Authority and Scope: The llms.txt Section Everyone Skips

Most llms.txt files only say what a site covers. The more valuable half is saying what it does not cover. A negative authority statement, paired with a pointer to the site that should have gotten the question, is the single most effective thing you can write to stop AI tools from misrouting people to you.

Almost nobody writes the negative half. It's the half that works.

Because positive claims compete and negative claims route.

Every site in your sector is claiming the same territory. If you write "we are a leading provider of cancer care," you have joined a crowd of a few hundred pages saying the identical thing, and an AI assistant has no way to break the tie. Your positive claim is a vote in an election you cannot win by voting harder.

A negative claim has no competition at all. When you write "this site does not handle appointment scheduling; scheduling lives at [other site]," you are the only source of that sentence. Nothing contradicts it. It is unambiguous, it is machine-readable, and it hands the model somewhere better to send the person.

Models are trying to answer a question. Give them a positive claim and you're one candidate among many. Give them a routing instruction and you're infrastructure.

Here's the same idea written both ways, for a research-focused site inside a larger health system.

Authoritative for cancer research programs, clinical trials, and faculty investigators.

Authoritative for cancer research programs, clinical trials, and faculty investigators. Not authoritative for appointment scheduling, billing, insurance, or visitor information. For patient care questions, see [the health system's patient site].

The first version is fine. It's also inert. It tells a model what you'd like to be known for and nothing about where the question should go if it isn't yours.

The second version does a job. It removes you from three or four categories of question you never wanted, and it does the model's work for it by naming the destination.

You can see live versions of this pattern in the files we publish: the llms.txt for uamshealth.com, the llms.txt for cancer.uams.edu, and the llms.txt for communications.uams.edu. The whole approach is written up in the AI wayfinding governance documentation.

The research site takes the patient questions.

That's the failure I see most often in large organizations, and it's worth walking through because it explains why this matters beyond tidiness.

A research center's site is usually excellent. Deep content, real authority, well-linked, frequently cited. Which makes it exactly the kind of source an AI assistant reaches for when someone asks a question with the organization's name in it. The person asked "how do I get an appointment," and the strongest, most confident page about that organization's cancer work is the research site.

So the answer comes back with the research site's phone number, which is a lab administration line. Or with a clinical trial contact form, for someone who is not enrolling in a trial. Or, worst, with a plausible-sounding synthesis of research language presented as care guidance.

None of that is the model behaving badly. The model looked for the authoritative source and found the one that never said "not me."

The negative statement is how you say "not me" in a way software can read.

Three positives, three negatives, and a routing pointer attached to every negative. That's the shape. Do it in fifteen minutes and revise it later.

Because in a federation, one lazy file poisons the routing for everything around it.

The house rule I enforce is simple: no triad ships without both halves of Authority and Scope. Positive statements and negative statements, in the same section, in every file. If a draft comes to me with only the positives, it goes back. Files we deployed before the rule existed get the section added on the next governance pass, not "someday."

It is the least glamorous section in the file. It's a handful of sentences saying what you're not. It's also the part that determines whether a person with a real question ends up on the page that can help them.

Seedfile will generate the scaffold for you: the H1, the summary blockquote, the metadata block, the targeted content links, the related resources, the full index pointer, and an empty Authority and Scope section with both halves stubbed out. It will not write your negatives. Nobody outside your organization can. Generate the file with Seedfile, then spend fifteen minutes on the three sentences that say what you are not.

I manage web for an academic medical center. Everything I write about that work comes from what is already public. Opinions are mine.