History of SEO Terms on Wikipedia

Part one

The changing of the guard

Naming decisions record what a term meant. Usage records whether anyone still says it. Measuring each phrase at density per 1,000 words — never raw counts, since the corpus grows fivefold — gives every piece of trade vocabulary a full 23-year curve. The clearest reading is peak year, and how much of that peak survives today.

The drift has a direction, and it is not a rebrand. The vocabulary that entered with the field described what a practitioner does to a page: meta keywords, keyword density, title tags, cloaking, doorway pages. The vocabulary of the 2020s describes what a system does: machine learning, artificial intelligence, large language models, structured data. The subject of the sentence changed.

Every phrase, one curve each

The phrase “keyword density” peaked in 2010 and has lost about four-fifths of its frequency since — the single most-asked question about this dataset, and the answer is a date. But the pattern above it is sharper. Meta keywords, cloaking, doorway pages and page rank all peak in 2004–2006, the corpus’s earliest years, and never recover. The vocabulary that arrived with the field is the vocabulary that died.

Two spelling cases are worth separating. pagerank as one word is the largest early term in the whole set, peaking at 1.387 per 1,000 words in 2006, and retains 46%. page rank as two words retains 10%. The proper noun survived; spelling it out as a generic technique did not. Likewise serps retains 36% while serp singular retains 51% — the plural was jargon, the singular became a noun.

Social media is the largest term in the entire set by an order of magnitude, peaking at 2.684 per 1,000 words. And the terms peaking in 2025 and 2026 are the live vocabulary: artificial intelligence, machine learning, user experience, structured data, large language. Two of them — large language and core web vitals — do not appear anywhere in the corpus before 2022.

Below a third of peak a phrase has effectively left the working vocabulary, even though it still appears in historical passages. That 35% line is a presentation choice, stated so you can move it: the full ordering runs continuously from 5% to 100%.

Each phrase in the two tables below links to its entry in our SEO glossary, which shows these dates beside its definition.

Part two

Replacements and arrivals

Trade language rarely just dies — it gets replaced. Tracking the old form and the new form of the same idea through the same corpus puts a date on the moment the industry’s language actually turned over.

The hyphen died in 2011. “email marketing” overtook “e-mail marketing” that year and never fell behind again — a spelling reform you can date to the year. The bigger renaming took longer: digital marketing overtook internet marketing in 2016, the field’s older name for itself giving way in place.

The other two panels show the limit of the record, honestly drawn. website already led web site when the corpus opens in 2004, and web crawler already led search engine spiders — those races were decided before our first snapshot, so the chart can only say “already over,” not when.

When Wikipedia let each phrase in

Wikipedia does not coin language; it admits it. A phrase’s first appearance in this corpus is an acceptance date — the year the encyclopedia judged the term established enough to use — and lining those dates up shows acceptance arriving in waves, not a trickle.

The striking feature is the gap: nothing new got in for seven years. Of the 27 still-current phrases on the timeline, 23 arrive by 2008 — then from 2009 through 2015, not one new phrase enters the corpus. Admission resumes with mobile-first (2016) and general data protection regulation (2017), and the AI-era vocabulary lands as a wave in 2022. The freeze is a claim about this study’s vocabulary — the tracked list plus the discovered risers — not about every phrase in Wikipedia, but it matches the acceptance thesis exactly: an encyclopedia admits a field’s language in bursts, when the language has already settled elsewhere.

How trade language dies — 15 phrases the data surfaced

The list above was chosen by hand. This one was not: every two- and three-word phrase in the corpus was compared between 2004–2009 and 2021–2026, and these are the ones that fell hardest — each appearing in at least eight articles early on, so no single editor’s habit can masquerade as industry-wide jargon. Read together, they show that trade language dies three distinct ways.

Compression. Robots exclusion standard keeps under a fifth of its old frequency; world wide web and markup languageroughly a third. Nobody writes those out any more — they write robots.txt, the web, HTML. The concept survives; the words collapse into an abbreviation.

Product death. Yahoo search marketing, open directory project, ask jeeves, live search — the jargon died with the products it named. And the three-name reflex of the era, google yahoo msn, is nearly gone: naming a search engine used to mean naming three, and now it means naming one.

The trade’s own words, fading in place. Internet marketing — the field’s older name for itself — keeps a third of its former use. And search engine spiders, the exact phrase Sherwoodseo created as a redirect at 23:45 on the founding night, keeps under a third. The jargon he wired into the encyclopedia is now fading out of it.

And sometimes the trade’s word wins. The most formal name in the early corpus, unsolicited commercial e-mail, keeps 2% of its 2004 frequency — the encyclopedia now just writes spam, the word the trade always used. Hyper text markup language spelled out is gone entirely, last seen 2019. Compression wins even against formality.

Part five

Where the sources go

A citation is a link out of the encyclopedia. Tracked over 23 years, those links tell a second story about search-marketing sources: most of them do not last. The same corpus that retired the founding vocabulary also retired the founding citations — language and sources turn over together.

More than half of every domain the encyclopedia ever cited on these topics has since been dropped, and the typical source survives just 4 of 23 years. But the attrition is not uniform — it depends heavily on when a source first appeared.

The earlier a source arrived, the faster it left

Every cohort is tracked over the same eight-year window, so the comparison is fair. Of the domains first cited in 2006, fewer than half survived even a single year, and 30.1% lasted the eight. The 2018 cohort held 55.4% over the same span. Stretch the 2006 line all the way to today and only 25.3% remain. Early search-marketing citations were disposable in a way later ones are not — the blog posts, forum threads and defunct tools of the mid-2000s, cited once and replaced.

