For Whom the Algorithm Ranks: What Ernest Hemingway Can Teach Us About GEO

A Lesson From Hemingway

There's a story about Ernest Hemingway: every morning he would sit down and write. Not when inspiration struck, not when he felt particularly creative, but every single day, treating writing with the discipline of a bodybuilder rather than the temperament of an artist.

Bear that in mind as you read the rest of this article. I promise it will all make sense by the end.

An Industry Built to Explain the Unexplainable

One of the recurring patterns in technology is that every genuinely disruptive innovation immediately generates an ecosystem of companies promising to explain it, optimise it and, ultimately, monetise it. Artificial intelligence is no exception.

In fact, one could argue that the commercialisation of AI has progressed considerably faster than our understanding of it. We barely share a common vocabulary for describing what large language models actually are, yet an entire industry has already emerged claiming to know how to make them recommend your company.

Every week I come across platforms promising to improve your "AI Score," increase your "LLM Visibility," optimise your GEO, or unlock the hidden mechanisms that determine whether ChatGPT, Gemini, Claude or Perplexity will mention your brand. The message is almost always the same: if you don't start investing in AI optimisation today, you're going to disappear from tomorrow's internet.

The more evidence I collect, however, the more I find myself questioning not the existence of Generative Engine Optimisation itself, but the way it is increasingly marketed.

To Be Clear: GEO Is Real

I am not arguing that GEO doesn't exist. Quite the opposite. I've probably been talking about this topic longer than many of the companies currently selling GEO solutions.

Five years ago, I was already arguing that conversational interfaces would progressively replace traditional search as the primary way users interact with digital information. At the time, of course, nobody called it Generative Engine Optimisation, because the acronym simply hadn't been invented yet. The concept, however, was already there.

In 2023, I dedicated an entire chapter of my book, We Are the Glitch, to this transition. The argument was relatively straightforward: we were moving from a web where users actively searched websites to one where artificial intelligence would increasingly serve as the interface through which websites were discovered. In that scenario, the scarce resource would no longer be ranking, but recommendation, and recommendation would necessarily depend on the quality, accessibility and contextual relevance of the knowledge available to AI systems.

Looking back, I still think that diagnosis was broadly correct.

What I did not anticipate was how quickly an entire consulting industry would emerge around the assumption that this represented a fundamentally new optimisation problem, rather than, in many respects, an old publishing problem wearing new clothes.

Those are two profoundly different propositions.

A Disclosure

Before anyone misunderstands my position, let me add some context: I currently sit on the advisory board of a company building GEO technology. I won't mention its name, because I don't want this article to read as disguised marketing, but the irony is that this experience has reinforced, rather than weakened, my scepticism toward many of the claims currently circulating in the market.

One of the very first conversations we had wasn't about AI rankings at all. It was about identifying gaps in knowledge and helping hotels publish the information that AI systems were already looking for.

That distinction, I increasingly believe, is the entire point.

AI Doesn't Work Like Google, Only Smarter

One of the biggest misconceptions surrounding GEO is the assumption that AI works like Google, just smarter. In fact, I suspect a large part of today's GEO industry is unconsciously applying twenty-five years of SEO mental models to a technology that behaves in fundamentally different ways.

For more than two decades, SEO trained us to think in terms of rankings, authority scores and domain ratings. Those concepts made perfect sense within a deterministic search engine whose primary task was ranking documents.

Large language models, however, perform a completely different cognitive operation: they retrieve fragments of information from multiple sources, weigh those fragments differently depending on the context of the conversation, synthesise them into a coherent answer, and then repeat the entire process every time a new prompt is submitted.

That may sound like a subtle distinction, but it isn't. The consequence is that asking the exact same question twice doesn't necessarily produce the same answer. From a machine learning perspective, this is perfectly normal. From an SEO perspective, it's almost heresy.

Which is precisely why I wonder whether much of today's GEO conversation is, at its core, an attempt to translate a probabilistic system into deterministic language. We keep talking about "ranking in ChatGPT," not because that's necessarily how ChatGPT works, but because ranking is the only conceptual framework our industry has inherited from Google.

The problem with metaphors is that they eventually become prisons. If we insist on treating large language models as though they were search engines, we may end up optimising for concepts that don't actually exist.

