Two Hacker News projects just told you your website has a second audience
As of August 2026, two projects hit the Hacker News front page inside the same 48 hours, solving the same problem from opposite ends. acceptmarkdown.com earned 164 points and 95 comments for a plain idea: when a request arrives with an Accept header saying the client wants Markdown, serve a clean Markdown version of the page instead of the full browser payload. WebMCP, written up on August 4, 2026 with 55 points and 55 comments, goes further — it lets your website publish named actions an AI agent can call directly, like checking availability or starting a booking, instead of scraping your forms and guessing.
For an operator, the takeaway is simple: the software reading your website is now a real audience, it does not see what your customers see, and developers are building the plumbing to serve it deliberately.
What an agent actually sees when it hits your site
When a customer asks ChatGPT, Claude, or Perplexity for the "best roofing company near me," the model does not open your homepage in a browser. It works from whatever its crawler pulled out of your HTML — often buried under navigation, cookie banners, chat widgets, and JavaScript that never ran. If your service list is an image, it is invisible. If your prices sit in a PDF, they may as well not exist. The Markdown-feed idea drew 95 comments because it removes all of that: heading, text, list, done.
This is the same reason a business gets left out when a customer asks an AI assistant for a recommendation. The model cannot name what it could not cleanly read.
The trend is not slowing. In the same week, Perplexity published its "Portable Computer" push for local-first AI, and a developer put a 35-billion-parameter Qwen model on a Raspberry Pi to run a car assistant (144 points). More agents, running in more places, all needing structured input. Sites that are easy to read will get named; sites that are not will get skipped.
Do these three things this week
1. See what the machine sees
Paste your homepage URL into ChatGPT or Perplexity and ask it to list your services, service area, and pricing. Then do the same for two competitors. If the model gets your business wrong or comes back thin, that is the version of you it hands customers. This costs ten minutes.
2. Make the page legible
Put your services, cities served, and prices in real text with real headings — not baked into a hero graphic. Add a plain-text or Markdown version of your key pages, and an llms.txt file that points to them. This is the same move acceptmarkdown.com automates, and for most sites it is an afternoon of work.
3. Decide about actions
WebMCP is the early edge of agents not just reading your site but doing things on it — pulling a quote, holding a slot, submitting a job request. Before turning that on, be clear about which actions an agent may complete with no human looking. That is the judgment call we walked through in our post on approval gates for employee AI agents: read-only is safe to automate today; anything that books revenue or makes a promise needs a gate.
The build-vs-hire read
Adding a Markdown route and an llms.txt file is genuinely small — a competent developer does it in a day. The ongoing part is what breaks: keeping that feed correct every time you change a price or drop a service area, watching whether ChatGPT and Perplexity actually start citing you, and wiring a booking agent that knows your real calendar and your real rules. One-time setup is cheap. The maintenance and the monitoring are the job, and that is the part owners underestimate.
There is also a cost to getting it wrong. A Markdown feed still showing last year's pricing, or an agent that books jobs outside your service area, does damage faster than no feed at all — because the machine presents it with confidence and the customer believes it. That is the case for treating this as a maintained system, not a one-time checkbox.
That is why our agent stack is split into modules — visibility, content, booking — instead of one big build. You can stand up the reading layer yourself and add the pieces that need upkeep when the manual version starts costing more time than it saves.
What this means if you're weighing AI marketing or an agent build
The cheap, urgent move is making your site cleanly readable by machines this week, because that is what decides whether an AI assistant can recommend you at all. The build-vs-hire line falls at maintenance: do the one-time setup in-house, and bring in help for the monitoring and the agent actions that have to stay correct over time.
Want to know what ChatGPT, Claude, and Perplexity say about your business right now? Get your free AI Visibility Report — it takes 24 hours and shows exactly where you stand.