Last week, a row in our analytics crossed a line we did not expect to see so soon. The source: chatgpt.com. The medium, as Google Analytics now labels it: ai-assistant. Over the past seven days, ChatGPT referred exactly as many sessions to this site as LinkedIn did, and about a third as many as Google organic search. For a traffic source that did not exist on our reports a year ago, that is a milestone worth pausing on.
The people behind that row never typed our name into a search box. They asked an assistant a question, the assistant read the web on their behalf, composed an answer, and named us in it. The click that followed was not discovery. It was verification: a human arriving to confirm what an AI had already told them.
That, in one anecdote, is why generative engine optimization (GEO) has moved from conference-panel material to something we now treat as seriously as SEO. This post is what our own data taught us, and what we changed on this very website because of it.
The referrer nobody planned for
GA4 recently started classifying assistant-driven visits under a dedicated ai-assistant medium, which makes the trend easy to watch: open your traffic acquisition report and look for sources like chatgpt.com, perplexity.ai, or gemini.google.com. On our site, three things stood out immediately.
The volume is no longer a rounding error. Session for session, ChatGPT matched LinkedIn, a channel we actively post to. Nobody at Crewdle spends an hour a day marketing to ChatGPT.
The visitors arrive pre-qualified. A search click starts an evaluation; an assistant click ends one. These visitors asked a question, received a synthesized answer with us in it, and clicked through to check pricing, documentation, or a specific claim. They land deeper in the funnel than almost any other cold channel.
What you can measure is the floor, not the ceiling. Most assistant answers never produce a click at all: the user reads the answer and moves on, satisfied. And when an AI agent fetches your page directly, it does not execute your analytics script, so it never appears in GA4 at all. The ai-assistant row is the visible edge of a much larger shift.
Search happened in front of you. Answers happen without you.
Classic search was a negotiation you could watch: ten blue links, a ranking to climb, a click-through rate to optimize. Generative engines collapse all of that into a single composed answer built from a handful of sources. You are either one of those sources or, for that conversation, you do not exist.
This changes what winning means. In SEO, position three still gets traffic. In an assistant’s answer, there is no position three. The model either found your facts, understood them, and trusted them enough to repeat them, or it recommended someone else. GEO is the discipline of making sure it can do the former.
What GEO actually is (and is not)
GEO is not a bag of tricks to fool a model, and the fastest way to fail at it is to treat it like 2010-era keyword stuffing. Models synthesize; they do not count keyword density. What GEO comes down to is being legible, quotable, and retrievable:
- Answer-shaped content. Assistants assemble answers from passages, not pages. Sections that open with the question a buyer would ask, followed by a direct answer in the first sentence or two, get lifted into responses. Buried conclusions do not.
- Unambiguous facts, in plain text. Product names, prices, dates, and capabilities stated clearly, close together, without marketing fog. If a model has to infer what you sell, it will describe a competitor who said it plainly.
- Structured data. Article and FAQ schema (JSON-LD) gives engines a machine-readable summary of what a page claims. It is the cheapest legibility upgrade available.
- An llms.txt file. A plain-text map of your site written for language models, sitting at /llms.txt, pointing them at your canonical pages and descriptions.
- Fast, static HTML. Assistant fetchers read raw HTML on a budget. Content hidden behind client-side rendering is content a model may simply never see.
- Facts worth citing. Models quote sources that say something specific: original numbers, concrete how-tos, real positions. Generic content gets synthesized away; specific content gets named.
What we changed on this site
None of this is theoretical; you are reading a website that practices it. Every page here ships as static HTML rendered at build time, so a fetcher gets the full content in one request. Blog posts emit Article and FAQPage structured data, and end with an FAQ written in the same question-first shape assistants use. We publish an llms.txt describing the site in model-friendly plain text. And because assistants answer in the user’s language, we maintain full English and French page trees, so a question asked in French gets answered from our French pages, not from a competitor’s.
None of these changes required a platform migration or a budget line. They required deciding that a machine reading the page is a first-class audience.
The visitor after this one will not be human at all
Here is the part that makes GEO more than a marketing tactic. Today, an assistant reads your site and refers a human. The step after that, already underway, is an AI agent browsing on a human’s behalf: comparing your pricing to alternatives, reading your docs to check an integration, shortlisting vendors while its owner sleeps. We see this firsthand at Crewdle, because the agents our customers hire do exactly this kind of research as part of their jobs.
When that agent evaluates your site, there is no design, no brand video, no persuasion. There is only whether your pages state, in machine-legible terms, what you do, what it costs, and why you are credible. The businesses that are legible to models will be in every shortlist those agents produce. The ones that are not will be invisible, and will never know it, because invisible visits leave no analytics trail.
The takeaway
AI assistants are no longer a hypothetical traffic source; on this site they already match LinkedIn, and the unmeasurable share is larger still. Treating GEO as seriously as SEO costs little: watch the ai-assistant medium in GA4, restructure key pages so each section answers one question directly, state your facts plainly, ship FAQ and Article schema, publish an llms.txt, and keep your content in static HTML. Your next visitor may be a person who trusted an AI’s recommendation. The one after that may be the AI itself. It pays to be ready for both.
Ask about this article 
