Give agents pages and tools they can use.
Agent-ready content like markdown, the right MCP tools and APIs, and each agent sent to the path where it finishes the job. You see which agents arrive and set what they may do.
Agents get the pages built for people.
Personal agents reach you through your website, your MCP server and APIs, and your customer-facing agent. Almost all of them get the experience built for people. A fee sits behind a pop-up, a button has no label, or a tool asks for a code the agent can't find. The task fails, and no error tells you.
Few businesses offer agents anything better. In Cloudflare's scan of the top 200,000 sites, only 3.9% answered a request for markdown. Fewer than 15 published an MCP server card or API catalog.
An agent that can't finish is a customer you lose.
What personalization means here.
The format and the path an agent gets. An agent can get markdown where a person gets a page, or a tool where a person gets a form. Prices, fees, stock and policies stay the same for everyone.
You also see which agents arrive, how sure we are about each one, and what each one may do.
How it works.
Setup goes from lightest to heaviest. Nothing sits in front of your site to start.
- 1
Start with your logs
Run our script that removes personal data, then send seven days of website and MCP logs.
Nothing to install.
- 2
Ship files in your own repo
Markdown, structured data, tools and the files agents look for.
They live in your own code and accounts.
- 3
Add a watch-only layer
A CDN worker or Next.js middleware next to your bot tools, and a small package in your MCP server.
It watches and changes nothing.
- 4
Go live, one path at a time
Each change is tested in runs before it ships.
If the website layer stops working, traffic goes straight to your site.
What's inside.
01 · Agent-ready content
Content agents can use, with the same facts people see.
Agents that ask for markdown get it at the same address, with a link back to the page. Structured data goes on the page people see. A lighter page goes to verified agents, and only where runs show a gain.
- Markdown for the agents that ask for it
- Structured data for prices, stock, fees and policies
- Works through Cloudflare, Adobe or Akamai, or our middleware
- Search crawlers always get the page people see
- 34% → 68%
/fare-rules
Markdown for the 3 agents seen asking
- 61% → 74%
/flights
Structured data for fares and fees
- Pulled
/baggage
Markdown pulled: the fee didn’t match
- Left off
/seats
Lighter page tried, no real gain
Finish rate in runs, before and after. Search crawlers always get the page people see.
02 · Tools and APIs
The right tools and APIs, where agents can reach them.
We map each key job against the ways an agent can do it. Where a tool is missing, we design it for agents, generate it from your API and set up sign-in. Then we get it listed where agents look for tools.
- A few tools per job, named and described for agents
- Sign-in set up and tested with each agent
- Listings for ChatGPT, Claude, Muse and the MCP Registry
- Every file agents look for, written from one source
| Job | MCP tool | Who can reach it |
|---|---|---|
| Change a flight | Built · 9 of 10 pass | ChatGPT, Claude · Muse in review |
| Add a bag | Needs a fix | ChatGPT, Claude |
| Cancel and refund | Missing | – |
Files agents look for
MCP server card, API catalog and A2A agent card, published from one source.
03 · Routing
Each agent sent to the path where it finishes.
For each agent we can identify and each job, we find the path that finishes most often and the way to point the agent there. Both are tested before anything ships. Then live traffic is split until one path wins.
- A page flow, an MCP tool or a shorter flow
- Path tests and pointer tests before launch
- A winner only after 200 sessions per path
- People and search crawlers see no change
Verified and declared agents
LiveFare search → /fares/compact
84% vs 58%, new path wins
Claude, connector installed
LiveChange a flight → MCP tool
86% finished, live
ChatGPT agent
Not routedAdd a bag → app pointer
Pointer followed 1 of 5
People and search crawlers get the site as it is.
04 · Same-facts check
Every path gives agents the same facts.
We compare prices, fees, stock and policies on every path: the page, the markdown, structured data, tool replies and what your customer-facing agent said. A version that doesn't match is pulled, and the page people see is served.
