# Your site, ready for the agentic web > Agents already read, choose and buy. The checks measure how they read you, the fixes close the gaps, and a real agent proves it by finishing a customer's errand, with a replay that is yours to publish. Measure, watch, fix, prove: one loop. *Built for the machine-readable web* ## How it works ### One site. Every signal. Explore the full loop: discover your site, measure readiness, test agents, track AI answers, generate fixes and verify improvements. Select a stage to explore. ### Turn a domain into a map. Discover your pages, group them into templates and identify your industry. See the site an agent actually has to navigate. ### Find the friction before an agent does. Measure technical readiness with weighted checks and page-level evidence. Industry plugins ask the extra questions your business needs, beside the AX score. ### Give an agent a job. Keep the proof. Run real customer tasks, inspect every step and see where agents succeed or get stuck. Replay the evidence and verify the result after a fix. ### See who makes it into the answer. Ask AI assistants the questions your buyers ask. Track mentions, inspect citations and compare your presence with competitors across answer engines. ### From a finding to a file you can ship. Prioritize issues and generate fixes grounded in your pages. Take the work into your editor with MCP or automate it through the API. ### Prove the improvement. Keep watching. Rescan, compare before and after, and rerun agent tasks. Scheduled scans and regression alerts keep your team informed as your site changes. ### It knows what kind of site it is measuring The kind of site is detected from the shape your URLs already have and confirmed by a model, then shown, explained, and correctable by hand. A wrong guess about your business is yours to overrule, and nothing overwrites your answer on its own. ### A site is pages, not a URL Your sitemap collapses into a few dozen page templates: 259 URLs under one shape are one page type, not 259 pages. Discovery derives the templates, a model names them, and the exemplars that stand for most of the site are the ones measured, so the score describes your site, not your home page. ### Whether the links still resolve Anyone can check that a file exists. We follow what it says. Every link your llms.txt publishes is fetched to see whether it still answers, and the templates the file never mentions are reported by name. Coverage runs in both directions. ## The scorecard ### Every check, honestly weighted Each check carries the weight of how much it changes what an agent can do: an llms.txt outweighs a sitemap because an agent leans on it harder. Some run without moving the score, and say so: rewarding a file that lies is the one thing a measurement must not do. This list is read from the engine itself, weights and all, so it grows when the engine does. - **HTML Rendering** (weight 11): How to fix HTML rendering for agents that do not run JavaScript: SPA-shell symptoms, semantic landmarks, a single H1, alt coverage, and a noscript fallback. - **Robots.txt** (weight 9): How to configure robots.txt for AI crawlers: the missing file, unconfigured crawlers, wildcard blocks, and Sitemap directives. - **Agent Access** (weight 9): AI crawler user-agents blocked by your WAF despite robots.txt allowing them, or served reduced content. Align firewall rules with your declared crawler policy. - **Agent Operability** (weight 7): How to make a page an agent can operate rather than only read: real controls, accessible names, labelled fields, and the entry blockers that hide everything behind them. - **Structured Data** (weight 6): How to add JSON-LD structured data for AI agents: missing blocks, invalid JSON, schema.org context, @graph, entity types, and BreadcrumbList. - **AI Directives** (weight 6): How the page-level AI directives work: noindex, nosnippet, max-snippet, noarchive, nocache, data-nosnippet, and why Google-Extended does not do what most people think. - **LLMs.txt** (weight 5): How to create a spec-compliant /llms.txt file: the missing file, headings, blockquotes, sections, links, and llms-full.txt. - **SEO Basics** (weight 5): Foundational head-tag hygiene every agent uses: title length, meta description, canonical URL, html lang, charset, viewport, and hreflang. - **OpenAPI Spec** (weight 4): How to create an OpenAPI specification for AI agents: the missing file, version, info, paths, and servers. - **HTTP Headers** (weight 4): How to configure HTTP headers for AI agents: security headers, AI discovery Link headers, and