- AI search gets you into the answer. Agentic distribution is the extra work so an agent can call your product and finish a task.
- Content handles the story, the prompt set, and the copy inside skills. Engineering handles the interfaces, login, and whether a session holds up.
- Start by measuring the prompts buyers already use, then watch an agent try one complete task.
Most of the AI search work we do with clients stops at two questions: does the model mention you, and does it recommend you? We have written about mention scores vs. recommendation scores, what your library teaches the models, and how third-party pages talk about you when the story lives off your site.
What agentic distribution adds
Agentic distribution is the next layer. A buyer, or a coding agent sitting in their repo, is trying to use the product. We check that the way an agent would: HTTP GET of robots, sitemaps, llms.txt, HTML, OpenAPI, GitHub, MCP registries, and Wikipedia. No browser session. Some of the work is copy. Some of it is the product. Put each item on the right pile, even the ones you skip for now.
If people in Cursor are saying "add search to this app," measure those prompts. I would get those answers right before you build an MCP server or a new API.
Can an agent find you
Start with whether an agent can find you. If a search for your company misses, the model will pick whoever Wikipedia and Google already treat as the category. The homepage should say what the product does, in HTML a crawler can read, return 200, and still make sense with JavaScript off. robots.txt should let through the paths that matter, including user-triggered agents like ChatGPT-User. If your firewall treats every non-browser request as an attack, the agent never gets a page back.
- Show up for your company in Search Console. If you are missing there, the rest of the list is working around a hole.
- A Wikipedia / Wikidata page, if the category article exists and you are absent from it. Models still pull from that page.
- One docs host. Two live hosts both returning 200 means agents fetch both and mix the copies.
- Unknown URLs should 404. A junk path that 200s to Get started will get cited as if it were real.
- Sitemaps for marketing, docs, and community, submitted in Search Console. Fix or drop the ones that error.
- CLI and SDK on the public package index, MCP in the registries agents search, and a user skill on the indexes Cursor already looks at. Contributor AGENTS.md is for people hacking on your repo. Cursor is looking for "scaffold, deploy, debug."
What the public pages say
The public pages should describe the product as it exists today. Older vs-pages, an llms.txt that lists brand assets instead of how to install, and a docs host on the old brand will outnumber the live story. Link the CLI, MCP, and OpenAPI from the site you already rank for. Agents are bad at guessing a separate API hostname.
- llms.txt as a short index of jobs: when to use you, how to install, and where CLI, MCP, OpenAPI, and examples live. Stay under ~30k characters. Put the long dump in llms-full.txt and in per-page .md files.
- Markdown copies of the important URLs, and a 404 body that points back at the index.
- JSON-LD, canonicals that match the host you want fetched, plus about, contact, and privacy.
- An OpenAPI spec that fetches, linked from marketing llms.txt or well-known. A spec that only lives on the API hostname might as well be private.
Getting in without a browser
An agent has to get in without anyone sitting in Chrome. If the only way in is a person clicking through a consent screen, the session stops there. When a call fails, the body should say what was missing. A bare unauthorized status is hard for a script to recover from.
- CLI for setup and writes. MCP for docs search (anonymous) and product state (auth that matches the job). Streamable HTTP for remote clients.
- Call each tool the way you would title a page, specific enough that the next call is obvious.
- Rate-limit headers that show remaining quota.
- /.well-known/mcp.json on the marketing host, not only on docs.
- Leave ARD catalogs, RFC 9727 catalogs, and OAuth metadata until 404s, OpenAPI, and a usable llms.txt are in place. If Cloud stays API-key-only, say that in llms.txt instead of staffing RFC 8414.
Setup and a complete task
The gap I see most often is setup: API keys, webhooks, environments. Listing an interface is a start. Have an agent try to open a project, run the core action, and read status. pip install plus a local run is a different test from opening a Cloud workspace. Some steps should require a person. Write those down. Otherwise the team, and the agent, will treat every path as unattended.
- One install or local run before anyone creates an account.
- Admin and lifecycle endpoints, on top of the feature endpoint.
- Logs that can tell Cursor, Claude, MCP, and the CLI apart from a person in the UI. Rate limits sized for a retry loop. The next sprint comes from what broke last week.
- Machine-payable APIs, WebMCP, A2A cards, MCP Apps, NLWeb, and ChatGPT directory listings can wait. Recommendation, identity, 404s, and a working install come first.
Who does the work
Content handles the prompt set, third-party pages including Wikipedia, the story on the URLs agents already land on, examples, skill copy, and llms.txt. Engineering handles CLI, MCP, SDK, OpenAPI, login without a browser, whether a session survives a real task, lifecycle APIs, rate limits, logging by client, real HTTP 404s, and one docs host. The person watching ChatGPT mentions is the wrong person to debug why an MCP connection dropped. Split the work.
Where to start
You can skip items. Showing up in the answer and finishing a job are separate pieces of work.
- Measure the prompts your buyers already use, including coding agents if that is how they try you.
- Fix the public pages, and the third-party pages you can influence, that currently misstate the product.
- Publish a machine-readable contract and, if you can, one call before anyone signs up.
- Add the interfaces those agents already use, and test them the way an agent would.
- Confirm one end-to-end task completes, including setup.
- Turn whatever broke into the next docs and interface work.
If you want a smaller first project, I wrote about how to get started with agentic marketing: add agents to the system you already have.
