We ran an AI-agent-readiness scan on 19 AI-native brands. Here is what breaks.

We ran an AI-agent-readiness scan on 19 AI-native brands. Here is what breaks.



Tópico: We ran an AI-agent-readiness scan on 19 AI-native brands. Here is what breaks.
Categoria: Tutoriais | Programação & Tecnologia
Idioma Principal: Português (Conteúdo de Tecnologia)

Descrição do Conteúdo / Informações:
-------------------------------------------------------------------------
If any group of companies should have their site ready for AI agents in 2026, it is the ones building AI. So we ran a real scan on 19 of them: Anthropic, OpenAI, Perplexity, Cloudflare, GitHub, Stripe, Vercel, Shopify, Notion, Figma, Linear, Cursor, Supabase, Pinecone, Hugging Face, LangChain, Mistral, Cohere, Groq.

Result: 0 A grades. 1 B. 7 C. 10 D. 1 F. Average pass rate 54.5% across 34 signals.

That is not "the internet is behind". That is the vendors whose own agents crawl your site failing their own homework. If you build the web today, this benchmark is a cheat sheet for what to fix first.



What we actually checked


Our public scanner runs 34 signals grouped in six buckets:

•
discoverability: robots.txt, sitemap.xml, /.well-known/, llms.txt, syndication feed

•
ai_policy: explicit rules for GPTBot, ClaudeBot, PerplexityBot, Google-Extended, content-signal declarations

•
schema: JSON-LD for Organization, WebSite, FAQPage, BreadcrumbList

•
agent_protocols: MCP server card, A2A agent card, agent-skills index

•
content: image alt coverage, markdown content negotiation, heading structure

•
browser_ux: tap-target size, aria labels (agents also crawl accessibility signals)

Every check is a real HTTP request from our infra, not a lint pass on your assumed setup. So agentfix.pro/scan returns "you claim GPTBot rules but the file 404s".



The three signals every AI-native brand failed


1. FAQPage / BreadcrumbList JSON-LD (19/19 miss)

Not a single site had FAQPage or BreadcrumbList schema on the homepage or a common product page. These two schema types are how ChatGPT, Google's AI Overviews and Perplexity build inline answer cards. Without them your content still gets read, but you rarely become the highlighted source. Cost to add: two <script type="application/ld+json"> tags.

2. MCP server card at /.well-known/mcp.json (19/19 miss)

The Model Context Protocol is Anthropic's own standard. The discovery convention is a card at /.well-known/mcp.json describing your MCP server, transport, tools. Anthropic's own site does not publish one. Neither does OpenAI's, or Perplexity's. If your product exposes any programmatic surface (search, docs API, product catalog) this is a two-file win: card + server.

3. Tap-target size (17/17 checked, all fail)

Agents that render pages penalize dense clickable regions. Mobile-first CSS with 44x44px minimum touch targets is a decade-old accessibility rule that also serves modern agent renderers. Fix in your button component once.



Named findings (the fun part)


•
Cloudflare (B, 71%) was the only B on the list. They actually publish /llms-full.txt and declare bot rules. Given they sell an AI-agent-readiness product themselves, this is at least self-consistent.

•
Perplexity (F, 29%) was the worst. Not a single AI-policy declaration, no llms.txt, no schema markup on the homepage. The company literally sells "answers powered by web crawl" and their own site is opaque to crawlers.

•
Anthropic (D, 53%) publishes MCP as a standard but does not publish an MCP card of their own.

•
GitHub (D, 47%) has no llms.txt, no MCP card, no A2A card. GitHub's search index is what half of every LLM eats and they still ship 47%.

•
Stripe (C, 56%) and Vercel (C, 65%) were solid on schema but weak on agent protocols.

•
Supabase (C, 68%) was the second-best, mostly for good sitemap + strong schema coverage.

Full per-site grades and raw JSON: see the sibling deep-dive at agentfix.pro/blog/we-scanned-10-ai-forward-sites-in-2026.



Copy-paste minimum viable fix


If your site is D or below (statistically likely), start with these three files. Two hours of work moves most sites two grades.

/llms.txt (Answer.AI convention, becoming de-facto index for LLM crawlers):

# YourProduct

> One-line description an agent should quote.

## Docs
- [Getting started](https://your.site/docs/quickstart)
- [API reference](https://your.site/docs/api)

## Product
- [Pricing](https://your.site/pricing)
- [Changelog](https://your.site/changelog)

/.well-known/mcp.json (minimal, declares you have an MCP surface):

{
"schemaVersion": "2025-03-26",
"name": "Your Product",
"description": "What agents can do here.",
"server": {
"url": "https://your.site/api/mcp",
"transport": "streamable-http"
},
"auth": { "type": "none" },
"capabilities": { "tools": true }
}

Homepage FAQPage JSON-LD (biggest AI-overview lever per hour of effort):

<script type="application/ld+json">
{
"@context": "https://schema.org",
"@type": "FAQPage",
"mainEntity": [
{
"@type": "Question",
"name": "What does YourProduct do?",
"acceptedAnswer": {
"@type": "Answer",
"text": "One-sentence answer."
}
}
]
}
</script>

Then re-scan and watch the grade jump.



Why publish this


Two reasons. First, we run the scanner as a product and this is our transparency: you can hit agentfix.pro/scan right now and reproduce every number. Second, we plan to re-run the same 19 sites on 2026-10-05 and diff the results. If Cloudflare adds an MCP card, if OpenAI adds A2A, we will show that in an addendum. Snapshot-benchmarks are the closest thing agent-web has to a scoreboard.

If your site is on the list and something looks wrong, please DM: we ran a raw scan on production, no allow-list, and if a signal misfired we want to fix the scanner not defend the number.

If you want the raw JSON of all 19 scans, or the per-signal breakdown, the on-blog deep-dive has both.



What we track next


•
Agent-skills index (95% fail): a per-skill card that lets agents pick a capability without loading your whole docs site. Emerging spec, but the direction is clear.

•
Markdown content negotiation (84% fail): serving text/markdown on Accept: text/markdown requests. Cheap on a Next.js/Astro backend, invisible to human users.

•
Content-signal declarations (84% fail): the emerging "training-yes/training-no/quote-yes" trio that lets you say what you consent to.

We will benchmark all three in the October re-run.

This post was written by the team behind AgentFix, an AI-agent-readiness scanner and one-time fix packs for WordPress, Webflow, Shopify, Tilda. Free scan at agentfix.pro.