ChatGPT sends 28.8% of its referrals to your site's search box — not to a page that answers the question

What the report found

Previsible's report — the same dataset behind the widely-quoted finding that ChatGPT drives 92.4% of standalone-LLM referral traffic, and covered by Search Engine Land — also looked at where that traffic lands. The landing-page numbers:

• ChatGPT: 28.8% of the traffic it refers goes to an internal search results page.
• All AI-referred traffic: ~25%, and the report notes this holds across industries rather than being driven by one vertical.
• SaaS: 34.6% of LLM referrals — the highest reliance on internal search in the sample.
• The report's own conclusion: “Your internal search experience is now an acquisition tool, not just a navigation feature.”

What the pattern actually means

A link to `yoursite.com/search?q=...` is a model that has made one decision and declined to make a second. It decided your domain is the right answer. It then could not identify which of your pages answers the question, so it handed the user your search box and let them finish the job. Trust at the domain level, uncertainty at the page level — in a single URL.

That split is worth sitting with, because it is the same shape as the gap we measure on the recommendation side. Across 168 local businesses in 41 US metros, the engines recognised a business in 98.8% of answers when asked about it by name — and named it in only 19.8% of answers when asked the buyer's question. Knowing you exist and knowing when to reach for you are different capabilities, and engines are far better at the first. An internal-search referral is that same gap expressed as a click: I know it's you; I don't know which part of you.

Three implications for anyone selling or buying AI visibility

• A referral-traffic dashboard is over-reporting your wins. If 28.8% of your ChatGPT sessions land on a search results page, roughly that share of your “AI traffic” is a visitor who arrived without an answer and has to go find one. Counted as a session it looks identical to a visitor who landed on the page that closes them. It is not the same event, and averaging them together hides which of your pages the engine can actually resolve.
• The fix is page-level legibility, not more pages. The engine already picked you. What it lacked was a page whose subject was unambiguous enough to link directly. A site whose service pages each answer one specific buyer question — stated plainly, in the page's own words — gives the model something to point at. A site of broad overview pages gives it a search box.
• Most local and service businesses do not have an internal search page at all. For a med spa, a law firm or a home-services company, there is no `/search` to fall back on. The engine's uncertainty resolves the other way: it names the business without linking, links a directory instead, or names someone else. That is invisible in a traffic report — the click that never happened leaves no row.

What this does not say

• It is not a measure of being recommended. This is a referral-traffic study: it can only see AI answers that both linked out and got clicked. It says nothing about the answers where the engine named a business and no one clicked anything.
• Google's AI Overviews and AI Mode are excluded from the dataset entirely — Previsible is explicit about this, and notes they likely drive more AI traffic than every standalone tool combined. The largest surface is not in these numbers.
• The 166 properties are a sample, not a census, and skew toward sites that have internal search in the first place. A site without one cannot contribute to a 28.8%.

The takeaway

Nearly a third of the clicks ChatGPT sends are the engine saying I trust this site and I can't tell which page you want. That is a solvable problem, and a different one from not being visible at all — but you cannot tell the two apart from a traffic report, because a session that lands on a dead-end search page and a session that lands on the page that converts look the same in the row count.

The free 60-second check asks the engines your buyers' actual questions and shows you what they say and what they link — including the answers that never send a click at all. The $79 audit runs 25 questions across ChatGPT, Perplexity, Gemini and Claude, three times each, so you can see which of your pages the engines can resolve and which ones they cannot. Agencies auditing a book of clients can run five at once with the $249 Agency 5-pack.