A share of the people evaluating your organisation now arrive as software. Not browsing — working: comparing suppliers, checking a requirement, assembling a shortlist, answering a question someone asked in a chat window.
The useful framing is not that you have a new audience. It is that a filter has been inserted between you and a person, and the person will never see what the filter rejected. Being agent-ready means surviving that filter and being represented accurately on the other side of it.
The good news is that almost everything that helps is unglamorous, and almost everything sold as agent-optimisation is not where the value is.
Four levels, in the order that matters
| Level | The agent can… | Effort | Who needs it |
|---|---|---|---|
| 1. Discoverable | find you at all | Low | Everyone |
| 2. Understandable | tell what you offer and who you are | Low–medium | Everyone |
| 3. Navigable | reach specifics without executing scripts or guessing URLs | Medium | Most |
| 4. Transactable | do something on a person's behalf | High | Few, for now |
Vendors sell level 4. The returns are at levels 1 to 3, and they are cheap. Sequence accordingly.
Level 1: discoverable
Nothing exotic. Stable URLs that do not change when marketing redesigns something. A sitemap that reflects what actually exists. A robots.txt that permits the content you want cited — and a deliberate decision about AI crawlers rather than an accident.
Get robots.txt right in particular, because its failure mode is invisible. A crawler obeys exactly one group, its most specific match, and does not merge that with the wildcard group. A well-intentioned per-crawler Allow: / block therefore exempts that crawler from every exclusion below it. We made that exact mistake here; the website guide shows the before and after.
Level 2: understandable
This is where most organisations have the largest and cheapest gap.
Every page needs to say what it is — its own title, its own description. Obvious, and the single biggest defect found on this site: 33 pages covering eleven professional programmes all rendered the bare site name as their title and shared one site-wide description. The per-page copy existed in the database and was never rendered. In a browser the pages looked fine.
Your organisation needs one identity, stated once and referenced everywhere. Emit an Organization node site-wide with a stable absolute @id, and have every other structured-data node reference it rather than restating a slightly different version. That is what lets an engine resolve your name to one entity instead of several near-duplicates. It is the highest-value single item on this page.
Say what you are not. Boundaries — not legal advice, not a substitute for professional judgement — are worth stating explicitly, because an agent summarising you will otherwise infer scope from context.
Level 3: navigable
An agent that does not execute JavaScript sees your server-rendered HTML and nothing else. If your page ships an empty container filled after hydration, that page is blank to a meaningful share of the software reading it. Fetch your own page with a plain HTTP client and look at what comes back inside <main>; the check takes a minute and settles the question.
Beyond that: predictable URL patterns, real hyperlinks rather than script-driven navigation, and nothing important behind a hover, a modal or a CAPTCHA that exists only to gate reading.
Two optional additions at this level:
- llms.txt — a curated markdown index of your important resources. Cheap and unproven, and worth doing only if you can generate it, because a stale one is worse than none. See llms.txt: what it is and whether you need one.
- A read-only API, if you hold structured data an agent would genuinely want — a directory, a catalogue, a dataset. It turns scraping into reading, and gives you rate limiting, a stable contract and a place to state licensing terms.
Level 4: transactable, and why to wait
Letting an agent act — book, buy, submit, enrol — is a different category of problem, and the reason is not technical.
Once software can act on a person's behalf, you inherit questions with no settled answers: how you know the agent genuinely represents the person, what it may do without a fresh confirmation, what happens when it acts on a misunderstanding, and who is liable. The security position also changes sharply, because an agent that acts can be induced to act wrongly — see prompt injection, which is an architectural property of language models rather than a bug awaiting a patch.
For most organisations the right position today is: be excellent at levels 1 to 3, and let agents hand off to a person for anything consequential. That is not timidity. It is the same judgement you would apply to any new channel that can commit you to something.
What not to bother with
- Custom
ai:meta tags. No engine documents support for them. - Marking up everything you can. Schema types you cannot support with real page content are a liability, not a signal.
- Writing separate content for machines. Duplicate surfaces drift apart, and the drifted one is the one that gets cited.
- Agent-optimisation as a subscription. The work is on your side of the wire and is mostly one-off.
The test worth remembering
Every technique in this cluster also makes the page better for a person: clearer structure, stated conclusions, visible sources, honest titles. When a technique only helps the machine, that is usually a sign you have drifted from making yourself understood into trying to game something.
The first has been rewarded for thirty years. The second has never lasted.
Where to go next
- Doing the work. How to make a website readable by AI agents — the implementation guide, written from doing it here and reporting what broke.
- The index file. llms.txt: what it is, and whether you need one.
- Before you let agents act. Prompt injection: why it happens and what actually stops it.
- Governing it. AI governance, and human oversight that is more than a rubber stamp — because approval fatigue is the failure mode of every agent confirmation step.