The Web Has Always Been Wrappers: A 30-Year History of SEO

There is a question I get asked constantly now: has AI changed everything about search? As is my nature, I did the research before treating my instinct as fact. The history of SEO runs clear back to 1994, and it says something more interesting than yes or no.

At the center of every page ever published sits the same thing: a piece of knowledge that the page exists to convey. A product. An answer. An idea. Everything the search industry has built in thirty years — every tag, every protocol, every acronym — is a wrapper around that knowledge, added so machines can interpret what humans already understood.

Every era of search added a wrapper. None of them replaced the page.

The History of SEO: Thirty Years, Six Wrappers

The directory and crawler era, 1994 to 1997. Yahoo began in 1994 as a hand-built directory. AltaVista launched in December 1995 with a full-text index of sixteen million pages and handled three hundred thousand searches on day one. To be interpreted by these machines, pages needed their first wrappers: the title tag, the description, the meta keywords tag. The keywords tag matters to this story because of how it died. It was stuffed and abused into worthlessness, and in September 2009 Google confirmed it ignores the tag entirely. That is the exception that proves the rule of this essay: wrappers that carry real information persist, and wrappers that carry manipulation get discarded.

The link era, 1998 to 2004. Brin and Page's 1998 paper introduced PageRank, and the wrapper moved outside the page: links as endorsement. Your knowledge was now interpreted partly by what others said about it. That wrapper never went away. Third-party authority remains the trust spine of every engine today, including the AI engines deciding which store to cite.

The site era, 2005 to 2009. Google introduced Sitemaps in June 2005, and in November 2006 Google, Yahoo, and Microsoft jointly adopted the open Sitemaps protocol — competitors standardizing a wrapper because everyone benefits when knowledge is legible. Canonical tags followed in February 2009, letting sites declare which version of a page is the real one. Rich snippets arrived in May 2009, reading early semantic markup.

The social and semantic era, 2010 to 2013. Facebook unveiled Open Graph in April 2010: a wrapper telling social platforms how a page should look when shared. Then June 2, 2011: Google, Bing, and Yahoo jointly launched schema.org, with Yandex joining that November. Rival engines again agreed on one vocabulary, this time for what things are. Google's Knowledge Graph followed in May 2012 — "things, not strings" — and the Hummingbird rewrite in 2013 tuned the whole engine toward entities and meaning. This is the era when machines stopped indexing text and started understanding things: a product, with a price, in stock, from a brand.

The entity and feed era, 2013 to 2020. JSON-LD became a W3C Recommendation in January 2014, and by 2016 Google's documentation recommended it as the preferred structured-data format. Commerce got its own wrapper: Froogle launched back in 2002, evolved into Google Shopping, and in April 2020 free product listings returned — whole catalogs, machine-readable by design.

The answer era, 2023 to now. Google announced generative search in May 2023 and rolled AI Overviews across the US in May 2024. The generative engine optimization paper that named GEO was published at KDD in August 2024. The llms.txt file was proposed in September 2024 — a candidate wrapper, not yet a standard, and honesty requires saying no major AI crawler has committed to it. Then the agentic protocols arrived in a rush: Anthropic's Model Context Protocol in November 2024, Google's Agent Payments Protocol in September 2025, ChatGPT's instant checkout with Stripe days later, and WebMCP accepted as a W3C community draft with a Chrome preview by February 2026. Machines now assemble answers from pages and complete purchases from them.

The pattern, four beats

Walk those eras and the same four things happen every time.

1. Each wrapper builds on the ones before it and replaces none of them. Schema.org did not retire the title tag. AI Overviews read the schema, the links, the sitemaps, and the prose, all at once.
   

2. Each starts optional and becomes table stakes. Structured data was an enthusiast's hobby in 2011. Today it is the difference between being parsed and being understood.
   

3. Each rewards early adopters, then pays parity. The stores that adopted feeds early owned Shopping results for years. The advantage always converts to a requirement.
   

4. Each gets refined until the next one arrives. Sitemaps 0.84 became 0.90 became universal. The wrappers under construction today will look the same in hindsight.
   

What this means for the acronym soup

GEO. AEO. Agentic optimization. Each is real, and each is the newest wrapper behaving exactly like every wrapper before it: built on top of everything that already exists, optional today, table stakes tomorrow.

Which is why I call the approach we take Visibility Optimization. The idea is bigger than any tool. It says: solidify each foundational layer, in order, beginning with the innermost one — the idea the page is intended to convey. Then the prose that conveys it. Then the metadata that names it. Then the structured data that formalizes it. Then the third-party authority that corroborates it. Roughly eighty percent of that work is shared across every surface a shopper might ask, and the research keeps bearing that out. A store that has done the layers in order is prepared for each new pillar of AI search, whatever acronym it arrives wearing, because the new pillar is always a new wrapper around the same knowledge.

The newest evidence points the same way. Kevin Indig's analysis of 1.2 million AI answers found that 44.2% of ChatGPT citations come from the first 30% of a page. The machines still read the innermost layers first.

History's verdict is consistent: nobody who skipped a layer kept their visibility through a transition.

The honest close

One thing genuinely does not transfer between eras: measurement. Rank tracking tells you nothing about whether ChatGPT recommends you. Every prior wrapper eventually got its measuring stick, and the current era's stick is still being built. That gap is where the real work of this era is, and anyone who tells you it is already solved is selling you a wrapper of a different kind.

The page endures. The knowledge at its center endures. The work has always been the same: make what you know legible to whatever machine is asking. It was true when the machine was AltaVista. It is true now that the machine talks back.

Sources and dates

All dates verified against primary sources: Google Search Central announcements (meta keywords, 2009; Sitemaps protocol, 2006; canonical tags and rich snippets, 2009; schema.org, 2011), the Stanford PageRank paper (1998), the W3C JSON-LD Recommendation (2014), the KDD 2024 GEO paper (arXiv 2311.09735), Answer.AI's llms.txt proposal (2024), and vendor announcements for the Model Context Protocol, Agent Payments Protocol, and WebMCP (2024–2026).