What agentic readiness actually mean
Agentic readiness is task-completion capability. Indexing and citation work get an agent to your page. Agentic readiness decides what happens in the sixty seconds after it arrives with a job to do.
Your GEO program may be doing exactly what you hired it to do. It just stops at the door.
The distinction matters because different teams fix the two problems. Getting retrieved and quoted in AI answers is content and entity work, which is what AI visibility optimization services cover. Getting a task finished is engineering work: rendering, forms, markup and access rules. A site can be excellent at the first and useless at the second.
An agent that stalls halfway has no way to tell you it stalled. Your analytics won’t record the attempt either. That’s why the only reliable signal available today is watching one try.
Run this test yourself: a 15-minute agent walkthrough
Open a browsing agent, give it one real task on your own site, and watch. That’s the whole method. No scanner, no signup.
Write the protocol around the capability, not a product name. ChatGPT Atlas carried agent mode until OpenAI shut it down on 9 August 2026 and folded browser automation into its desktop app, as TechRepublic reported that month. A checklist naming a specific agent product goes stale in weeks.
One tool that does the job today is Claude in Chrome, Anthropic’s browser extension, available on all paid plans. Anthropic’s own description is the useful part: it “reads the page, clicks, types, navigates, and fills forms, while you watch and decide what happens next.” Any agent with those four verbs will run this test.
Pick one task that actually matters
Choose the task a real buyer would delegate. Not a reading task.
Good candidates: find the price for a named plan, submit the contact form with a specific message, book a demo slot for next Tuesday, download the security whitepaper.
Pick one. Running four tasks badly tells you less than running one properly.
Give the agent the task, then keep your hands off
Start the agent on a search engine or a blank tab. Don’t start it on your homepage: discovery is part of the task, and handing over your URL hides the first failure.
Use plain buyer language, naming your company the way a buyer would. “Find out what the mid-tier plan costs at Acme and book a demo” is a real instruction. An instruction like “test my site” gives the agent nothing to do.
Then stop helping. The moment you take the mouse back, the test is over and you’ve learned nothing.
Log the failure, not the feeling
When it stalls, record four fields: the URL, the step it was attempting, what was on screen, and the agent’s own words about what it couldn’t do. That last field is the most useful line in the log.
“It seemed confused” isn’t a finding. “It couldn’t select a value from the custom dropdown on /pricing and stopped” is a ticket. File that one.
The five things that block agents mid-task
These five failure modes each have a fix a front-end team can ship.
| Blocker |
What the agent hits |
The fix |
| Client-side rendering |
Content injected after page load is missing from the accessibility tree the agent reads |
Server-render or pre-render the pages that carry pricing and forms |
| Forms and CAPTCHA |
Unlabeled fields, custom dropdowns and bot challenges stop the run at the final step |
Label every field, use native form controls, keep a challenge-free path on low-risk forms |
| robots.txt written for search crawlers |
Directives that predate agents, blocking or ignoring them by default |
Declare AI-agent preferences explicitly instead of inheriting a 2019 file |
| Missing structured data |
Price, availability and contact details have to be inferred from prose |
Mark up identity, offers and contact points so they can be read rather than guessed |
| No MCP or API surface |
Nothing machine-readable describes what the site can do |
Publish an API catalog or an MCP endpoint when the task is transactional |
Three of those terms get used loosely. Structured data is schema.org markup that states facts about a page in a fixed format, so a machine reads the price instead of parsing it out of a sentence. robots.txt AI-agent directives are lines in your existing robots.txt declaring which AI systems can access the site, and for what purpose. MCP or API access is a machine-readable interface that lets an agent call your service directly, rather than clicking through an interface built for humans.
Adoption is thin. Cloudflare’s April 2026 scan found that 78% of sites have a robots.txt file while only 4% declare any AI usage preference in it, and markdown content negotiation passes on 3.9%. The fewer-than-15 figure for MCP and API surfaces is the far end of the same scan. These are the 200,000 most visited domains: the well-resourced end of the web.
Which leaves llms.txt, the file those pages all recommend. Publish it if you like. It’s cheap. Just don’t expect it to work yet.
Ahrefs analyzed 137,210 domains in May 2026, found 28% publishing an llms.txt file, and found 97% of those files got no requests at all that month. That sample skews technical, so treat 28% as an upper bound. Google’s John Mueller called the convention “purely speculative for now” in a Reddit discussion rather than in Google documentation, noting the file has existed for years without AI systems using it (Search Engine Journal, June 2026).
The accessibility tree explains most of that table. An agent doesn’t see your design. It reads the same structured representation a screen reader uses.
The accessibility work most B2B teams have deferred is also the agent work. That overlap is the cheapest first move for the B2B SaaS teams Rampiq works with.
The agentic readiness checklist
Every item maps to a step in the walkthrough. Run the test first and let it tell you which ones you need.
- Discovery. The agent found you from a cold start, without being handed your URL.
- Rendering. Your pricing, product and contact pages return their content in the initial HTML response.
- The critical form. The agent completed one real form without a human touching the keyboard.
- Labels and controls. Every field on that form has a programmatic label, and no step depends on a custom control the agent can’t operate.
- Access. Your robots.txt states an explicit position on AI agents, whatever that position is.
- Markup. Organization, offer and contact details sit in structured data, not only in prose.
- Machine interface. If the task is transactional, an agent can reach it through an API or MCP endpoint.
- Measurement. You review server logs rather than the analytics dashboard, because a server-side agent fetch never fires the JavaScript tag. Tools that track AI visibility over time cover the citation half; the completion half still needs the manual test.
Order matters more than count. Items 1 to 4 decide whether a task finishes. Items 5 to 8 decide whether you can see it happening.
One note on how we run this at Rampiq: our technical fixes ship alongside GEO content briefs and referral dashboards, on a monthly cadence with the client’s in-house team rather than as a one-off audit. We work it that way because blockers come back as sites get rebuilt, and a fix list with no re-test date expires quietly.
If you’d rather not run the walkthrough by hand across a large site, you can run a full AI search & GEO audit with us instead. It covers the same ground at scale, against your priority pages.