What the Data Actually Shows

The empirical evidence increasingly points in exactly that direction. Over the past few weeks, two independent studies have been published which, taken together, paint a rather different picture from the one often presented by GEO vendors.

The first, conducted by Kollective Technology, analysed more than 13,500 prompts across ChatGPT, Gemini, Perplexity and Google AI Overviews. Its conclusions are striking: ask ChatGPT the exact same hotel question twice, and the number-one recommendation changes 45% of the time. When comparing different AI platforms, they agree on the top recommendation only 4% of the time.

Think about what those numbers actually imply. If the same model changes its own answer almost half the time, while different models almost never agree with each other, then the very idea of a universal AI ranking becomes increasingly difficult to defend. We continue to talk about "being number one in ChatGPT" as if ChatGPT behaved like Google circa 2015, when in reality the retrieval process itself is dynamic, contextual, and inherently probabilistic.

The second study, published in Search Engine Journal, analysed more than 107,000 AI responses across eight different AI platforms. Its most striking conclusion wasn't which brands ranked highest, but the fact that almost 90% of the analysed brands weren't mentioned at all.

So What Are We Actually Optimising For?

Taken together, these studies suggest something I believe gets lost amid much of the current GEO excitement: we're collectively behaving as if AI discovery were simply another version of Google Search, where rankings are relatively stable, and optimisation means climbing incrementally through a deterministic list of results.

Large language models simply don't behave that way. Which naturally raises another question: if the system itself is probabilistic, what exactly are we trying to optimise?

The answer, at least based on everything I've observed empirically over the past twelve months, is both surprisingly simple and profoundly unfashionable. What AI systems appear to reward has remarkably little to do with technical sophistication, and a great deal more to do with informational richness. Put differently, they don't seem to care nearly as much about whether a page has perfect schema markup or beautifully crafted JSON-LD as they do about whether that page contains knowledge worth retrieving.

Over the past year, I've been monitoring AI mentions across dozens of clients, comparing how ChatGPT, Gemini, Claude, Perplexity and Google AI Overviews retrieve and cite hotel content. Going into the project, I expected homepage authority, technical SEO, structured data or booking-engine optimisation to play a decisive role in determining which pages surfaced most frequently.

What surprised me wasn't even which hotels appeared, but which pages AI systems consistently chose to retrieve. In the overwhelming majority of cases, they weren't homepages or landing pages painstakingly optimised for conversions, nor were they pages packed with every conceivable schema implementation. They were blog articles and FAQs, editorial content. If I had to estimate, I'd say that roughly 95% of the AI mentions I monitor originate from this type of content.

Thinking Like a Traveller, Not an SEO Consultant

Once you stop thinking like an SEO consultant and start thinking like an actual traveller, this outcome becomes almost inevitable. Nobody opens ChatGPT and asks, "Book me Room 203." People ask:

"What should I do in Rome this weekend?" or "What's happening in the city in October?" or "Which boutique hotel would you recommend if I'm attending this conference?" or "Where can I experience authentic local food?"

These aren't transactional queries but rather exploratory conversations and uncertainty reduction. They're classic top- and mid-funnel behaviour, where people are still trying to understand a destination rather than compare prices or room categories. That distinction matters enormously because it fundamentally changes what an AI system needs to answer the question.

Which leads to what is perhaps the most obvious point in this entire discussion, and yet somehow the one that receives the least attention: before a large language model can retrieve knowledge, somebody has to produce it. LLMs don't create expertise out of thin air; they compress, synthesise and reorganise expertise that already exists somewhere else.LLMs don't possess expertise: every answer is assembled from knowledge someone else has taken the time to write and publish. In this regard, AI is making authorship more valuable than ever. The better the models become at finding useful information, the more they depend on people willing to create it. If nobody has written a thoughtful article explaining why the Tomb of Seneca on the Appian Way is worth a visit (spoiler alert: it is), it becomes rather difficult for an AI model to recommend it with any confidence. If no hotel has ever connected that monument to its own location, services, or guest experience, the model has very little material to build an answer.

Which is why I increasingly suspect that what many people currently call "AI optimisation" is, in reality, something our industry has known for decades: publishing. The web has quietly become a knowledge ecosystem again, and knowledge is precisely what AI systems consume.

