When AI Searches, Where Does It Go?

Ido

Last time we tossed out two numbers: 14.4% of ChatGPT responses use query fan out, and Forrester's finding that 94% of business buyers use AI search when buying. Today I want the part we left hanging. When a model does go searching, where does it actually go?

Ben cohen

And the answer is more interesting than I expected. Ahrefs, in a study by Louise Linehan, looked at 66.6 million free, logged out GPT 5.6 Luna responses and 15 million fan out queries from August and September. Only 14.4% of responses had any fan out, averaging about 1.6 searches. So roughly six in seven prompts got answered without searching the web at all.

Ido

Which means your page only matters in that smaller slice. And a caveat they make themselves: it's one model, free sessions, a two month snapshot. Retrieval changes constantly.

Ben cohen

Fair. But inside that slice, 6.8% of fan outs used the site: operator, spread across about 116,000 domains. And among the most searched ones, 75.9% pointed at brand sites. Apple, Ikea, Nike, Adidas, Samsung.

Ido

Right, and the vocabulary backs it up. The word official showed up in 13.1% of English fan outs in September, up from 10.1%. Price went from 6.0% to 8.9%, which is a 47% relative jump in a month. The model is basically saying, go ask the company directly.

Ben cohen

So the brand site is back in charge? Because that sounds like great news.

Ido

Only for half the journey. In a separate test of 20,000 US prompts, factual questions about pricing, features, or company details cited at least one official source 84.4% of the time. For comparisons, like A vs B or which is cheaper, that dropped to 68.3%. And looking at the citations themselves, 63.4% went to official sites for factual questions, but only 46.4% for comparisons.

Ben cohen

So for facts, your site is the reference book. For comparisons, the model wants a jury. Which is my interpretation, to be clear, the study reports the pattern, not the intent.

Ido

Same reading here. And it changes the work. Your site has to be the cleanest possible source of facts, and then you also need to be talked about, accurately, by sources you don't control.

Ben cohen

The business side is the uncomfortable part. A buyer builds a shortlist inside the assistant, and you never see it. When someone does click, the referrer is often stripped, so it looks like direct traffic. Ahrefs suggests two rough methods: multiply AI referred conversions by average customer value, or just ask new customers how they heard about you and extrapolate. Imperfect, but better than guessing.

Ido

Now the trap I want to flag, from Pedro Dias in Search Engine Journal. Research on model memory shows a model can recite your product description word for word when you give it the first sentence, and still fail when a buyer asks which product fits their workflow.

Ben cohen

So repeating our tagline everywhere isn't a strategy.

Ido

Not a reliable one. In one small synthetic biography experiment, accuracy jumped from 9.7% to 96.2% when facts were worded differently and entity names were repeated. But that's a tiny model on invented people, not a promise about commercial systems. And in a Mistral 7B test, giving the model relevant passages at answer time hit 87.5%, versus 50.4% from base training and 58.8% with paraphrased training. Caveat again: those questions came from the same text the retrieval could supply, and it was multiple choice.

Ben cohen

So the lesson is, don't count on being remembered. Be easy to retrieve.

Ido

Which brings us to infrastructure. Think of your website as a lavish restaurant. Great lighting, wallpaper, the whole experience, and humans love it. A machine just wants the menu.

Ben cohen

And clean markdown is the menu. Strip out the navigation, scripts, repeated interface pieces, and heavy page builder markup, and you leave the facts in short, extractable blocks. I'd start by looking at your own page the way a machine sees it, with no styling at all.

Ido

Then llms.txt. It's a file at your root that gives models a curated map of your site, with a fuller version that includes the actual content in markdown. Useful today for reducing parsing noise, especially for agents ingesting your structure. But does adding it make you visible? No. I haven't seen evidence that models rank you higher because the file exists. It's an experimental piece of a bigger strategy.

Ben cohen

Same with schema. Organization, Product, FAQ, Service, it gives explicit context. But it has to match what's visible on the page. Generic, duplicated, or auto generated schema that contradicts your copy hurts trust, and it can't rescue weak content.

Ido

And then the third audience. A website now serves a human, a machine that interprets, and an agent that wants to act. If an agent can read your pricing but can't fill your quote form because the labels are vague or a script blocks it, you're only halfway visible.

Ben cohen

So what can someone do this week?

Ido

Check how assistants describe you. Make your core facts explicit in plain text: what you do, who you serve, pricing, specs. Add accurate schema. Check robots.txt, because blocking AI crawlers without understanding it can mean disappearing from the shortlist. Then test a real journey with an agent. Tools like LLMagnet can track prompts and compare what you claim with what models say, but you can start manually.

Ben cohen

And what not to do: don't stuff keywords for AI, don't publish hundreds of thin pages, don't treat one file as magic, and don't make the human experience worse. Also, don't measure citations alone. Check accuracy and relevance too.

Ido

One prediction, as an opinion: in two to three years, sites will routinely serve different representations of the same content to humans and agents.

Ben cohen

Mine: shortlists will decide more deals than clicks, so companies will track visibility the way they once tracked rankings.

Ido

Here's your action. Pick five to ten buyer prompts. Mix factual ones, like pricing and specs, with comparison ones. Ask ChatGPT, Claude, and Perplexity. Note whether they use your site, cite third parties, or get you wrong.

Ben cohen

The goal isn't just visits. It's making sure machines can find, understand, trust, and eventually use your brand correctly. If this helped, follow Be Visible and send it to someone building a website.