From Search Rankings to AI Recommendations: Is Your Brand Visible?

Ben cohen

Here's a scenario. Someone on a remote team needs a project management tool. They open an AI assistant, type one sentence about their team size and budget, and get back a shortlist of three brands with a line or two on each. Then they ask about pricing. Then which one plays nicely with their calendar. Twenty minutes later they've basically decided, and they haven't visited a single company website.

Ido

And if you're brand number four, you weren't rejected. You just weren't in the room.

Ben cohen

Exactly. There's no page two to scroll to. So the question for today is simple. How does a brand stay visible when the thing standing between the customer and the web is an assistant?

Ido

Welcome to Be Visible. I'm Ido Navarro. I build websites and, these days, agentic tools that let people manage them in plain language.

Ben cohen

And I'm Ben Cohen. I look at this from the growth side, and I also built a free tool that shows what a site looks like to AI systems once you strip out the styling and the JavaScript. Which is going to come up, because it's a humbling thing to look at.

Ido

It really is. So let's start with what changed. For twenty years, visibility meant being found. You rank, you get the click, the person lands on your page and your design does the persuading. Now the answer gets assembled before anyone clicks. So visibility also means being understood correctly, being picked as a source, being summarized fairly, and being recommended.

Ben cohen

Let me make that less abstract, because it's already happening to branded searches. Ahrefs looked at 232.8 million US desktop Google crawls from July to September 2026. As of late September, AI Overviews appeared on about 83 percent of branded keyword searches. In July the monthly average was 61 percent.

Ido

Branded searches. The one place you thought you owned.

Ben cohen

Right. And in a smaller panel of 100 big brand names, when an AI Overview showed up, it sat in position one about 76 percent of the time. The most common outside sources it cited were Wikipedia, YouTube, and LinkedIn. So someone types your company name and the first thing they read may not be your homepage. It may be a summary built partly from places you don't control.

Ido

I want to flag the caveat, though. Google changes this stuff all the time, and Ahrefs says so itself. That's a snapshot, not a law of nature. But the direction is hard to ignore.

Ben cohen

Fair. And there's the other side, the ChatGPT side, which connects to what we covered last time, when we talked about where AI goes when it searches. Ahrefs analyzed 66.6 million logged out ChatGPT responses from August and September. Only 14.4 percent contained those background searches, the fan out queries. Most answers came without searching at all.

Ido

Which is interesting, because it means two different games. If the model answers from what it already learned, you need to be part of what it learned. If it searches, you need to be retrievable at that moment.

Ben cohen

And when it does search, a lot of it goes straight to brand sites for facts. Among the most searched domains, about three in four of the site specific searches targeted a brand's own website. In a separate test of 20,000 US prompts, factual questions about pricing or features cited at least one official source 84.4 percent of the time. For comparison questions, like A versus B, that dropped to 68.3 percent, and more of the citations went to other sites.

Ido

So the pattern is, for facts, the model wants you. For comparisons, it wants what other people say about you.

Ben cohen

That's my reading, yes. One model, one two month window. I wouldn't tattoo it on anything.

Ido

Okay, so the obvious pushback. Is this really different from SEO? Because half the advice I read sounds like SEO with a new hat.

Ben cohen

I think the honest answer is, it's built on top of SEO, not instead of it. Olaf Kopp wrote a piece in Search Engine Land that I like, because he says the useful question isn't whether GEO replaces SEO. It's what you're trying to achieve. He splits it into three goals. Get your content cited. Get your brand recommended. Get your products selected.

Ido

And each one has a different owner inside a company. That's what jumped out at me. Content cited is an SEO and content job. Brand recommended is closer to PR and brand. Products selected is ecommerce and product data. It's not one team's checklist.

Ben cohen

Let's take them one at a time. Pillar one is what he calls LLM readability. How easily can a model process your content and use it as a source. And there's an order of operations. Document retrieval first, readability second. If your page was never crawled, indexed, and pulled into the pool of candidates, the prettiest paragraph in the world doesn't matter.

Ido

That's the part people skip. They go straight to clever formatting. But if a crawler can't reach the page, or the content only appears after heavy JavaScript runs, you're done. The Semrush guide makes a point of this, that many AI crawlers still struggle with client side rendered content.

Ben cohen

That's what my tool shows people, honestly. You paste in a page, and what the AI reader sees is sometimes a nav menu, a cookie banner, and half a sentence.

Ido

So once it's retrievable, what does readable mean? Because the models don't read a page like a person does. They split it into chunks, find the chunks that match the question, and stitch an answer from those pieces, often without the surrounding page.

Ben cohen

Which changes how you write. A paragraph that defines something, compares something, or states a fact should make sense on its own. If it says 'as mentioned above' or 'this is why', that paragraph falls apart when it's lifted out.

Ido

The Semrush example is good. Hard to extract is 'There are several reasons this method works. After trying it, most people find their eggplant tastes better.' Easy to extract is 'Salting eggplant for 15 minutes before cooking removes bitterness and excess moisture.' Same idea. One of them survives being quoted.

