GEO vs. SEO: How to Stay Visible to AI Assistants

Ido

Picture someone I'll call Dana. She runs a small physical therapy clinic, and it's a Tuesday night. She opens an AI assistant and types: compare three scheduling tools for a small clinic, keep it under a certain monthly budget, and tell me which one plays nicely with my accounting software.

Ben cohen

And this is a made up example, to be clear. But I bet it sounds familiar to anyone who's used these tools.

Ido

Right, it's a scenario, not a study. The assistant comes back with a neat comparison. Pricing, strengths, weaknesses, and a recommendation. Dana emails the one it recommended and books a call. Total number of vendor websites she opened before that email? Zero.

Ben cohen

And here's what hurts. The two vendors that weren't recommended will never know. There's no bounce in their analytics, no lost lead in the CRM. The shortlist got built in a place they couldn't see. And the winner will look at the demo request and think their marketing worked, when really the assistant did the first three rounds of the sales process.

Ido

So the question for today is a strange one. How does a brand stay visible when the thing standing between the customer and the web is an AI assistant?

Ben cohen

Welcome to Be Visible. I'm Ben Cohen, and I spend most of my time on how companies get found, and how that turns into actual customers.

Ido

I'm Ido Navarro. I come at it from the building side. Websites, WordPress, page structure, and now agents that help people manage sites in plain language. Ben asks what it means for the business, and I get to say what has to change in the site itself.

Ben cohen

We'll try to keep this practical. And honest, because this space has a lot of confident claims. So along the way we'll separate four things: what works today, what's still experimental, what's likely to matter, and what's currently just a hypothesis.

Ido

Let's start with what changed. For about twenty years, visibility meant being found. You ranked, someone clicked a blue link, and your page did the persuading.

Ben cohen

Right, and the whole economy was built around that click. Rank, click, land, convert. Every number on the dashboard assumed the person showed up.

Ido

Now a generative system sits in the middle. It reads a bunch of sources, retrieves some live, leans on what it learned in training, and writes an answer. And for your brand to win there, a lot of separate things have to go right. You have to be discovered. Then understood correctly. Then seen as credible. Then quoted or referenced. Then recommended at the moment of decision.

Ben cohen

That's five gates, and ranking only used to be one of them. And there's a sixth that people haven't thought about yet. Can the assistant, or an agent working for the person, actually do something with you? Book, buy, configure.

Ido

Yes, and we'll get there. One more shift though, because it changes how you measure. These systems are probabilistic. Ask the same question twice and you might get different answers, different brands, different order.

Ben cohen

So the idea of ranking number three for a keyword kind of dissolves. You're not standing at a fixed position in a list. You're appearing or not appearing across a lot of conversations, and the useful numbers look more like mentions, citations, and share of voice.

Ido

Which brings up the question everyone's thinking. Is this really different from SEO? Or is GEO just SEO with a new hat?

Ben cohen

That's my question too, honestly, because I don't want to tell a founder to throw away what's working. And I'd push back on the framing that SEO is dead. If an AI system retrieves live information, it still needs to reach your page, and a fast, crawlable, authoritative site still helps.

Ido

I agree, so let's call it layers. The foundation is the same. Can the page be accessed, is it fast, is it structured sensibly, does it have real authority. SEO fundamentals stay. What GEO adds is a question about extraction. Can a model pull a clear answer out of this page, and would it be right if it did?

Ben cohen

And that's where the business side shows up for me. Think about what an assistant needs to say about you in one sentence. What you do, who you serve, and why anyone should trust you. If your homepage says we empower synergy for modern teams, the model has nothing solid to repeat. Or worse, it fills the gap with something from a three year old directory listing.

Ido

That's the entity piece. A model needs to know that this company is this kind of thing, that this person wrote this article, and that this product belongs to that company. It sounds obvious, but a lot of sites are vague about it. Authors listed as admin, no about page that says anything concrete, product names that change from page to page.

Ben cohen

And trust doesn't only come from your own site. This is more of an interpretation than a proven rule, but it makes intuitive sense. If your company is described consistently on review sites, in industry articles, on community platforms, then the picture a model forms of you is more stable. Some vendors in this space say third party mentions and places like Reddit matter a lot. Nobody outside the labs can tell you the exact weighting, so treat it as a strong hypothesis, not a formula.

Ido

I'd add that the content itself has to be easy to lift. Direct answers near the top. Clear definitions. Comparisons that are actually comparisons, with real attributes. Original data or real examples, because a model has no reason to cite the hundredth page that repeats the same generic advice.

Ben cohen

Which is also a conversion point. A page that answers the question cleanly is easier for a human to trust too. I get nervous when people hear GEO and think of tricks. The honest version is mostly clarity.

Ido

Okay, so let's go under the hood, because this is where it gets practical. First thing, and it's almost embarrassingly basic. Check your robots.txt.

Ben cohen

This is the one I want every listener to do after the episode. Go to your domain slash robots.txt and look.

