If you thought chip companies were ahead of the curve on everything tech, here's a reality check that's gonna sting. A developer ran an agent-readiness scanner across 106 semiconductor, analog, power, RF, sensor, MCU, memory, passive component, connector, EDA, distributor, and test-and-measurement websites—and the industry collectively failed. The average score came in at 42 out of 100. Zero sites earned an A. Only one scraped by with a B.

The Audit

The scanner evaluated each site across 20 different signals: robots.txt compliance, llms.txt presence (a dedicated index for AI agents), MCP server endpoints, structured data markup, and content negotiation headers that would let agents request markdown instead of wall-of-HTML. Think of it like checking whether a website speaks the language your AI assistant is actually trying to use when it's shopping for parts on your behalf.

Winners and Losers

Quectel took top honors with a 71—barely cracking a B by academic standards but miles ahead of everyone else. Right behind: onsemi, Nuvoton, and Silergy at 69; MediaTek at 63; AMD and Infineon tied at 60. NVIDIA managed a respectable 55, while Texas Instruments (39), Analog Devices (34), and Qualcomm (39) all flunked. On the bottom end: Cadence, Sensata, Goodix, and Anritsu each landed at 16—absolute fail territory.

The Three Fatal Flaws

Almost every site shared the same three blind spots. First: zero markdown content negotiation. Send Accept: text/markdown in your request headers and you still get slapped with a wall of rendered HTML. Second: no llms.txt file—a simple index that would tell an agent which pages actually matter instead of forcing it to crawl everything blindly. Third, and worst: not a single site exposed an MCP server endpoint for calls like search_parts or get_datasheet, leaving agents stuck scraping raw DOM like it's 2015.

The Token Tax

Here's what makes this painful in practice. A typical chip company homepage costs between 40,000 and 90,000 tokens to process—because of all the markup bloat an agent has to chew through before extracting actual content. Strip that same page down to clean markdown and you're looking at 1,000 to 2,500 tokens. That's roughly 95% waste. Every buyer query through Perplexity or Claude is burning compute on garbage because these companies never optimized for how their content gets consumed.

Key Takeaways

  • The entire semiconductor supply chain is essentially invisible to AI-powered purchasing agents
  • Major players like TI, ADI, and Qualcomm are scoring in the 30s—barely functional for agentic workflows
  • llms.txt and MCP support remain non-existent across all 106 sites tested
  • Token inefficiency means buyers using AI tools get slower responses and higher costs when researching these products

The Bottom Line

This is a massive self-inflicted wound. Buyers are moving to AI assistants for research—semiconductor companies that don't adapt will simply stop getting recommended. Fixing this isn't even hard: add an llms.txt, support markdown content negotiation, and expose one damn MCP endpoint. The infrastructure exists. Nobody's using it.