What is GEO (Generative Engine Optimization)?

It’s worth noting that, according to an analysis conducted in 2022 on over 60,000 search results, the top ten results on Google have an average score of around 78—and fewer than 15% exceed 90. This doesn’t mean that reaching 90 is necessary; it means that those who reach 90+ have already achieved a technical result above most of the sites analyzed.

One example is the performance metrics of a website we built for kivutpois.fi

(source: Core Web Vitals e ranking — SupportHost, 2022)

Unlike the Performance score, where a low result often points to real speed problems and can go hand in hand with worse Core Web Vitals, hitting 100 in Accessibility, Best Practices, and SEO on Google PageSpeed is often relatively easy. It means you've passed the automated checks included in those scores, not that the site is necessarily perfect in every respect.

You can have all three values at 100 even on a painfully slow platform. Performance, on the other hand, is much more affected by the site's actual speed. If the hosting is poor, performance can take a serious hit too.

And that's where both the Performance score and some Core Web Vitals can suffer.

So saving a few dozen or a few hundred bucks a year on hosting can turn out to be the wrong choice.

If you want to know how your site's hosting platform performs, ask the agency to show you results achieved with their actual sites, or go to Google PageSpeed and check your site yourself.

The TL;DR: Not Just for Readers, But for AI

Do we have a summary for people in a hurry? Having well-structured, clear, and easily summarizable content can make it easier for automated systems to analyze too. If your agency never talks to you about E-E-A-T, they might not be the best solution.

The TL;DR—“Too Long; Didn’t Read”—is a clear synthesis of the page's main points.
It's not just a summary for the reader: it's also a way to make the central theme of the content more explicit.

But careful—it's not enough to write "TL;DR" and hope it works. Google doesn't require special markup for AI Overviews or AI Mode.

A small technical example of how we use it in this article:

<aside data-ai-summary="true"><p>
<strong class="ai-tldr-label">TL;DR:</strong>
Your summary here.</p></aside>

data-ai-summary is an experimental attribute we use as an additional layer in the code. It doesn't replace semantic HTML, and we can't say that Google gives it any particular advantage.

We can also include an abstract in our JSON-LD, reporting the same synthesis used in the TL;DR:

{
"@context": "https://schema.org",
"@type": "Article",
"abstract": "Your TL;DR here."
}

This way, the same synthesis is present in both the HTML content and the structured data.

JSON-LD and the llms.txt File

One thing that genuinely paid off was rebuilding the kivutpois.fi website. I talk about it in more detail here.

Just 12 hours after going live, I saw the pages get indexed on Google—which doesn't always happen. But what really got me excited was watching the data grow over the following week. And instead of the usual link, I started seeing structured data showing up in the search results.

So I decided to test it from an AI perspective, to see if the work had actually paid off.

I asked the kind of question a regular person would ask: "I'm in Lapland near Kemijärvi and I need an osteopath. Who do you recommend?"

During the first week, I kept seeing the same links as always. But after about three weeks, in the tests I was running, kivutpois.fi was showing up at the top of the suggestions, right next to Lapin Fysioterapia Atlas AY—the company behind kivutpois.fi.

Even today, in my tests, kivutpois.fi often shows up at the top of AI search suggestions for other tourist destinations too, including the Pyhätunturi ski resort, where Lapin Fysioterapia Atlas AY has a clinic.

At that point, I tried something different. I asked: "What can you tell me about the page siti.italiadigitalweb.it? What JSON-LD and data-ai do you see on that page?"

That's when things got frustrating. Every single time, the answer was: "I don't see any JSON-LD on the page, or any data-ai attributes."

I tried everything—and yeah, it was frustrating. But I kept digging, and eventually I got a better idea of what might be causing it. So I'm sharing it here.

While working on the JSON-LD graphs to make them more complete and advanced, I realized there was an important difference between what actually exists on the page and what some chat tools are able to show you when they run a direct fetch.

When I validate the structured data on validator.schema.org and Google's Rich Results Test, everything checks out. If I check the page source, the structured data is right there—in the <head>, written correctly, like this:

<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@graph": [
    {
      ...
    }
  ]
}
</script>

So, the JSON-LD is there. It's on the page. It's written correctly.

But it's a different story when I ask a chat tool to inspect a page directly and run a fetch.

In several of my tests, the version of the page returned to the chat didn't contain scripts or JSON-LD, even though they were clearly present in the original page source.

This doesn't mean the JSON-LD doesn't exist or that a crawler can't read it. It simply means that the representation of the page that reaches the chat isn't always the complete HTML we'd see by opening the page source.

And that's exactly the difference that completely threw me off in my first tests.

The llms.txt File

There's also a newer tool that I find particularly interesting: llms.txt.

It's an emerging convention and, in its most common form, the file lives in the root of your website—just like robots.txt.

It's written in Markdown and is designed to give LLMs and agents a simple version of the site's structure and its most important resources.

Google Search has made it clear that it doesn't use llms.txt for ranking. But Chrome Lighthouse has now added a check for the file in its Agentic Navigation section.

So I don't see it as a replacement for SEO, the sitemap, or structured data. I see it as an additional layer for systems that decide to use it.

Version 2 Arrived on August 10, 2026

In the meantime, the llms.txt proposal has evolved. On August 10, 2026, it was updated to version 2, which no longer focuses only on an llms.txt file in the root of the site.

The new version also allows llms.txt files inside individual sections of a site and proposes making Markdown versions of pages available, for example through an index.md file.

