Auditing AI Overviews, Brand Mentions, and the Reality of llms.txt

Ido

Here's a number that should bother anyone who owns a brand. Ahrefs analyzed 232.8 million US desktop Google results, from July to September 2026. As of September 29th, AI Overviews appeared on 82.91% of branded searches. In July, the monthly average was 61.3%.

Ben cohen

Wait, branded. So somebody types your company name, not a category, your actual name, and Google writes the summary before they ever see your site.

Ido

Yes. And in their separate sample of 100 big brands, 63 had an overview by the end of September. Forty eight of those sat at position one, that's 76.2%. So your own domain gets pushed down by a paragraph you didn't write.

Ben cohen

And the paragraph borrows from other people. Wikipedia, YouTube, LinkedIn. Wikipedia is mostly neutral, fine. But YouTube and LinkedIn are places where a brand can barely steer the story. Ahrefs even pointed out that a video might carry outdated product information.

Ido

Right. Wikipedia showed up in about 72% of those external citations, YouTube about 39%, LinkedIn about 23%. That's one sample, so I wouldn't treat it as a law. But the direction is clear. Your storefront now has an intermediary interpreter.

Ben cohen

Welcome to Be Visible. I'm Ben.

Ido

I'm Ido. Last time we looked at how AI engines synthesize brand queries. Today is the tactical sequel. How do you audit what they say, and how do you make your site easy to represent correctly?

Ben cohen

Start with the audit, because most people skip it. Marcus Miller's ASC framework, analysis, SEO, content, begins with a very unglamorous step. Just ask.

Ido

Manually, and in a private window. Otherwise you're seeing what the AI thinks of you as a known user, not as a stranger. Fresh chat for every prompt. Then three stages. Awareness, like what's the best way to fix a problem. Consideration, like best providers in your category, or competitor alternatives. Action, like is your brand any good, or you versus a competitor.

Ben cohen

And you record three things. Were you mentioned, who else was, and which sources were cited. That third one is the gold, because it shows where the story actually comes from.

Ido

Which leads to something uncomfortable. Ahrefs looked at 75,000 brands, and backlinks barely registered for appearing in AI Overviews. What tracked more closely was how often a brand gets talked about across the web, linked or not. The top quarter by mentions got more than 10 times the citations of the quarter below.

Ben cohen

So the old link building reflex isn't the lever. Mentions are. But I'd add the business side. Cyrus Shepard studied over 400 sites. Winners had proprietary assets, 92.9% versus 57.1% of losers. Task completion, 83.7% versus 50.2%. Tight topical focus, 75.9% versus 61.3%.

Ido

Though that's correlation. Still, it says the website isn't decoration. Put the answer near the top, keep the HTML clean, name your services the same way everywhere, and make schema match what's visibly on the page. Schema that drifts from the content is worse than none.

Ben cohen

Okay, the question everyone's waiting for. Does adding an llms.txt file make me visible?

Ido

No, and I want to be careful here. Jeremy Howard proposed it. It's a markdown file at your site root. A title, a short summary in a quote block, then lists of links with brief notes, pointing to clean markdown versions of your pages. It's a curated index for inference, meaning when an agent needs information to help a user. It's not crawl permission. Robots.txt controls access. This is more of a guided map.

Ben cohen

Think of your website as a lavish restaurant. Lighting, furniture, music. llms.txt points to the menu, and clean markdown is the menu. Plain text, here's what we serve and what it costs. The AI doesn't have to inspect the decor.

Ido

That's why it matters that pages are noisy. Navigation, scripts, repeated interface pieces. The proposal itself says converting that back to clean text is difficult and imprecise.

Ben cohen

But let's be honest about the evidence. Google Search Central said llms.txt isn't necessary for its AI search surfaces. John Mueller said it's about agent functionality, not search discovery. And OtterlyAI's server logs over 90 days found 84 of 62,100 AI bot requests hit the file. That's 0.1%.

Ido

Right. No major AI platform has confirmed using it for search. So it's an agent readiness experiment and a documentation layer. Where it's clearly used today is developer docs. Lighthouse even checks for one. For a normal business, it's a cheap experiment, not a silver bullet.

Ben cohen

Same with tools. Something like LLMagnet can track which prompts you appear for, show gaps between your site copy and what AI says about you, and generate an llms.txt. Useful. But it's the manual audit made repeatable, not magic.

Ido

And it can't fix a weak business or vague copy. If your site doesn't say who you serve, no file will say it for you.

Ben cohen

So the week's plan. Run your twenty prompts in a private window. Note the cited sources. Fix the gaps on your own pages first, then shape the places that talk about you. Then try the menu.

Ido

And test the menu. The llms.txt guidance says to give an agent only that file and ask it questions about your business. If it answers wrong, your menu is broken.

Ben cohen

That's the takeaway. You've lost narrative sovereignty, not permanently, but you only get it back by knowing what the interpreter says. If this helped, follow Be Visible and send it to someone who owns a brand.

Ido

Ask three AI systems the same buying question about your company today. Compare. See you next time.