Thirty years of SEO, and a machine that keeps changing its mind

If Google Tag Manager is a name you keep meeting without quite knowing what it covers, the page on understanding Google Tag Manager lays the groundwork.

In the beginning there was a list of files

In 1990, the first search engine did not search web pages, because there were none yet. It was called Archie, and it indexed file names on servers. Not their content. Their names. That was already enormous.

Four years later the web exists and two schools compete. On one side Yahoo, a directory maintained by hand by humans filing sites into folders, like a library. On the other, engines that read the text of pages and file them on their own.

The human directory lost, obviously. Not because it was worse, but because the web grew faster than any human capacity to file anything.

This is when the craft was born, with a name that never really changed. Who coined it exactly is disputed and I have never found a source that settles it. It hardly matters: the principle was already there. Write a page so that a machine understands what it is about.

Google arrives and invents reputation

In 1998, two Stanford students had a simple, brutal idea: rather than judging a page on what it says about itself, judge it on what other pages say about it.

That is PageRank. A link becomes a vote. A page receiving many links from pages that are themselves well linked rises. The web starts to look like an enormous network of recommendations.

That idea is still alive today, and it matters to understand why: almost everything a modern search engine does is a layer placed on top of that 1998 intuition. We added meaning, context, security, speed, artificial intelligence. We never removed reputation.

The golden age of cheating

Between 2000 and 2010, organic search looked like a western.

The rules were known, the defences weak, and imagination limitless. People repeated a keyword forty times on a page. They wrote white text on a white background to hide it from readers and show it to robots. They bought links by the thousand in directories nobody ever read.

It worked. That is the hard part to admit, but it worked very well.

On 16 November 2003, Google launched the Florida update, right before the holiday season. Thousands of retailers watched their traffic vanish at the worst possible moment of the year. The message was clear: the rules can change overnight, and nobody will be warned.

Twenty three years later, that remains the only SEO lesson never once contradicted.

Google learns to read

The following decade is a series of blows to the shortcut industry.

Panda, on 24 February 2011, went after content farms, the sites producing a thousand bland articles a month to catch every possible query. Penguin, on 24 April 2012, went after bought links. Hummingbird, on 22 August 2013, taught the engine to understand the meaning of a question rather than the words composing it.

Then came Mobilegeddon on 21 April 2015, demoting sites unreadable on a phone, and RankBrain in October of the same year, the first machine learning brick inside ranking. BERT, in October 2019, taught the engine the small words that change everything: prepositions, negation, word order.

One way to summarise those twenty years: Google spent them catching up with the reading level of an attentive human. Every update made one trick obsolete and one obvious thing relevant.

Then it learns to be demanding

From 2021 the subject shifts. It is no longer only about understanding the page, but about judging the experience it offers.

Core Web Vitals arrive on 15 June 2021 on mobile, then in November on desktop. The engine starts measuring your display speed, your layout stability, your responsiveness to a click. A slow site becomes a penalised site.

On 25 August 2022, the Helpful Content Update pushed the logic to its conclusion and asked a disarmingly simple question: was this page written to help someone, or to catch a search engine?

Many sites did not like the answer.

2024 to 2026, the machine starts answering

And then the ground moved for good.

In May 2024, Google launched AI Overviews in the United States: a written summary placed above the links. AI Mode, a full conversation with the engine, followed. In France, both have been live since 22 July 2026.

What that changes fits in one image. For twenty five years a search engine was a signpost: it showed you where to go. It is becoming a guide that answers in your place, and sometimes mentions where its answer came from.

The figures are unambiguous. The July 2025 Pew Research study, run on 900 US adults and 68,879 searches, measured a click on a classic result in 8% of visits when an AI summary is present, against 15% without. Seer Interactive observed a 61% drop in click through rate on general questions.

But there is a nuance the alarmist headlines systematically forget. Transactional searches, the ones where someone wants to buy, book or get in touch, are far less affected. People who want to understand stop at the answer. People who want to act still click.

If your site sells something, you are less exposed than you think. If your site lives on informational traffic and advertising, you are more so.

So what do you actually do

Here is what I answer when asked, and none of it is a trick or a secret.

Stop chasing the latest novelty. Every tactic born from an update dies with the next one. By the time you have implemented the technique of the month, three thousand people will have commented on it on LinkedIn, so it will already be saturated.

Look after what you already have. On the large majority of sites I take over, there is more to gain from fixing existing pages than from creating new ones. Pages Google does not crawl, two pages targeting the same query and cancelling each other out, content sitting dormant without a single internal link.

Write passages that stand on their own. A generative engine does not quote a page, it quotes a paragraph. If your paragraph opens with “as we saw above”, it is unusable. That is a change in writing, not a tool to buy.

Measure, even badly. Since June 2026 Search Console gives your impressions inside AI answers, without the clicks or the queries. It is frustrating and incomplete. It still beats nothing, and I explain in a dedicated article how to measure your visibility in AI answers with what exists today.

And accept not knowing everything. Nobody, including the people selling courses on the subject, knows exactly how an AI Overview picks its sources. Anyone telling you otherwise is selling you something.

What never changed

Thirty years, a dozen major turning points, thousands of silent adjustments. And yet one constant runs through the whole story, from the first Yahoo directory to Google AI Mode.

The engine is trying to guess what a human really wanted, and to give them the best available answer.

That is all. Every update since 1998 has only brought the machine closer to that goal, closing the door on cheats along the way. The day an engine can perfectly judge the usefulness of a page, SEO will not be dead. It will simply have gone back to what it should always have been: doing something well, and saying it clearly.

Until that day there is work to do. The jungle is loud, the rules change, and plenty of people shout wrong things very confidently.

If you want clarity on your own site, that is exactly what the Organic Visibility Diagnostic 90 days does: a state of play, a list of actions ranked by impact, and a roadmap. And if you would rather understand how I work first, the freelance SEO consultant page covers it.