A Case Study: The Hotel With the Worst Website and the Best Visibility

Perhaps the best illustration of this comes from one of my own clients. Ironically, one of the hotels that consistently enjoys the highest AI visibility among all the properties I monitor also has one of the weakest technical websites I've seen in years. Its Authority Score is in the single digits; there are thousands of technical issues that any SEO consultant would immediately want to fix; and, objectively speaking, the website is far from a showcase of technical excellence.

If technical optimisation were really the decisive variable, this property shouldn't perform particularly well. And yet it appears again and again across multiple AI platforms, far more often than its competitors.

The explanation, in my opinion, has almost nothing to do with technical GEO and almost everything to do with editorial discipline. The owner publishes a new piece of good content EVERY SINGLE DAY. She writes about local events, recommends restaurants, explains neighbourhoods, publishes seasonal itineraries, and answers questions that prospective guests haven't even realised they have yet.

Over time, she has built a knowledge base. From the perspective of a large language model, that's an extraordinarily valuable asset, because if an LLM's job is to retrieve useful information, organisations that consistently produce useful information naturally become easier to retrieve.

That sounds almost embarrassingly obvious, which perhaps explains why we tend to overlook it.

The Blog Is Back

There's another irony here that I find difficult to ignore. For almost twenty years, we've been declaring blogs dead, myself included. As Google evolved into an increasingly transactional ecosystem, businesses gradually shifted their attention away from editorial content and toward conversion optimisation. Particularly in hospitality, writing destination content came to be regarded as something vaguely old-fashioned: useful perhaps for SEO fifteen years ago, but no longer worthy of serious investment.

Artificial intelligence appears to be reversing that trend. A conversational model attempts to synthesise enough reliable information to answer a question, and inevitably rewards organisations that possess rich, structured, experience-based knowledge.

In retrospect, perhaps we (perhaps I) should have expected this.

Even Google Agrees

Interestingly, even Google seems to be pointing in exactly the same direction. In its recently updated Search Central documentation on AI-powered search experiences, it explicitly states that optimising for AI search is still SEO. More importantly, it dismisses many of the techniques currently marketed as “revolutionary GEO practices,” including llms.txt, content-chunking strategies, AI-specific rewrites, excessive attention to structured data, and several other supposed optimisation "hacks."

Instead, Google's guidance repeatedly returns to a principle that has remained remarkably stable for more than two decades: create original, useful, experience-based content, because that is ultimately what search systems (including AI-powered ones) are trying to surface.

As someone in love with words, I find that refreshing. The company operating the world's largest search engine appears to be saying something considerably less glamorous than many GEO vendors.

Echoes of 2003

In many ways, this reminds me of the SEO industry around 2003. Back then, everyone was obsessed with keyword density, meta keywords, PageRank sculpting, and countless technical tricks that, in hindsight, mattered far less than we believed. Eventually the industry matured and rediscovered something almost embarrassingly simple: search engines were getting progressively better at rewarding useful content, while becoming progressively worse targets for technical manipulation.

I wouldn't be surprised if, ten years from now, we look back at today's GEO discussions in much the same way. We'll probably smile at AI Scores, llms.txt files and endless optimisation checklists, before quietly admitting that the organisations which consistently won were often those that simply became the best publishers in their category.

Could all of this change? Absolutely. Anyone pretending to know exactly how AI retrieval will work twelve months from now is either extraordinarily optimistic or extraordinarily confident. And if none of us truly knows where this is heading, perhaps we should also be slightly more sceptical of anyone claiming to have already solved it.

Investing in What Survives

Personally, I'd rather invest in assets whose value survives regardless of how AI evolves. A library of thoughtful, authoritative content will remain valuable whether the future belongs to ChatGPT, Gemini, Claude, Perplexity or whatever comes next. A genuinely useful article doesn't stop being useful because somebody invents a new optimisation acronym.

Which ultimately brings me back to the question I think we should have been asking from the very beginning: perhaps the real issue isn't whether AI will recommend your business. Perhaps it's whether you've produced enough useful knowledge to deserve being recommended in the first place.

Those are two very different questions, and confusing one for the other is, I suspect, what has allowed an entire GEO industry to emerge around optimisation, when much of the real opportunity still lies where it has always been.

The Real Paradox of GEO

Maybe that's the real paradox of GEO: the more sophisticated our discovery systems become, the more they reward something remarkably unsophisticated.

Hemingway (I told you it would make sense in the end!) would probably recognise that.
Not because he understood schema.org.
Because he understood the discipline of writing.