CORS for .well-known resources. - **Content Negotiation** (weight 4): Serve Markdown to AI agents via Accept: text/markdown, send Vary: Accept, avoid relabeled HTML, and advertise rel=alternate fallbacks. - **Usage Policy** (weight 4): How to state your AI usage terms in a way machines can read, and how to stop five mechanisms contradicting each other: Content Signals, AIPREF, RSL, TDMRep and robots meta. - **HTTP Hygiene** (weight 4): How to make your server tell an agent the truth: soft 404s, redirect chains, refused HEAD, missing charset, and 429 without Retry-After. - **Agent Card (A2A)** (weight 3): How to create an A2A agent.json card: the missing file, invalid JSON, required fields, skills, protocol version, and optional fields. - **Security.txt** (weight 3): How to create a security.txt file per RFC 9116: the missing file, required fields, expired dates, and optional fields. - **MCP (Model Context Protocol)** (weight 3): How to create a Model Context Protocol configuration: the missing file, server metadata, tools, resources, and CORS. - **TLS / HTTPS** (weight 3): Serve over HTTPS, redirect HTTP to HTTPS, and configure HSTS with preload and includeSubDomains. Agents downgrade trust on insecure origins. - **Sitemap** (weight 2): How to fix your sitemap: missing sitemap.xml, invalid XML, empty urlset, missing lastmod, stale entries, sitemap-index handling, content type, and the 50k URL limit. - **RSL License** (weight 2): Declare Really Simple Licensing terms via a robots.txt License directive and publish a valid license.xml with the right namespace, permits vocabulary, and payment types. - **Crawl Efficiency** (weight 2): Enable Brotli or gzip compression, support conditional GET with ETag and Last-Modified, and keep pages a size an agent's budget survives. - **Agent Skills** (weight 2): How to publish agent skills a site can be operated by: the index, SKILL.md frontmatter, digests, and the naming rules the specification enforces. - **Meta Tags** (weight 1): How to add AI meta tags, rel=alternate links, rel=me identity links, and Open Graph tags. - **Auth Discovery** (weight 1): How an agent finds its way past a 401: RFC 9728 protected-resource metadata, RFC 8414 server metadata, PKCE and dynamic registration. - **AI Catalog** (reports, does not score): How to publish an agent resource catalog: ai-catalog.json and ard.json, entry types, and why an entry pointing at nothing is worse than no catalog. - **WebMCP** (reports, does not score): How to declare a form as a callable tool with WebMCP: toolname, tooldescription, parameter descriptions, and the deprecated namespace. - **Commerce Discovery** (reports, does not score): How to publish a UCP commerce profile so an agent can complete a purchase: services, schemas, payment handlers and keys. ## What the subscription is for ### The three things a free scan will not do for you The scan is free, and it stays free. What a subscription buys is the work that begins once the number lands: the missing file written from your own pages, the rival measured in the same breath as you, the regression that reaches your inbox before it reaches a customer. - **One-click autofix**: One click turns a finding into the file that resolves it, generated from your own templates and handed back ready to ship. - **Industry plugins**: The questions your kind of site owes an agent, asked and scored beside the core number, never folded into it. - **Competitor comparison**: The sites you lose deals to, measured in the same run as yours, so the gap you read is the gap that exists. ## The product ### Ten rooms, each with its own page Every watched site gets the same ten rooms, and each one is explained in full on its own page: what it does, how it stays honest, and where its limits are. - [The overview of a watched site](https://axrush.com/about/overview) - [Pages: the whole site, not the home page](https://axrush.com/about/pages) - [What each industry plugin asks](https://axrush.com/about/plugins) - [What an agent task measures](https://axrush.com/about/agent-tasks) - [Share of answer, measured](https://axrush.com/about/answers) - [What we write for you](https://axrush.com/about/optimizations) - [Competitors, measured beside you](https://axrush.com/about/competitors) - [Alerts: when it moved, and what moved it](https://axrush.com/about/alerts) - [Scans: the record, kept and exportable](https://axrush.com/about/scans) - [Scheduled scans: measured on its own](https://axrush.com/about/schedules) ## The autofix ### Then it writes the files for you The same measurements generate what your site should