AI Referrals Are Falling. Your Website's Job Just Changed

AI engines like ChatGPT, Gemini, and Google's AI Overviews now answer most early research questions on the spot, without a click. So the informational traffic your blog used to pull is shrinking, while the visitors who do land on your site show up later in their decision and closer to buying. That changes your website's job. It is no longer where buyers start their research. It is where they finish it.

This article covers what the traffic data actually shows, what it means for how your site should be built, and what to hand your web team this quarter.

What's actually happening to your traffic

The pattern is simple. When someone asks ChatGPT or Google "how much does a water heater replacement cost" or "tankless vs. tank", the AI pulls from many sources and presents a synthesized answer directly. The searcher gets what they need without visiting anyone's website. Marketers call this zero-click search, and it hits top-of-funnel educational content hardest.

That is the loss side. The other side is more interesting.

Schema App, a structured data company, recently published numbers from its own website that show both halves of the shift. As reported by Andrea Badder, Schema App's AI referral sessions to top-of-funnel content gradually declined. Over the same period, their branded search impressions rose 114% quarter over quarter and branded clicks grew 75%. Non-branded impressions rose 91% and non-branded clicks grew 50%.

Badder's reading: AI was answering more early-funnel questions before people needed to click through. The research still happened. It just happened inside AI platforms. And when those researchers eventually visited the site, they arrived with much of their discovery and evaluation already done.

One company's analytics could be a fluke. The broader data says otherwise. A study by LLM Scout analyzed 15,252 AI queries and 90,232 extracted citations across ChatGPT, Claude, Gemini and Perplexity between September 2025 and January 2026. It found that the number of outbound links included per answer is falling across the board. Claude lost nearly two-thirds of its links per answer in three months. Gemini and Perplexity each lost roughly half. ChatGPT barely moved, because it was already operating at a low-citation baseline. When every major platform moves the same direction at once, that is a change in how the products are built, not a seasonal dip you can wait out.

This also means the scoreboard most businesses use is measuring the wrong thing. Counting total sessions treats a lost top-of-funnel click as pure damage, when the same shift is producing more qualified arrivals through branded and direct traffic. Judge the site on booked jobs, calls and form fills, and track branded search separately from non-branded.

For a home-service business, the change looks like this: a homeowner asks an AI assistant what a 200-amp panel upgrade involves, roughly what it costs, and how to choose an electrician. By the time they search your company name or click through to your site, they are not learning what a panel is. They are deciding between you and two competitors.

Your next visitor already did their homework

Most local business websites are still built for a visitor starting from zero: a broad homepage, a deep services menu, educational pages meant to be found through search. That structure now serves a visitor who mostly does not show up anymore.

Design instead for the pre-informed buyer. In practice:

None of this means education has no place on your site. It means education is no longer the front door.

The technical layer: how AI engines learn to cite you

If buyers are doing their research inside AI platforms, the important question becomes: what determines whether the AI mentions your business during that research?

A large part of the answer is structured data. Optimizing for this is what people mean by answer engine optimization, or AEO, and it is a related but distinct discipline from classic SEO. Schema markup is code added to your pages that labels your content in a form machines read reliably: this is a business, here is its service area, these are its services, these are its reviews. A knowledge graph connects those labeled facts to each other, so the relationships between your business, services, locations, people, and content are explicit instead of implied by prose. If any of this terminology is new, our plain-English AI glossary covers the vocabulary.

This matters because AI engines assemble answers from sources they can parse with confidence. A page of unstructured text forces the engine to guess what your business is and does, and guessing engines tend to skip you or describe you wrong. Structured, connected data removes the guesswork.

Schema App has a concrete data point here: after strengthening entity linking, which connects related entities in their markup, their visibility in Google's AI Overviews increased 32%.

The falling citation density makes this more urgent, not less. If every AI answer now links to fewer sources, the competition to be one of those remaining sources gets sharper. Being easy for an engine to parse and confident about is how you stay in a shrinking set.

An honest caveat: markup is not a guarantee of citation. Nobody outside these platforms controls what gets cited. What structured data does is make your business legible, so that when an engine weighs sources for an answer, yours is one it can actually use. And as Badder points out, when an AI references your expertise, the recommendation can happen even when no click does. Trust starts forming before the visit.

Give ready-to-buy visitors answers, not a scavenger hunt

The second technical shift is conversational search on your own site. Visitors trained by ChatGPT expect to ask a question and get an answer. Hunting through static pages for it feels like work.

An on-site assistant grounded in your own content, meaning your services, pricing, FAQs and policies, answers the last questions people have before they buy: "Do you service my town?" "Is financing available?" "How soon can someone come out?" Schema App does this with a Knowledge Assistant built on NLWeb, drawing on the same structured content layer that feeds the AI engines. The knowledge graph does double duty: it makes you citable off-site and answerable on-site.

The caveat matters more than the feature. This is only worth doing if the assistant is grounded in your actual business information. A generic chatbot that cannot answer "do you offer weekend appointments" is worse than no chatbot. Data layer first, conversational layer second.

What to hand your web team

In order:

Where this leaves your website

AI search has taken over early-stage discovery. Your website's primary job is no longer top-of-funnel education. It is converting pre-informed buyers who are ready to take action. The businesses that come out ahead will work both sides of that journey: legible to AI engines while the buyer is researching, and frictionless for the human who arrives once the research is done.

If you want to see how visible your business currently is to search and AI engines, and where the gaps are, a free Rhize-Scan will show you in a few minutes.

Sources

• Schema App, Andrea Badder: Why AI Referrals Are Falling (And Why This Isn't All Bad News). All Schema App figures cited above are their measurements from their own website.
• LLM Scout: Why AI Referral Traffic Is Falling. Citation-density analysis across 15,252 queries and 90,232 citations, September 2025 to January 2026.