Just because your page ranks first on Google does not mean that the AI agent can read it. Detection tools like isitagentready ask another question: when the crawler behind ChatGPT, Perplexity, Claude or Google AI Overviews crawls your website, can it cleanly get the text you want to be quoted without executing JavaScript or guessing the layout. We run this test for our clients, and it is common for pages in the top rankings to receive failing scores. The problem almost always lies in the machine reading layer, not the quality of the content.
isitagentready detects machine reading, not human eye reading
Traditional SEO tools look at keyword density, external links, and page speed scores. Tools like isitagentready change the perspective: think of yourself as an AI agent that does not have a browser and only makes HTTP requests, and see what it can get back from your page. The difference is critical. When a user opens a page with a browser, JavaScript will be executed, images will be loaded, and the layout will be arranged; however, most AI crawlers capture the original HTML and do not wait for your front-end framework to finish running. If your main content relies on JavaScript to render, the agent may be left with an almost blank skeleton.
This is why a website with high scores in Lighthouse may not be suitable for being cited by AI at all. Fast speed and smooth interaction are indicators for people to see; whether it can be cleanly extracted by a machine is another set of standards. The following eight items are the skeleton of this standard.
Eight AI Agent Readability Metrics
- Crawler access rights: whether robots.txt and llms.txt allow AI agents such as GPTBot, ClaudeBot, PerplexityBot, and Google-Extended.
- Server-side rendering: The main content exists in the original HTML, without executing JavaScript.
- HTTP health: whether the response code is 200, whether there are redundant redirects, and whether the canonical is correct.
- Semantic HTML: Whether the title hierarchy is reasonable and whether semantic tags such as article, main, nav, etc. are used.
- Structured data: Whether to mark Article, FAQPage, Organization and other schemas in JSON-LD.
- Metadata: title, meta description, Open Graph are complete and the description is accurate.
- Extractable answer structure: Whether key information is presented in easily extractable forms such as questions and answers, definitions, tables, etc.
- Markdown friendliness: Whether the main content can be cleanly converted into plain text or Markdown, without navigation and style noise.
The first three items determine whether you will be seen.
The number one metric is crawler access because it is a veto. If your robots.txt blocks GPTBot, or the CDN you use blocks non-browser User-Agent by default, then it will be meaningless no matter how well you do the last seven items, and the agent will not be able to get in at all. The correction method is very straightforward: check whether robots.txt clearly allows mainstream AI crawlers, and add a copy of llms.txt, listing the pages and brief descriptions you most want to be cited in plain text. Many companies don’t even know they’ve blocked AI crawlers, and this is the problem we most often discover on day one.
The second item is server-side rendering. Use "View Source" instead of "Inspect Element" to open the page. If you can't see the text, the agent probably can't either. Single-page applications (SPA) are the worst-hit areas. All content is loaded by the front end, and the original HTML only has an empty div. The solution is to import server-side rendering or static generation so that the text is delivered completely in the first response. The third item of HTTP health is to clean up the foundation: confirm that the page returns 200 instead of soft 404, remove layers of redirects, let canonical point to the only correct URL, and prevent the proxy from catching duplicate or expired versions.
The middle three determine whether the content will be understood.
After getting in, the agent must be able to distinguish which paragraph is the title, which paragraph is the main text, and which paragraph is the navigation. Semantic HTML gives it this map: there is only one h1 on a page, there is no jump in the title level, the main content is included in main or article, and the sidebar and footer are marked with corresponding tags. Using a series of semantically meaningless divs for layout is incomprehensible to the human eye, but machines will only see a bunch of homogeneous blocks. Add a layer of clear annotation to the structured data, use JSON-LD to tell the agent "this is an article, who is the author, and the date of publication", or use FAQPage to mark question and answer pairs. This step is to change "let the machine guess" into "tell the machine directly".
The sixth metadata is often regarded as a trivial matter, but it is the agent's first clue to determine the topic of the page. The title should accurately describe what this page is about, rather than being stuffed with keywords; the meta description should be able to be quoted as a sentence alone; the Open Graph tag determines what your link will look like when it is paraphrased. These three fields add up to less than 20 lines of code, but they directly affect whether the agent is willing to include your page in the answer.
The last two determine whether you will be cited.
Being understood is not enough, the agent must be able to extract a paragraph and use it directly. The extractable answer structure refers to: writing the key conclusion into a complete and self-sufficient sentence, preferably immediately after the corresponding question or subscript, rather than appearing after three paragraphs of elaboration. Tables, definitions, and step lists are all forms particularly preferred by agents because they have clear boundaries and are easy to cut. Markdown friendliness is the final quality check: after converting the page to plain text, whether the main text can still be read smoothly, or whether it is fragmented by the navigation bar, advertisements, and social buttons. The higher the ratio of content to noise, the higher the chance of being cited cleanly.
After testing, which one should be repaired first?
It is not necessary to practice the eight items at the same time. The sequence is based on the three levels of "being seen, being understood, and being cited": first solve the two one-vote items of crawler access and server-side rendering, then add structured data and semantic tags, and finally polish the answer structure. When we do GEO audits for B2B SaaS customers, the first two layers can usually bring most pages to the passing mark within thirty days, and these two layers often bring about the fastest changes in visibility, because they solve the flaw that "agents can't read it at all."
isitagentready gives you a score, but the score does not tell you which item is most critical to your business or how much resources should be invested in repairing it. If you have already run the test but are not sure which one to start with, you can make an appointment for a thirty-minute GEO diagnosis. We will use your actual website address to run through these eight items and point out the gaps that should be fixed first.