be publishing (/llms.txt, /llms-full.txt, /.well-known/ai-catalog.json, /.well-known/agent-card.json, /.well-known/mcp/server-card.json), derived from your own templates and their labels, deterministic, and diffable against whatever you serve today. Every link in them is a page that exists. It will also tell you not to publish something. Where a value cannot be read from your site, it is not invented; and a file that advertises a service you do not run (an MCP server, an API) raises your score and makes your site worse. The generator says so instead of writing it. - [What we write, and where it stops](https://axrush.com/about/optimizations) ## Industry plugins ### Industry plugins: beside the score, never inside it What a shop owes an agent is not what a publisher owes one. Plugins ask the extra questions for your kind of site and answer beside the core score, so installing one never reads as a regression. Not applicable is an answer, not a failure. A plugin whose every check is met adds a green plus after the score: 100 with two perfect plugins reads 100++. One that falls short keeps its plus in amber or red, so an installed promise never disappears from the sentence. - **E-commerce**: Whether an agent could actually buy from you: products and offers with price, availability and currency, legible without executing JavaScript. - **Publisher**: Whether an agent could cite you: articles with authors and dates, and structure that carries the meaning instead of the styling. - **Portfolio**: Whether an agent could put you forward: who you are, what you do and how to reach you, stated where a machine can read them. - **Social presence**: Whether your presence elsewhere is discoverable from your own pages, and says the same thing everywhere it appears. - **Local presence**: Whether an agent could send somebody to you: address, hours, booking and a pin a map accepts, declared where a machine reads them. - **Docs & product**: Whether an agent could learn and operate your product: documentation, pricing, changelog and support, each one door from the front page. - **Organization identity**: Whether an agent could tell you from a lookalike: legal identity, typed contact points, and the profiles that are actually yours. - **Listings**: Whether an agent could read one of your listings as data: typed, fielded, priced, dated and canonical. ## The proof ### Everything else measures how your site is built. This measures whether an agent can use it. We send a real agent to your site with the errand a customer would give one: find this product's price, reach the checkout, say when you open. It reads your pages the way an agent does, with no JavaScript and no privileges, and we keep every step it took. What comes back is not a score. It is a recording. - **Measured, not claimed**: The same task, run several times, reported as a rate with the count beside it. Something with a temperature was asked to do a job, so the honest unit is a frequency and not a tick. - **The answer nobody else finds**: A price an agent reports that is on none of your pages is named as a finding, one no checklist can see. - **Then prove the fix**: A failure carries its diagnosis and the file or block that resolves it, and the rerun is recorded beside it. The task success rate sits beside the AX score, never inside it. ## The other outcome ### A site can be flawless and still go unnamed. This measures the part you do not control. The buyer who used to open a search box now asks an assistant, and gets back a handful of names. Share of answer puts your category's real questions to the engines, each through its vendor's own search-grounded API, and keeps what came back word for word, together with the sources it cited. The verdict is arithmetic over that text: cited and at what rank, named with no link, or absent. Passes are bought one at a time, and every paid site gets one free. Asked 47 times across 3 engines on September 6, 2026, no answer named axrush.com. 4 other domains were. We publish our own zero for the same reason we publish our own score. - [How share of answer works](https://axrush.com/about/answers) ## After the fix ### Then it keeps watching Sites drift: a deploy drops a header, a CMS rewrites a template, a file goes stale. The point of measuring twice is what moved between. - **Schedules**: Daily, weekly or monthly per site, promoted into the scan queue without you pressing anything. - **Trends**: Every check sparklined on the same 0–100 axis, so a movement is a movement and not a rescaled one. - **A changelog, not a delta**: Never “−20” when it can