The idea is simple: an agent can use llms.txt to understand where to look for information and then follow links to the pages it needs, possibly using a Markdown version that's easier to read than the full HTML.

Version 2 also introduces links that can tell agents where to find the Markdown version of a page and its related llms.txt file.

(source: llms.txt v2)

And What About data-ai? I'm Still Testing It

When I started experimenting with data-ai attributes, between October and November 2025, I'd found and cataloged around fifteen of them.

By August 2026, my registry already contained 17 with verified public use, plus many others found in specific products and implementations or still waiting to be verified.

Google clearly says that no special markup is required for its AI features. And that's fine.

But the question I asked myself was this: if these attributes weren't useful to any system, why do I keep finding new ones being used by different products? The fact is, today we know that some products really do use them.

The point is that each system can use them for its own purposes. There still isn't a universal vocabulary.

In my tests, I've also noticed that sites where I've used this structure get cited more often in AI responses. I can't prove that this is because of the data-ai attributes. It could be a coincidence, or it could be the result of everything else that was done on the site.

But the behavior has repeated itself enough times to make me keep testing them.

That's why I use them as an additional layer when they have a precise meaning, without replacing semantic HTML, accessibility, or Schema.org. If some of these markers become more important in the future, at least I won't have to start everything from scratch.

The Next Test: A Markdown Version of the Pages

Here I'm moving into an experiment I still have to run and that, interestingly enough, goes in a direction similar to the one suggested by the new version of llms.txt.

Cloudflare, for example, already makes its documentation pages available in Markdown too. By adding /index.md to a page URL, you can get a version that's much easier for an AI agent to read.

The idea interests me because, in my tests, not every chat tool is able to retrieve HTML, scripts, JSON-LD, and Markdown files in the same way.

What I've noticed instead is that TXT files are read very easily, even when the content inside them is written in Markdown.

That's why over the next few months I want to run a slightly different test: place an index.txt version alongside the HTML pages, with the same factual content written in Markdown and the important relationships made easier to read.

The index.txt format isn't part of the llms.txt v2 proposal, and I don't know yet whether it'll bring any benefits. It's an experiment.

The important thing is that AI systems aren't given information that's different from what people can see. The format changes, not the reality we're describing.

Once I have enough data, I'll publish the results and we'll see whether it was worth it.

Concrete Strategies: What Actually Works

Let's be clear: no miracles. Anyone promising immediate or guaranteed results is someone I'd be very careful with.

A good SEO and GEO setup can create a very solid foundation, but the work doesn't end when we publish the site. In fields where people are constantly looking for information, continuing to publish genuinely useful content can make a big difference over time.

(source: Generative engine optimization (GEO): How to win AI mentions — SearchEngineLand.com)

Interestingly, memes and jingles have started appearing on social media promising to get a site to top AI positions with just a few clicks.

Just a few days ago Diletta Beligotti, my colleague and expert Social Media Manager, wrote to me:

"Hi Roberto, I'm monitoring agency ads on Facebook and look what's showing up. They're starting to sell AI positioning to businesses."

My immediate response was:
"Yeah, but the way they tell it in the ad is pure snake oil."

To actually get mentioned by AI, there can be a ton—an absolute ton—of work behind it.
So be extremely careful with promises of easy or guaranteed results.

Often these are clickbait ads with vague or misleading information, like:
"My small business gets XX new customers a day thanks to ChatGPT," accompanied by promises of similar results with a few clicks.

I'd stay away from these promises.

If instead you want to invest in a serious strategy to increase your chances of being found and cited by AI too, below I'll explain what the work actually involves.

Editorial Plan and Quality Content

The foundation is always the same: a well-built platform that's fast, technically solid, and has consistent structured data that helps search engines better understand content, entities, and relationships.

But this is just the foundation.

A platform, even well-developed, if it contains very little content will inevitably have fewer chances to answer the many questions people may ask in its field.

Pages need to tell something useful. They need to offer real value and answer concrete questions. A simple "showcase site," however well made, will rarely be able to cover all these topics.

Anyone seriously intending to emerge in the local services landscape can work on:

• a site with coherent, well-implemented structured data;
• content that demonstrates real experience and expertise;
• consistent information across the site and other digital properties;
• experimental attributes like data-ai-* when they're relevant and purely additive.

The data must be cross-referenceable and verifiable. Consistency between content, technical structure, and online presence is important.

But we can't determine in advance how many weeks or months it'll take before a site gets cited by an AI or moves up for a specific search.

And this is where, for many projects, the editorial plan comes in.

It's not about publishing random articles or content written only to intercept certain keywords.
It's about offering a real service to your audience.

An effective editorial plan is born from people's real questions, from doubts that emerge in daily life, and from concrete problems seeking solutions.

If you're a professional in the medical field, for example, you can regularly analyze how to prevent, manage, or recognize certain conditions.
Whatever your field, the key is using your experience to help—and really help.

This Is the Real Secret!

Many make the mistake of focusing exclusively on acquiring technical authoritativeness or external links, forgetting that a quality platform must first and foremost be useful to people.

For a moment, let's set aside algorithms and keywords.
Let's put people at the center.

When attention is directed at human beings, articles have a much better chance of being genuinely useful and shareable.

Naturally, the expert will need to focus on producing genuinely valuable content.
It will then be the developer's or technical team's job to integrate it correctly into the platform, structure it clearly, and add the relevant structured data.

FAQ: The Most Common Questions About GEO