Agentic Optimization: Being Found by AI Agents, Not Just AI Answers

Three layers, and most brands optimise one

There are now three distinct ways a machine can put your product in front of demand, and they use completely different inputs.

Layer	Who is reading	What you optimise	What winning looks like
SEO	A person scanning results	Pages, links, rankings	A click
GEO and AEO	A person reading a generated answer	Answer-shaped passages, schema, entity signals	A citation
Agentic Optimization	An agent doing a job, with no person watching	A capability manifest and its descriptions	Being called

The first two are about being read. The third is about being used. That is a bigger difference than it sounds, because an agent that has decided to book a flight does not want an article about booking flights. It wants something it can invoke.

Why your website is invisible at the third layer

When an agent needs a capability it does not open a search engine. It asks a discovery service a plain-language question, gets back a list of machine-readable resources, picks one, and connects to it directly over that resource's own protocol. Your marketing site is not in that loop at any point.

This is the part that surprises people who have done years of good SEO. Ranking is not a transferable asset here. The discovery layer has its own index, built from manifests published on domains, and a site with a million backlinks and no manifest is simply absent from it.

The standard, in one paragraph

Agentic Resource Discovery (ARD) is the specification for this layer, developed in the open with participants including Google, Microsoft, Hugging Face and GoDaddy, published under Apache 2.0. You serve a JSON document at /.well-known/ard.json on your own domain. It lists your resources: MCP servers, agent cards, OpenAPI services, skills. Registries crawl it, agents query the registries, and your capability becomes selectable. It deliberately stops there: ARD handles discovery only, and once an agent has chosen your resource it connects over that resource's native protocol.

It is worth being precise about what this is not. It does not replace MCP, which is how an agent calls a tool. ARD is how it finds one worth calling. You need both, and they never do the same job.

The opportunity, with the actual numbers

This is the rare case where the competitive picture can be stated rather than guessed at. A published crawl of 375,998 domains in September 2026 found 207 serving an ARD manifest at a well-known path. Of those 207, 176, or 85 percent, were still on the path the specification superseded on 26 August 2026, which a fully conformant consumer is not obliged to check at all.

Compare that with the situation at the layer above. Getting cited by an AI answer engine means competing with every well-resourced content team in your category. Being discoverable by an agent means competing with a few dozen domains, most of which have not noticed the specification moved. The counts above are served as JSON at ardregistry.net/api/stats.json if you want to check them rather than take our word for it.

What to do, in order

1. Publish a manifest at the current path

Serve JSON at /.well-known/ard.json on the domain you actually want credited. The old ai-catalog.json path was superseded in August 2026. If you already publish one, moving it is the highest-value hour available to you, because 85 percent of your competition has not.

2. Write the queries, not the category

Each resource carries a list of natural-language requests it answers well. This is the field that decides whether a valid manifest is ever returned, and it is the one most publishers fill in with a marketing noun phrase. "Document processing" matches every document tool ever written. "Extract line items from a scanned invoice" describes a job somebody actually has.

The rule that catches most mistakes: every query needs a verb, and you should be able to say it out loud to a colleague without them asking what you mean. Three to eight per resource. Write what you are genuinely good at, because claiming a query you cannot answer spends the one impression you were given proving you were the wrong choice.

3. Describe your tools for a stranger, not for your own docs

If you run an MCP server, the tool names and descriptions are read by a model that has never seen your product. A tool called send described as "sends via the configured channel" is unusable. The same tool called send_sms, with parameters named to_phone_number and message_body, is self-explanatory. Parameter names carry real weight because they are frequently the clearest statement of what a tool actually does.

4. Get into the registries

Publishing the manifest makes you discoverable in principle. Being in an index makes you discoverable in practice, and there are several public ARD registries, each with its own submission door. Submitting to several costs nothing. The part nobody can take away from you is the manifest on your own domain, which is what reaches all of them over time.

5. Then measure it

Publishing without checking is how the 85 percent happened. Ask a registry your own questions and see whether you come back, and who comes back above you. If you are not returned for a query you wrote about yourself, the sentence is not describing you as clearly as you assumed.

How this interacts with the GEO work you have already done

It compounds, in one specific direction. Off-site brand mentions correlate roughly three times more strongly with being recommended by an AI system than backlinks do, and AI answers overwhelmingly prefer third-party sources over a brand's own pages. That is about being talked about. Agentic Optimization is about being reachable. A brand that is well known and unreachable loses the task to a brand that is reachable, because the agent can only call what it can find.

One practical note that catches people: most AI crawlers do not execute JavaScript, so anything injected client-side is invisible to them. A manifest is a static JSON file, which sidesteps the problem entirely, and it is worth checking that the rest of your evidence is in the HTML your server sends rather than assembled in the browser.

The honest limits

Two things are worth saying plainly, because the space is young enough that confident claims should be treated with suspicion.

Nobody can tell you agent search volume. Discovery services do not publish query logs, and any vendor quoting you a monthly figure for how often agents asked for your capability is showing you their own test traffic or a model output. What is measurable is position: whether you are returned, and who is above you.

And the standard is young. The manifest path changed once already, in August 2026. Anything you publish now should be treated as something you will revisit, which is an argument for generating it from a source you control rather than hand-editing JSON you will forget about.