Which Websites Survive AI Search and Which Ones Don't
The sites that come through this intact all have one thing in common. They own something a language model can’t just restate. Everything else is downstream of that.
Start with the damage, because it’s worse than the averages suggest and it isn’t distributed evenly. Over roughly two years, small publishers lost about 60% of their search referrals. Mid-sized sites lost around 47%. The big ones lost 22%. Google Search page views across the web fell about a third in a single year. Referrals from chatbots grew fast in percentage terms and still sit below 1% of total publisher traffic, which means nothing is arriving to replace what left.
The regressive shape of that is the part marketers keep missing. This isn’t a rising tide problem where everyone gets a little wet. Scale is the insulation. If you’re small, you’re taking three times the hit that a major brand is taking, and you have none of the direct traffic to fall back on.
What actually holds up
Community and user-generated surfaces. Forums, review platforms, Q&A, anywhere real people describe what happened to them. Models route opinion, comparison and use-case questions toward this stuff because a brand page tells you what a product does and a thread tells you how it went. These sites get pulled into answers rather than replaced by them.
Original data. Testing you ran. Numbers you collected. Photographs you took. A model can summarize your findings but it has to say where they came from, and readers who want the underlying material still have to come to you.
Direct audience relationships. Email lists, apps, subscriptions, the bookmark. Traffic that never touched a search results page can’t be intercepted on a search results page. Boring, unglamorous, and the single most durable asset in this whole mess.
Transactional and local intent. Booking a table, buying a part, finding the nearest place that’s open. An answer isn’t the outcome the user wants. The click is the point.
Tools. Calculators, configurators, converters, anything interactive. Hard to summarize, easy to bookmark.
Brand. Not brand as a vibe, brand as the thing that makes someone type your name into a box instead of describing what they want and taking whatever comes back.
What doesn’t
Explainers. Definitions. Listicles. “What is X” pages. Roundups of publicly available information, dressed up. Any page whose entire contribution is packaging facts that exist elsewhere in less convenient form. That convenience was the whole product, and something else does it better now, right there in the results.
The behavioural data is blunt about this. When an AI answer appears at the top, roughly one in four users ends the session entirely. That traffic didn’t go to a competitor. It stopped existing.
General news commentary sits in an awkward middle. Reporting survives. Reaction to reporting mostly doesn’t.
What to do about it, if you’re small
Pick one thing you can produce that nobody else has. It doesn’t have to be big. A recurring test, a dataset you maintain, a corner of a market you actually watch. Then build the direct channel while you still have traffic to convert, because a list you build at 40% of your old traffic is still a list, and a list you try to build at 8% isn’t.
Two harder questions worth sitting with.
The first is whether your model was ever a business or just an arbitrage on Google’s inefficiency. Plenty of content sites were the latter and the honest version of this conversation admits it. The window closed. That’s not a moral judgment, it’s an inventory check.
The second is about who’s buying. If you’re holding web assets, the buyer pool matters as much as the traffic. Anyone acquiring a site or a domain to run a content play into organic search is a shrinking category of buyer. Anyone acquiring a name because they want to be called that is not. Short, literal, memorable names that people say out loud are holding value. Keyword strings assembled to rank are the ones under pressure.
Direct traffic and brand recall were always the fundamentals. Search made it possible to succeed without either for about twenty years. Now it doesn’t.