The Agent Web Selling to Machines That Buy

Classic search matched strings. A page containing your words could rank for your words. Answer engines work on entities instead, which means the question is not whether your page mentions a phrase but whether the machine is confident it knows what thing you are.

Confidence is assembled rather than declared. Your title area says something. Your schema states something. Your llms.txt claims something. Your profiles and the pages other people write add their own versions. The model reads all of it and looks for agreement before it repeats any of it.

That assembly is why contradictions are so expensive. Publishing two versions of your name, your description or your numbers does not give a model a choice to make. It gives a model a reason to skip you, because quoting a source it cannot reconcile is the one outcome an answer engine is built to avoid.

The fix is filing work rather than talent. One name spelled the same everywhere, one sentence of description reused word for word, current numbers in every place they appear, and a permanent identity in schema that every page you publish points back to. If two names for you are genuinely true, schema has a field for that, and using it beats picking one and hoping.

There is a useful test while you work. Open a fresh chat and ask an engine who you are, then compare its answer line by line against your own pages. Anything it gets wrong is not a mystery about the model, it is a contradiction you published and forgot.

The entries below cover the shift end to end: the three protocols and what each one expects, the feed that quietly replaces your landing page, what happens when being read starts costing money, and what an ad slot inside an answer does to a citation you earned for nothing.

01The working notes