GEO: Why Search Engines Finally Learned to Recognise Real Quality

For twenty years, search rewarded the people who gamed it best. The new LLM-driven engines reward the people who actually know what they're talking about.

For two decades companies optimised their sites for algorithms that ran on simple rules. Keyword density, the number of backlinks and server speed decided who ranked where. Spam pages beat substantive ones because they were better at playing the positioning game, and anyone writing for human readers paid a tax to anyone writing for the crawler. That has flipped. At Xfaang we ran the GEO work for K MAG, a magazine with fifteen years of bylined, hand-written archive behind it, and it worked: the new engines reward the exact thing the old ones were blind to.

The new engines are built on large language models. ChatGPT Search, Google AI Overviews and the rest read a page closer to the way a person reads it, and judge it on what it says rather than on how well it was tuned. The practice that has grown around them is called GEO, Generative Engine Optimization, and it has almost nothing in common with SEO.

§ 01 · The shiftWhat the new rankers actually read

An old search engine asked which page contained the phrase and how many sites linked to it. A model-driven one asks whether the page was written by somebody who knows the subject. In practice it reads four things, and all four are hard to fake:

Author identity: is there a real person behind the text, with a track record somebody can check? User-side quality: the ratio of ads to content, the volume of pop-ups, the share of affiliate filler. Authenticity: was this written by a human or generated by a model? Verifiability: do the facts, dates and numbers line up with independent sources?

None of that can be produced in a quarter. A fifteen-year publishing history is either there or it is not, an editorial standard cannot be retrofitted onto an archive already full of thin SEO pages, and model-generated articles do badly against writers with real experience, particularly when the thing grading them is also a model.

Search used to reward whoever wrote best for the crawler. The crawler is no longer the reader.

§ 02 · What this means in practiceThe incentive structure flips

Sites that built their traffic on volume, thousands of low-effort posts aimed at long-tail keywords, now look like what they always were. Places that look small on the old metrics but carry a named author, a point of view and citations that hold up are the ones turning up in answers from the assistants that plenty of readers now use instead of a search box.

There is a defensive side to this too. Once authenticity counts as a ranking signal, checking whether a text was really written by the person whose name is on it becomes part of publishing rather than a curiosity. Tools that estimate whether a piece is human or model-written are turning into ordinary editorial equipment, the way spellcheckers and plagiarism scanners did.

A ranker that reads for meaning makes a content farm look like a content farm.

§ 03 · The workWhat we did at K MAG

I wrote the Polish original of this piece for K MAG, a magazine that has published only hand-written work with named authors for fifteen years. It is also the magazine Xfaang did the GEO work for, and the useful part of that story is how little of it was about tricks. The substance was already there. Most of the job was making sure the engines could see it: authorship, dates and sources where a ranker can read them instead of buried in a template.

The team there also built VerifAI, a tool that checks whether a text is authentic before it goes out. On its own that is a small piece of tooling. Next to fifteen years of named authors it is what an editorial standard looks like once spinning up a content farm stops being a competitive advantage.

K MAG looks at this shift with relief, and so do we. The magazine was built as if this day had already arrived. It has now.