say which check fell, on which template, and what the finding was. - **Alerts**: Rules you write: score drops, a check regresses, a competitor moves. Delivered by email, with delivery actually confirmed. ### Measured against the people you lose deals to Competitors are proposed from what your site says it is, verified to exist before anything claims they do, and measured at the same moment as you. A comparison against last week's number is not a comparison. ## It runs on its own advice ### Built for agents, reachable by agents The product takes its own advice. Every file it grades sites on, it publishes; every service those files advertise, it runs. The specification is tested against the routes, so a documented endpoint is an implemented one. - `POST /api/v1/scan · GET /api/v1/report`: Measure a domain, or read a stored report, over plain HTTP. - `POST /mcp`: An MCP server an agent can call: 4 anonymous tools over streamable HTTP, and 17 more with an organization key. - `POST /a2a`: An A2A endpoint with an agent card, for the agents that speak it. axrush.com is scanned by its own engine in production, and the report is public, whatever it currently says. - [/report/axrush.com](https://axrush.com/report/axrush.com) - [How an agent task works](https://axrush.com/about/agent-tasks) - [How share of answer works](https://axrush.com/about/answers) - [What each plugin asks](https://axrush.com/about/plugins) ## The state of the web ### What the leaders of six industries score Market leaders rather than best performers: sites everyone has heard of, measured on the same 26 checks as yours and rescanned every week. Each score opens the report it came from. ### Online shop - [ikea.com](https://axrush.com/report/ikea.com): 75/100 - [amazon.com](https://axrush.com/report/amazon.com): 49/100 - [etsy.com](https://axrush.com/report/etsy.com): 46/100 - [ebay.com](https://axrush.com/report/ebay.com): 40/100 - [zara.com](https://axrush.com/report/zara.com): 37/100 ### Software product - [stripe.com](https://axrush.com/report/stripe.com): 75/100 - [figma.com](https://axrush.com/report/figma.com): 68/100 - [notion.so](https://axrush.com/report/notion.so): 68/100 - [slack.com](https://axrush.com/report/slack.com): 63/100 - [atlassian.com](https://axrush.com/report/atlassian.com): 25/100 ### News outlet 1 site withdrawn on request - [nytimes.com](https://axrush.com/report/nytimes.com): 65/100 - [theguardian.com](https://axrush.com/report/theguardian.com): 64/100 - [bbc.com](https://axrush.com/report/bbc.com): 62/100 - [apnews.com](https://axrush.com/report/apnews.com): 36/100 ### Travel and tours - [booking.com](https://axrush.com/report/booking.com): 55/100 - [marriott.com](https://axrush.com/report/marriott.com): 48/100 - [expedia.com](https://axrush.com/report/expedia.com): 48/100 - [tripadvisor.com](https://axrush.com/report/tripadvisor.com): 43/100 ### Financial services - [paypal.com](https://axrush.com/report/paypal.com): 72/100 - [chase.com](https://axrush.com/report/chase.com): 64/100 - [visa.com](https://axrush.com/report/visa.com): 62/100 - [mastercard.com](https://axrush.com/report/mastercard.com): 36/100 ### Government or public body - [gov.uk](https://axrush.com/report/gov.uk): 71/100 - [irs.gov](https://axrush.com/report/irs.gov): 63/100 - [canada.ca](https://axrush.com/report/canada.ca): 57/100 - [europa.eu](https://axrush.com/report/europa.eu): 57/100 - [whitehouse.gov](https://axrush.com/report/whitehouse.gov): 8/100 Measured between Sep 4, 2026 and Sep 11, 2026, and rescanned every week. A score describes a site on the day it was read. Run one of these sites and would rather it was not here? Write to us, or say so in your robots.txt: a rule for AXRush takes the row down by itself, and the scanner stops coming. - [The whole index, with the method behind it](https://axrush.com/benchmark) ## Measured, not claimed ### Measured, not claimed Every report is a real measurement at a stable URL, including ours. We publish our own score whatever it says. ## Pricing ### Start on the free tier. Subscribe when a site is worth watching. Scanning anything, anonymously, is free and stays free. What you pay for is a site measured daily, and everything the product does about what it finds; watching one site on a weekly rescan costs nothing. - [See the price](https://axrush.com/pricing) ## Measure a site No account, no email, about eight seconds. - [Scan a domain](https://axrush.com/) The thesis behind all of this, argued from data: [Read the manifesto](https://axrush.com/manifesto) --- Markdown representation of https://axrush.com/for-websites