In 1999 I was optimizing pages for AltaVista, HotBot, Excite, and Yahoo. Google existed, but betting on it looked risky. The smart money said portals would win.
I have been wrong about many things since. That was the first big one, and it taught me the lesson that paid for everything after it: the engine you optimize for today is not the engine you will optimize for tomorrow. The only durable skill is understanding how an engine decides what to show.
Twenty-seven years later, the engine is a language model, and I am watching it learn, in public and at speed, the same lessons Google spent two decades learning. Some it has absorbed already. Some it is failing in ways that feel very familiar to anyone who was there the first time.
What the directory era taught: relevance is a claim, and claims get gamed
Before PageRank mattered, ranking was mostly a text-matching exercise. You said what your page was about, and the engine largely believed you. Keyword stuffing was not a trick, it was the job.
That system died for one reason: self-description does not scale as a trust signal. The moment ranking rewards a declaration, the declaration becomes worthless. Google’s answer was to stop asking pages what they were about and start asking the rest of the web.
LLMs in 2026 are partly back in the directory era. A model deciding which source to cite leans heavily on how content describes itself: structured claims, clear statements, confident framing. That window is open right now, which is why Generative Engine Optimization works. If history repeats, the window closes the way it did in November 2003, when the Florida update wiped out a generation of keyword tricks in a single weekend. The models will trust external corroboration over self-description. Build for that day, not for the loophole.
What PageRank taught: borrowed trust beats declared trust
Links worked because they were expensive. Someone else had to spend their credibility on you. For a few years, that kept ranking roughly honest, and then an entire industry industrialized the spending of fake credibility, and Google spent the next decade clawing it back: Penguin in April 2012, manual actions, the disavow era. I filed my share of reconsideration requests, for sites that deserved it and a few that did not.
The LLM version of this fight has barely started. Citation in a generated answer is the new link, and there is already a market forming around manufacturing citation-worthiness. The models have no public equivalent of Penguin yet. They will need one. When a model’s training or retrieval pipeline starts discounting coordinated citation patterns, a lot of today’s GEO tactics will evaporate overnight. I have seen that movie. I archived it, chapter by chapter, in roughly 700 Hebrew posts since 2004.
What Panda and BERT taught: the engine eventually reads like a human
Panda in February 2011 was the first time Google demoted content at scale for being hollow rather than for being manipulative. On language, Hummingbird in 2013 and RankBrain in 2015 laid the ground, but BERT in 2019 was when parsing meaning, instead of matching words, visibly took over the results page. Each shift bankrupted a tactic and rewarded something that had been true all along: write for the reader who actually has the question.
This is the lesson ChatGPT never had to learn, because it was born reading like a human. Which creates the strange inversion of 2026: content quality is table stakes, and the differentiator has moved to structure. A language model does not rank your page. It dismantles it into statements and decides, claim by claim, what survives into an answer. A page can be excellent prose and still dissolve into nothing citable. That is a new discipline I call statement-level visibility, and it is most of what I now test at algoholic.com.
What Google learned that the models have not: the ecosystem has to eat
Here is the uncomfortable one. Google, for all the complaints I have written about it over the years, understood that search lives or dies on publishers having a reason to publish. The deal was crude but real: content for clicks.
The answer engines have not solved their version of this. They consume the open web and return answers that keep the user inside the answer. In September I measured five large American domains: the pattern across them was branded demand growing while non-branded organic, the traffic that content actually earns, stalls or bleeds. The full numbers will run on this site once a second measurement confirms the pattern. That is how claims should work.
I do not know how this resolves. Licensing, attribution interfaces, traffic-sharing, regulation, or a slow hollowing of the open web. But I know from 27 years of watching this market that an engine which starves its own supply eventually changes course or degrades. Either outcome reshapes what visibility means. This is the question I find most worth measuring right now.
What never changed
Every era of this industry begins and ends with the same question: how does the engine decide what to show?
Google’s own people would phrase it from the other end: what does the user actually need. Both phrasings are the same discipline viewed from opposite sides of the results page, and the practitioners who last are fluent in both.
In 1999 the answer was keywords. In 2005 it was links. In 2012 it was quality signals. In 2019 it was language understanding. In 2026 it is retrieval and citation inside generated answers. The mechanics turn over every few years. The question has never moved.
I have heard that SEO is dead at least five times. It only changes shape. The people who survive each turn are not the ones with the best tricks for the current engine. They are the ones who keep asking the question after the engine changes.
Why this site exists
Algoholic started as a nickname in 1999, because I could not stop taking algorithms apart. It is now the lab where I do that in public: every claim measured, every experiment documented, every tool free to run. The deepest recovery I ever ran, a cannabis site buried by a YMYL tightening and brought back, is documented in the case studies here. The archive covers the road from AltaVista to generative retrieval, and the tools score the thing I believe matters next, whether your statements survive into machine-generated answers.
If you have been in this industry long enough to remember any of the eras above, you already know the next one will not be the last. The question stays the same. The measuring never stops.
Gilad Sasson, also known as Algoholic, is the CEO of nekuda Web Solutions and an Israeli SEO pioneer since 1999. Google AdWords certified since 2004. He publishes his research at algoholic.com.
Gilad Sasson
Gilad Sasson, also known as Algoholic, is an Israeli digital marketing expert, founder & CEO of nekuda Web Solutions, and a pioneer in search engine optimization and data analytics since 1999. Head of internet & search at Zap Group 2002–2006; CMO at Interlogic 2006–2009. Speaker at SMX Israel, TNW Amsterdam, Web Summit Dublin, DMIEXPO.