Due to the rapid adoption of retrieval-augmented generation, the question of whether an AI agent can read your site has become a question of whether your site exists at all in the channels where users increasingly look for answers. The stated reason most teams care about this is search ranking impact and traffic generation. However, this framing misses the mechanism. The real reason is that retrieval-augmented agents are now a primary discovery channel, and a site that cannot be read by machines cannot be discovered through that channel.
This post covers the conventional framing, why it is wrong, the actual mechanism, and what the cost of inaction looks like in concrete numbers. The reason for this structure is that the gap between the stated concern and the real mechanism is where the actionable work is.
The conventional reason teams invest in AI readability is that they want to rank in AI-generated answers. The assumption is that if a site is mentioned by ChatGPT, Perplexity, or another retrieval system, traffic will follow. As a result, teams treat AI readability as an extension of SEO: optimize for the model, get cited, get traffic. This is the surface claim. However, this framing misses the mechanism, because it treats the agent as a search engine that happens to use different ranking signals. The agent is not a search engine. It is a retrieval system that fetches, parses, and synthesizes content in real time, and it is constrained by a different set of plumbing than a crawler that builds an index.
The actual mechanism: a retrieval-augmented agent needs to fetch your page, parse the HTML into usable text, extract structured facts if they exist, and follow links to discover related content. If any of these steps fail, the agent does not rank your site lower. It cannot retrieve your content at all. Meaning the site is invisible to that query, not merely ranked below a competitor. This is a binary outcome, not a gradient. A site that is indexable by Google but not retrievable by an agent has a gap that is invisible in traditional analytics, due to the fact that the agent never produces a click or an impression that shows up in a dashboard.
It is through the fetch-and-parse pipeline that the cost of missing signals becomes concrete. The MachineRead audit corpus shows that of the 13 check groups, the average site passes roughly 5 of them cleanly, with the remainder scoring partial or fail. The 56 checked points include signals like schema.org JSON-LD (present on roughly 30% of audited sites), llms.txt (present on fewer than 5%), and explicit AI-bot permission in robots.txt (present on roughly 15%). These are not edge cases. They are the majority state of the web. As a result, the majority of sites are partially or fully invisible to agent retrieval, and the site owners do not know it because their traditional search analytics still show traffic.
The cost of not having machine-readable signals is not a lower ranking. It is absence from the retrieval channel entirely. The 8 default agent-readiness probes that MachineRead runs catch the most common gaps: missing structured data, blocked bots, absent sitemaps, no canonical URL, no llms.txt, no API catalog. Of the 21 full-scope probes, the default 8 already surface the majority of blocking issues. A site that fails 4 or more of these 8 probes is effectively unreadable by the current generation of retrieval-augmented agents, regardless of how well it ranks on Google.
So here is what I recommend. Run a MachineRead audit on your site. Look at the agent-readiness score, not just the SEO score. If you fail any of the 8 default probes, fix those first, because they are the binary gates. The SEO score is a gradient you can improve over time. The agent-readiness score is a floor: below a certain threshold, your site does not exist in the retrieval channel at all.
For the launch post with the full mechanics, see Launching MachineRead.