Ido

Because here's what happens. A security plugin, or a firewall setting, or an old rule someone added years ago, blocks bots it doesn't recognize. And the newer AI user agents, GPTBot, ClaudeBot, PerplexityBot, Google Extended, can get caught in that. If they're blocked, you can have great content and strong traditional rankings and still be invisible to those systems.

Ben cohen

Now, to be fair, blocking can be a legitimate choice. Some publishers don't want their content used for training. But the mistake is blocking by accident, or blocking without understanding that it also changes whether you show up in answers. Decide on purpose.

Ido

Exactly. Second thing is weight. Take a typical WordPress page built with a visual builder. The sentence you care about, say your pricing or your service description, can be sitting inside layers of wrapper elements, plus scripts, plus navigation and footer repeated on every page. Humans never see that. A model reading raw HTML has to wade through it, and context windows are limited, so wasted tokens cost time and money.

Ben cohen

I'll jump in, because I've got a small bias here. I built a tool, the GetAISO AI Content Reader, that strips a page of styling and scripts and shows you what's left. And the first reaction people have is almost always surprise. The page looks gorgeous in the browser, and the stripped version is a mess, or it's missing the very thing they thought was the main message.

Ido

One caution, since I come from the Elementor world. One WordPress guide I read says heavy page builders hurt GEO and recommends using them for landing pages, not blog content. I'd treat that as a claim about code bloat, and the real answer is more nuanced. A well built page with lean structure can be fine. A poorly built one with fifteen nested containers per paragraph is the problem. Measure the actual output, don't blame the tool by name.

Ben cohen

So what's the fix? Rebuild everything?

Ido

No. This is where clean markdown comes in. Think of your website as a beautifully designed restaurant. Great lighting, nice furniture, a vibe. Clean markdown is the written menu. The AI doesn't need to inspect the decor to know what's being served. It needs the dishes, the prices, the ingredients.

Ben cohen

And the menu doesn't replace the restaurant. People still want the experience.

Ido

Right, and this is the bigger idea. The same content may eventually be served in different forms. A rich visual page for the person, a simple text version for a machine, maybe a structured action layer for an agent. The llms.txt proposal even suggests offering a markdown version of a page at the same address with dot md added. Some companies report big token savings from serving markdown instead of HTML. One guide says up to ten times, but that's a vendor claim, so I wouldn't promise anyone that number.

Ben cohen

Let's talk about llms.txt directly, because I think this is the most oversold item in the whole space. Ido, does adding an llms.txt file suddenly make a website visible?

Ido

No. Short answer. And I want to explain what it is first, then why people are confused. It's a markdown file at the root of your site. The proposal came from Jeremy Howard of Answer.AI in 2024. The format is simple. One heading with your site's name, a short summary in a quote block, then sections of links, each with a one line description. There's a section called Optional for pages an agent can skip if it's short on space.

Ben cohen

So it's a curated map. Not a sitemap.

Ido

Exactly. A sitemap lists everything. This says, here are the twenty to fifty pages that matter, and here's when you'd want each one. And there's a companion idea, a file people call llms full dot txt, which doesn't just link to pages but includes the actual content in one payload, so an agent can take it all in with one fetch. That makes sense for documentation. For a marketing site, the index alone is usually enough.

Ben cohen

Okay, but is anyone reading it? Because I've seen claims in both directions.

Ido

And that's where the evidence is messy, so let me give you both sides. On the skeptical side, one analysis from a company called Limy looked at over five hundred million AI bot events across brands it monitors. Over a ninety day window, only 408 targeted llms.txt directly. A separate study from SE Ranking, covering 300,000 domains, found about 10 percent had one. And in July 2025, Google's Gary Illyes said Google doesn't support it, and John Mueller compared it to the old keywords meta tag.

Ben cohen

But then there's the other camp.

Ido

Right. A report from Presenc AI says Anthropic and Perplexity have publicly confirmed support, and that OpenAI shows observable effects without a public commitment. Meanwhile the Limy piece says no major provider has committed to using it as a signal in production search. Those two accounts don't agree, and both come from companies that sell visibility tooling. So my honest read is, nobody should promise you a ranking lift from this file.

Ben cohen

So why would a business bother?

Ido

Because it's useful somewhere else. Developer tools like Cursor and Claude Code reportedly look for these files when pointed at documentation. There are MCP servers built around exposing them. Documentation platforms generate them automatically, and the big AI labs publish their own for their developer docs. So it's less a search ranking play and more Business to Agent infrastructure. A clean entry point for software that acts on someone's behalf.

Ben cohen

Which reframes the ROI question for me. If you sell developer tools, it's a developer experience feature. If you're a local bakery, you're probably not losing sleep over it, and the effort belongs elsewhere. If you're somewhere in between, it's cheap insurance that you shouldn't confuse with strategy.

Ido

Yes. And if you do write one, write it by hand or at least edit it by hand. Concrete facts in the descriptions. Actual pricing tiers, not affordable solutions for every team. Keep it short. Put the file at the root, don't put it behind a login, and keep it consistent with your robots.txt. Then test it by giving an agent only that file and asking it questions about your business.

Ben cohen

That test is great, because it tells you whether the thing says anything useful.