The usual reason teams give for caring about AI readability is ranking: they want to be the site an assistant names, on the theory that a mention produces a click and the click produces revenue. That reason is not wrong so much as it is the wrong shape. It treats an agent as a search engine with unfamiliar ranking signals, and it leads teams to optimize for placement in a list. The more useful way to look at it is that retrieval happens before ranking, and retrieval is not a gradient. Either the agent can fetch and parse your page or it cannot, and if it cannot, there is no list to be ranked in.
The scale of the channel is worth stating plainly, because both the hype and the dismissal are easy to find. AI platforms drove around 1.13 billion referral visits in June 2025, roughly 357% more than a year earlier, and ChatGPT accounts for the large majority of that flow (Digiday). At the same time, AI referrals remain well under one percent of total web traffic (SE Ranking). So this is not yet a channel that replaces search, and anyone telling you to rebuild your site around it this quarter is selling something. It is, however, growing at a rate that makes the cost of being unreadable compound quietly.
The mechanism that decides whether you are in the channel is narrower than most SEO advice implies, and one finding illustrates it better than any other. Vercel and MERJ instrumented roughly 500 million AI crawler fetches and found no JavaScript execution at all: the crawlers for ChatGPT and Claude will fetch your JavaScript files, but they do not run them (Vercel). Meaning a fully client-rendered site returns what amounts to an empty shell to those systems. The page a human sees and the page the agent sees are different documents, and only one of them has your content in it. Google's own crawler renders JavaScript, and Gemini inherits that infrastructure, which is exactly why a site can look healthy in Search Console while being invisible to everything else.
This is the part that makes the gap hard to notice from the inside. A ranking drop shows up in a dashboard. A retrieval failure does not, due to the fact that an agent which cannot parse your page never generates an impression, a click, or a referrer for you to count. The absence looks identical to a channel that simply has not grown yet. As a result, the failure mode is not a site that knows it is losing AI traffic, it is a site that concludes AI traffic is not real because none of it is arriving.
Retrieval also has more than one gate, and they are sequential rather than additive. The agent has to be permitted to fetch, which is a robots and WAF question. It has to receive real content, which is a rendering question. It has to be able to identify facts on the page, which is a structured-data question. It has to be able to find pages nobody linked, which is a sitemap question. Failing any single gate ends the chain, and a strong score on the remaining ones does not compensate. That is why MachineRead reports a strict agent-readiness lens separately from the point total: 8 default probes out of 21 in full scope, scored as gates rather than blended into an average that would let a good SEO score hide a fatal fetch failure.
So here is what I project. The gates are worth checking in order, starting with whether an agent can fetch and read your page at all, since everything downstream is conditional on it. If your site is client-rendered, that is the finding that matters most, and server-side rendering or prerendering for crawlers is the fix. Treat structured data and discovery files as the second layer, worth doing but pointless if the first layer fails. And judge the whole effort on the honest timeline: you are not buying this quarter's traffic, you are making sure that when the channel is large enough to matter, your site is legible to it.
See also
- Launch post - what the audit measures and what it refuses to claim
- Methodology reference - the 13 check groups and how the gates are scored
- Related: bot access and schema.org JSON-LD