Ben cohen

Kopp gives some rules of thumb. Keep paragraphs short, he says under 250 words, one idea each. Use who, what, when, where, why questions as subheadings. Lead with the answer, then the explanation, then evidence. He also says LLM agents often only read the first 350 to 400 words of a long document, so the key statement belongs up top.

Ido

I'd treat that last number as a heuristic, not a spec. Systems differ and they change. But the habit is right either way. If your homepage spends the first screen on a slogan like 'Reimagining tomorrow', and the actual sentence 'We make scheduling software for dental clinics' shows up in paragraph six, that's a problem for people and machines.

Ben cohen

And a good test for humans. Can a stranger say what you do, who it's for, and why to trust you after the first few lines?

Ido

Which brings us to the infrastructure, and I'll use the analogy. A normal website is a beautifully designed restaurant. The lighting, the furniture, the music. Great for a person walking in. Clean markdown is the written menu. An AI just wants to know what's available and what it costs, without inspecting the decor.

Ben cohen

And the reality is that the page that wins design awards can be the hardest to read. Navigation repeated on every page, scripts, pop ups, tracking, layers of HTML wrapping one sentence. The signal is in there, buried.

Ido

So the idea is, serve a lean version. Plain headings, text, tables, links. Low noise. I think there's a real chance we end up with sites that present different representations of the same content, one for humans, one for search engines, one for agents. But I want to label that. It's a direction, and the details are not settled.

Ben cohen

Okay, and that's where llms.txt comes in, and the longer one, llms full dot txt. Explain it to me like I'm a business owner.

Ido

llms.txt is a plain file at the root of your site, written in markdown. Think of it as a short map. Who you are, your most important pages, each with a one line description. Llms full dot txt goes further. Instead of just pointing at pages, it can contain the actual content in one clean document. So one is the index at the front of the book, the other is the book with the noise removed.

Ben cohen

So does adding one suddenly make a site visible?

Ido

No. And I want to be blunt. It's not an official ranking signal, and I wouldn't claim any major AI company has confirmed they rely on it. It's a proposed convention. Where it can help today is practical. Developer tools, documentation, and agents or coding assistants that you point at your site, because you've given them a cleaner starting point. Beyond that, it's cheap, it forces you to decide what your most important pages are, and it might matter more later. But that's a hypothesis.

Ben cohen

I like that it forces the decision. If you can't write the ten most important pages on your site in a list, that tells you something about your site.

Ido

Right. Next layer, schema. Structured data gives machines explicit labels. This is an organization, this is its name, this is a product with this price, this is an article by this author, this is an event on this date, this is a local business at this address. Same for services, FAQs, and reviews.

Ben cohen

But there's a catch I've seen people trip on. Semrush says plainly that we don't actually know how structured data is used inside large language models, whether in training, retrieval, or answer generation. We know systems cross reference signals from several sources. So schema is helpful, not a guarantee.

Ido

And the rule is, schema has to match what's visible on the page. If the markup says the product is in stock for 49 dollars and the page says 59 dollars or sold out, you've built a contradiction, and contradictions reduce trust. Same for duplicated generic schema pasted on every page, or auto generated markup nobody reviewed.

Ben cohen

Also, schema can't rescue weak content. If the page itself is vague, labeling the vagueness precisely doesn't help.

Ido

Ha. Precisely labeled vagueness. That's going on a mug.

Ben cohen

Now pillar two, because so far we've been talking about your own site, and the data says that isn't the whole story. Kopp calls it brand context optimization. How does the model associate your brand with a category and a need.

Ido

The way he explains it, models represent concepts in a kind of semantic space, and brands that show up repeatedly near certain terms get linked to them. His example is Notion, which keeps appearing next to phrases like all in one workspace, team collaboration, notes and docs. So when someone asks about productivity tools, Notion is close by. That's his framing of how associations form, and I'd call it a useful mental model rather than a measured guarantee.

Ben cohen

It also explains the branded search data. Wikipedia, YouTube, LinkedIn. Those are places that describe you. If your own description is thin or inconsistent, someone else's version fills the gap.

Ido

And the Semrush guide has the monday.com example for entity clarity. The word monday appears in a million contexts. Consistent descriptions across your site and your profiles tell the model this is project management software, not the day of the week. Your LinkedIn, your Crunchbase profile, your directory listings should say the same thing as your homepage.

Ben cohen

Now, a caution on volatility. When Semrush tracked 2,500 prompts across Google AI Mode and ChatGPT, they found that between 40 and 60 percent of cited sources changed from month to month. So nobody can promise you a stable position. The brands that showed up consistently shared things like clear entities, extractable content, and presence on multiple platforms. It's about improving odds, not buying a slot.

Ido

And the proof it's real money. Semrush notes that for Tally, a bootstrapped form builder, ChatGPT became the number one referral source. That's a small company, not a giant, which is the encouraging part.