A growing share of the visitors to any website are no longer people. They are AI assistants fetching pages on someone's behalf, and unlike a person with a browser, an agent either can parse your site or it cannot. There is no partial credit. The traffic those assistants send is still small but climbing quickly: AI platforms drove roughly 1.13 billion referral visits in June 2025, up about 357% from the year before (Digiday). That is under half a percent of all web traffic today, so this is not the channel that pays your bills yet. However, it is compounding faster than traditional search ever did, and the sites that appear in it are the ones a model can actually fetch, read, and cite. MachineRead exists to tell you whether yours is one of them.
MachineRead is an audit tool. You enter a public URL and get back a structured report on the signals that AI agents, retrieval systems, and search crawlers use to discover, access, and understand a site. The Essentials audit is free to run and needs no account, because the checks behind it cost me nothing to perform, so there is no honest reason to put them behind a login.
Before the numbers, one design decision worth stating up front. Most audit tools blur measurement into recommendation, and the result is a report that tells you to fix everything without telling you which gaps actually block retrieval. MachineRead keeps those two things apart: it measures the signals, scores them, and leaves the prioritization legible so you can see what is load-bearing and what is merely nice to have.
The audit runs 13 check groups covering 56 checked points across three pillars: off-site signals, AI access and scrapability, and traditional SEO. Nine additional advanced rows are reserved but locked, meaning they are shown as zero-scored coverage gaps rather than quietly dropped or guessed at, because they require paid crawling and LLM verification that Essentials does not perform. On top of the point score, a separate strict agent-readiness lens exposes 8 default probes out of 21 full-scope probes, so the free audit already catches the majority of agent-native discovery signals. Four agent-consumable surfaces are checked directly: llms.txt, ai-catalog.json, api-catalog.json, and openapi.json.
The reason this is a distinct problem from SEO is that ranking on Google and being retrievable by an agent are not the same capability, and the gap between them is where sites lose the AI channel without noticing. A page can rank well in search and still fail agent retrieval, due to missing structured data, a blocked crawler, or an API with no machine-readable contract. The audit quantifies that gap in exact check names and counts rather than a single grade, because a number you can act on is worth more than a letter you cannot.
It is equally important to be clear about what MachineRead does not do. It does not measure your actual search ranking, traffic, backlinks, social traction, Core Web Vitals field data, model citation share, or conversion. It does not run paid crawlers, paid LLM calls, or browser-agent task simulation. It does not claim to know what an agent will do with your content once it has retrieved it. Essentials measures the signals that constrain retrieval, not the outcomes downstream of it. That boundary is deliberate, and holding it is what lets the report stay honest about the limits of what it can tell you.
Essentials runs on a free-tier stack, which is a design parameter rather than a limitation to apologize for. The backend is a FastAPI app on Render's free tier, the frontend is a static export on Cloudflare Pages, and no paid API is called during a free audit, which keeps the running cost of the current product, mail included, under the $20 a month I set as the ceiling. As a result the audit is fast and bounded, and you can run it against your own site right now. The same posture is why Essentials does not attempt deep crawling, JavaScript rendering, or LLM content analysis, since those require paid infrastructure.
That last point is also the honest answer to where this is going. The 9 locked rows are locked because verifying them costs real money per audit, and when those Starter and Pro checks ship they will need accounts and payment to cover that cost. Accounts will be OAuth only, so there is no password for me to store or leak, and payment will run through Stripe and Link, so there are no card details on my side either. What will not change is tracking. There is none today and there will be none later, regardless of tier.
As a first test, I pointed the audit at MachineRead's own static export. It returned 5 pass, 5 partial, 3 fail-deferred, and 9 locked. The 3 fail-deferred rows are transport artifacts of testing against a local HTTP server rather than product defects, and they are documented as such. I am reporting the failures rather than hiding them, because a tool that cannot detect its own coverage gaps has no standing to audit anyone else's.
So here is what I recommend. Enter your URL, pick the preset that matches your site type, and read the report for the gaps it surfaces rather than the score it prints. The score is a convenience. The gap inventory is the product. If your site fails the agent-readiness probes, start there, because those are the binary gates. If it passes them, the locked advanced rows show you what deeper coverage would add. Either way it runs in under a minute.
For the full methodology, see the methodology reference. For agent integration details, see the agent integration guide. Run the audit.