- Checked on every publish and on a schedule
- Reads what each agent was actually told
- A log of every agent version served
- Keeps agent versions within search engine rules
Add a checked bag
- Page people see$40 ✓
- Markdown$35 ✗
- Structured data$40 ✓
- add_bag tool reply$40 ✓
- Customer-facing agent$40 ✓
Markdown pulled. The page people see was served instead.
05 · Identification
See which agents arrive, and how sure we are.
Each agent session gets one of four levels: verified, declared, flagged by your bot tool, or not seen. Spoofed names are flagged. Who an agent acts for shows only when there is a real link, like a signed-in session.
- Reads the bot tools you already run
- Adds your MCP and API traffic
- Signed-in and returning agents
- Agents that look like people are never counted
Verified 1,120 · Declared 410 · Flagged 1,684 · Spoofed 96
- Flagged by your bot tool
Muse
Guest · Website
- Verified
ChatGPT agent
Signed in · Returning
- Declared
Claude
Signed in · MCP server
- Spoofed
Says “ChatGPT-User”
No signature · wrong network
Agents in a person’s own browser look like people. They’re not counted.
06 · Access rules
Set what each agent may do. Test it first.
One rule set across your website, MCP server and APIs. Each rule allows, limits, asks the person first, blocks, or sends a stop message that tells the agent where to go instead.
- Replayed on your recent traffic before it goes live
- Safe defaults for agents acting for signed-in customers
- Written into the bot and fraud tools you already run
- Any block that hits customers needs your sign-off
- Verified agent, signed-in member · change bookingAllow
- Any agent · refund over $500Ask first
- Fare search over 200 pages an hourLimit
- Spoofed name · anythingBlock
Test before it goes live
“Block declared agents at checkout”
Would have stopped 291 checkouts, with $61,300 in linked orders.
Suggestion: limit instead.
What it covers.
Every agent session answers the same questions.
Who it is
- Verified
- Declared
- Flagged by your bot tool
- Not seen, and never counted
- Spoofed names, flagged
Who it acts for
- A signed-in customer
- A guest
- A returning agent
What it gets
- Markdown
- Structured data
- A lighter page
- An MCP tool or API
- A shorter flow
Facts kept the same
- Prices
- Fees
- Stock
- Options
- Policy text
What it may do
- Allow
- Limit
- Ask the person first
- Block
- Stop message
Start with one question.
Tell us what you want to know about the agents that come to you. We start from your logs.
For example
- “Which agents visit us, and what do they try to do?”
- “Which of them are who they say they are?”
- “What should an agent get when it opens our pages?”
- “Do agents see the same prices and policies as people?”
- “Which agents should we let in, and to do what?”
- “Which jobs need a tool built for agents?”
From seven days of your logs, with personal data removed first: which agents came and what they tried to do.
Works with what you already run.
It works next to your CDN and your bot tools.
- Websites
- MCP servers
- APIs
- Cloudflare, Fastly, Akamai and Vercel
- Fingerprint, HUMAN and DataDome
- Signed agent requests (Web Bot Auth)
- ChatGPT apps, Claude connectors and Muse connectors
- Next.js
Questions.
Is this cloaking?
No. Agent versions carry the same facts as your page, and they are checked every time you publish. Search crawlers always get your normal page.
What does personalization mean here?
The format and the path an agent gets. The facts stay the same for everyone.
Can you see every agent?
No. Some agents run in the person's own browser and look like people. We show how sure we are about each session and leave the rest alone.
Do you sit in front of my site?
Not to start. We begin with your logs. Later, a small layer can watch next to your bot tools. If it stops working, traffic goes straight to your site.
Will agents use the tools you build?
Only the agents that can reach them. We build the tool, test it, and get it listed where agents look for tools.
How do you test changes with agents like Muse?
We run the real agents by hand. For hundreds of runs we use our model of each agent, built on the AI it runs on where we can reach it, and checked against the real agent every week. Every result says which.
What happens to our data?
You remove personal data with our script before you send logs. We delete raw logs after 30 days, and a one-page data agreement is ready.
Do I need your other products?
No. It works on its own.