, CC BY-SA 4.0, via Wikimedia Commons

What should a marketer in India do with this history?

Treat each era's lesson as still live, because none of the old rules were repealed; they were stacked. Links still matter, mobile still matters, helpfulness still matters, and now quotability matters on top.

  1. Earn mentions, not just links. Penguin taught that bought links get discounted. AI engines cite Reddit, YouTube and Wikipedia heavily, so presence in those places is the new link building.
  2. Write the quotable sentence. Open each section with a dense, sourced fact, as the GEO study and Google's own Overviews both favour.
  3. Publish in Indian languages. Google added Hindi to AI Mode in September 2025 and Bengali, Kannada, Malayalam, Marathi, Tamil, Telugu and Urdu in October 2025. Fewer brands compete there.
  4. Measure beyond clicks. With position-one clicks down by more than half on Overview queries, track branded search, citations in AI answers and assisted conversions. See what actually works in AEO and GEO.
  5. Do not mass-produce. Scaled content abuse is a named spam policy since March 2024. Volume made cheap by AI is exactly what the next update will target.

This is the same discipline that runs through the rest of my marketing desk: measure the outcome, not the output.

Questions people ask

When did SEO start?

Search optimisation began in the mid-1990s with the first crawler-based engines, when ranking depended mostly on words on the page. It became a distinct discipline after Google's 1998 PageRank paper made links the main ranking signal.

What was the biggest Google algorithm update?

By measured reach, Panda in February 2011, which Google's reporting put at almost 12% of search results. By long-term effect, Penguin (April 2012) and the March 2024 core update matter as much, because they ended link buying and mass-produced pages as strategies.

Is SEO dead because of AI Overviews?

No, but clicks per ranking are falling. Pew found link clicks fell from 15% to 8% of visits when a summary appeared, and Ahrefs measured a 58% lower top-result click rate by December 2025. Ranking still feeds the sources AI answers cite, so SEO now works partly as a route into the answer.

What is the difference between SEO and GEO?

SEO aims to rank a page in a list of results; GEO aims to have a page's content used and cited inside an AI-generated answer. The 2024 GEO study found statistics, quotations and source citations lift visibility in answers by about 30 to 40%, while keyword stuffing does not.

Does Google penalise AI-written content?

Not for being AI-written. Its March 2024 spam policies target scaled content abuse, meaning many pages made mainly to manipulate rankings, whether a person or a model wrote them.

Sources

  1. Brin and Page: The Anatomy of a Large-Scale Hypertextual Web Search Engine, Computer Networks (1998)
  2. Wikipedia: Timeline of Google Search (accessed Nov 2026)
  3. Wikipedia: Google Panda (accessed Nov 2026)
  4. Wikipedia: Google Penguin (accessed Nov 2026)
  5. David Segal, The New York Times (republished by NBC News): The dirty little secrets of search (2011; source for the J.C. Penney story)
  6. Google Search Central: March 2024 core update and new spam policies (2024)
  7. Google: New ways we're tackling spammy, low-quality content on Search (with April 2024 update) (2024)
  8. Google: Generative AI in Search, AI Overviews launch (2024)
  9. Pew Research Center: Google users are less likely to click on links when an AI summary appears (2025)
  10. Ahrefs: Update, AI Overviews reduce clicks by 58% (2026)
  11. Aggarwal et al.: GEO, Generative Engine Optimization, KDD 2024 (2024)
  12. Google India: New languages in AI Mode and Search Live (2025)
  13. exchange4media: The ad that never loaded (2026)
SEOGoogle algorithm updatesPageRankPandaPenguinAI OverviewsGEO
Sapta

Sapta (Saptarshi Sarkar) is a mechanical engineer turned marketer in Bengaluru and a solo touring rider. He writes the seven Saptaverse blogs. Work with him or read more of Marketblog.

Keep reading in Marketblog

From the other blogs