Type ‘best landscape architect for coastal stormwater design in Auckland’ into ChatGPT this morning. You get a confident paragraph back. Three firms named, a sentence each about what they are known for, a note about a recent project. The reader closes the tab. No link was clicked. The model did the work the firm’s website used to do, and the firm has no record of the conversation happening.
A new layer has appeared between the reader and the firm. It reads everything published, summarises it, and almost never sends traffic back. Google AI Overviews now appear above more than a quarter of all searches, up from around 13 percent twelve months earlier. A randomised field experiment running through late 2025 and into early 2026 found that when an Overview appeared, outbound clicks on the underlying queries fell by 38 percent, and the share of searches that ended without any click at all rose from 54 percent to 72 percent.
The traffic isn’t there
Chartbeat’s analysis of publisher traffic shows chatbots accounting for less than 1 percent of all page view referrals as of late 2025. ChatGPT referrals are growing by more than 200 percent year on year, but the absolute numbers are quiet. For queries that look like search, ChatGPT sends roughly 1.3 percent of users to an outside link. The equivalent number for a Google search result is around 29 percent.
The shape of this on any individual firm’s analytics looks like the line below.
A small practice that had a steady drip of organic search visits two years ago now has a fraction of that. The pages are still there. The work is still good. The route to the reader has changed.
Optimising for the model is a shrinking pie
A small industry has formed around ‘AI-optimised content’, offering advice about question-shaped headings, citation-friendly paragraph structure, and `llms.txt` files. Some of it is reasonable. Most of it repeats the trap of the last fifteen years of SEO: chasing the algorithm, writing for the machine, and watching the machine change.
The deeper problem is structural. Search engines refer. ChatGPT answers. LLMScout’s analysis of 15,252 AI queries between September 2025 and January 2026 shows links per response falling by half on Claude, Perplexity, and Gemini, with ChatGPT down too. Even if the optimisation worked, the pie is shrinking.
The infrastructure of outbound links inside AI answers is contracting. Models are increasingly happy to answer from absorbed knowledge rather than point the reader anywhere. A firm hoping to be cited is hoping for a share of a smaller pie, served to readers who mostly do not click through.
What works: a list
The firms holding up best through this shift are the ones who already had a direct line to their readers. An email list of a few hundred procurement officers, council planners, developers, and architects who hear from the practice once a month, with a short piece on a project, a problem, or a piece of fieldwork.
The newsletter is the part of the marketing system that does not depend on a search engine, a social platform, or a model deciding to send anyone anywhere. The reader subscribed. The email arrives. The relationship is direct.
A list of 800 named people inside a firm’s niche is worth more than ten thousand organic visits a year, because the list shows up on a schedule the firm controls. Procurement decisions take months. Public infrastructure projects take years. A monthly email keeps the practice in front of the people who eventually write the brief.
The firms that do best from here will be the ones who own the direct line to the reader, whatever the search bar decides to do next.
The first step is small. A signup form near the top of the site, with one honest line about what the email is and how often it arrives. A short first email when somebody subscribes. One more a month after that, written by the principal, about something the practice actually did or thought about. No automation, no funnel, no drip sequence pretending to be a person.
AI search is real and the click-through rates are not coming back. The firms that figure this out first will not be the ones who chase the algorithm into its next form. They will be the ones who, while the search bar reshaped itself, were already writing directly to a few hundred people who asked to hear from them.