The encyclopedia stopped talking to itself

External citations nearly tripled — 13.6 to 36.6 per 1,000 words — while internal wikilinks fell by a third. In 2004 these articles linked inward 6.14 times for every outward citation; by 2026 that ratio is 1.55. The subject went from something Wikipedia explained mostly by cross-reference to something it documented from outside sources — the same shift, measured a third way.

Part six

Who gets cited

A reference work shows what it accepts by what it cites. Start with the first bibliography the field ever had — and what became of it.

The launch article had no sources at all

The 24 February 2003 revision carries zero external links. It was written from the author’s own knowledge, including the claim about when the practice began. The first seven citations appear in the 2004 snapshot, and they are the opening reading list for the whole discipline.

6 of the 7 domains still answer. Only 2 are still operating as themselves. That gap is the whole point of checking. A link checker counting HTTP 200s would report six survivors: realseo.com returns a hosting-provider placeholder, seo-glossary.com returns 163 empty bytes, and threadwatch.org redirects into the company that absorbed it. Only searchenginewatch.com — publishing same-day news when we checked — and toprank.blogspot.com are still doing what they were cited for.

The one genuinely dead entry is seopages.com: its name no longer resolves in DNS at all. And the most durable is the one nobody would call a citation — google.com/webmasters/seo.html still resolves, twenty-two years later, by redirecting to Google’s current SEO starter guide. The platform kept its content addressable across two decades of reorganisation; most of the independent trade did not.

Concentration across the whole period

One hypothesis this study set out to test was that large platforms have come to dominate these pages. The data does not support it, and points the other way.

Between 2004 and 2026 the number of distinct domains cited grew from 363 to 3,685 — a 10× widening — while the top-10 share moved from 16% to 15.8%. By the Herfindahl–Hirschman index these pages are measurably less concentrated now than in 2004.

What changed is who

Every citation is classified against a curated list of platforms, publishers, vendors, academic and government domains. Roughly 78.2% of citations fall outside that list, so the shares below are quoted against all citations rather than against the classified subset — the honest denominator, and the reason this is two series rather than a pie.

Platform citations peak at 12.3% in 2005 and fall to 7%. Publisher citations rise from 4.5% to 12.2%. They cross in 2014, and publishers never fall behind again.

google.com was the single most-cited domain in 2004, at 4.9% of all citations. By 2026 it has dropped out of the top four entirely, displaced by nytimes.com. The press replaced the platform.

The encyclopedia did not stop covering search. It stopped leaning on the industry’s account of itself — including Google’s — and moved to journalism about it. Acceptance, in a reference work, means being written about rather than being the one writing.

Update · 24 September 2026

Second edition, September 2026

The first edition measured 36 phrases chosen by hand. The second grows the corpus so that our SEO glossary can date its own terms by the same method: 233 of its 640 terms now carry a date, 185 of them measured on this edition.

What changed

402 articles added. We searched Wikipedia for the glossary terms the first edition used too rarely to date. A candidate article joined the corpus only if at least two articles already in it link to it, or if Wikipedia files it under search, online marketing and advertising, analytics, reviews, e-commerce, maps and location, AI, or web design and performance, and under no category about entertainment or fiction, such as film, music, sports or video games.

The same yearly snapshots. Each added article was captured the way the first edition’s were: the last revision of each year, from 2004 or the article’s first year, through 2026. The 2026 snapshot is pinned to the moment the first edition took its last one, on 6 August 2026.

Each article counted once. Rebuilding the corpus showed that the first edition stored 8 articles twice, under names that differ only in the case of their first letter: Landing page, Online advertising, Robots.txt, Search engine, Search engine marketing, Search engine optimization, Search engine results page and Web analytics. Both copies hold the same revision, so the first edition counted that text twice, 176 snapshots and 308,749 words of it. This edition counts each article once. It also leaves out 8 articles that met the selection rule only because of that double count: Heat map, Mobile local search, Real-time business intelligence, .htaccess, Multimedia search, Timeline of web search engines, Art Technology Group and FAQ.

Off-topic articles left out. Checking every title in both editions against the categories Wikipedia files it under found 2 articles off the study’s subject, both in the first edition’s corpus. Danny Sullivan is the biography of the IndyCar driver, collected in place of the search journalist of the same name: 23 yearly snapshots and 23,165 words. John Mueller is the biography of the political scientist, collected in place of the Google search advocate of the same name: 21 yearly snapshots and 8,465 words. The figures published above, measured on the first edition, include them; this edition leaves them out.

The second edition holds 757 articles, 12,492 yearly snapshots and 12,307,889 words from 2004 to 2026: the first edition’s 355 on-topic articles, each counted once, plus the 402 added.

The original figures stand

The study’s 36 phrases keep the figures published above, measured on the first edition. Measured again with each article counted once and the off-topic articles left out, all 36 keep their first-seen year and 32 of 36 their peak year; retention moves by at most 18 points (paid links), and 35 of 36 stay in the same table, faded or enduring. On the whole second-edition corpus, where the added articles use some phrases in new contexts, 26 of 36 keep their first-seen year, 25 of 36 their peak year and 30 of 36 their table. Those readings show how much the figures depend on the corpus; they are not corrections. The figures above describe the first edition’s corpus, and they stand as published.

The glossary’s 197 other dated terms

Beyond the study’s own 36, the glossary dates 197 terms by the same arithmetic and evidence rules, 185 of them on this edition. The other 12, marked in the table, keep the date the first edition gives them: the added articles use those words in other senses, so on the wider corpus their dates would measure the wrong meaning. Each term links to its glossary entry, which shows the same dates. Retained is the share of a term’s peak use that survives in 2026, banded the way the glossary bands it.