The platform
Make AI agents succeed on your site
Ghost Agent Lab shows you every AI agent that visits, proves agents can finish your key journeys, tells you exactly what to fix, and puts a dollar figure on the result. One platform, four jobs: observe, test, improve and prove.
How it works
From first log line to proven fix
Four jobs, one product, for any kind of website. Scroll through what each one looks like on a sample store.
-
1
Observe
See every agent
Connect the access logs your CDN or server already keeps. Every crawler, AI assistant and browser agent shows up, with which company runs it, whether it's real, and what it read.
Agent requests by day last 7 days
- Mon
- Tue
- Wed
- Thu
- Fri
- Sat
- Sun
- AI and search
- Training
- AI assistants
- Agents
Top agents
Agent Requests Most read Status ChatGPT-User 6,210 /products/trail-runner-x Verified Claude-User 4,980 /collections/shoes Allowed PerplexityBot 3,105 /products/ridge-jacket 403 from WAF OAI-SearchBot 2,430 /sitemap.xml Verified GPTBot 455 /blog/ Blocked by robots.txt Sample screen: see every agent, for a fictional store. -
2
Test
Prove they can finish the job
Real AI agents work through checkout, sign-up, booking, search or a contact form on a schedule. When one fails, the replay shows the step where it got stuck, and why.
Checkout 0 of 3 passed daily · last run 06:00
Goal: “Find a men's trail running shoe in size 10 under $150, add it to the cart and get to the payment page.”
- 1Opened the home page
- 2Searched “trail running shoe”
- 3Opened Trail Runner X, $139
- 4Chose size 10
- 5Added to cart
- 6Couldn't press “Checkout”: a cookie banner covers it
Agent's reasoning at step 6: “The Checkout button is under a fixed banner. The banner's buttons have no accessible names, so I can't tell which one dismisses it.”
Sample screen: prove they can finish the job, for a fictional store. -
3
Improve
Fix what matters first
Findings and failed tests become one ranked to-do list, with who usually does each fix and how long it takes. Ship it, and it's re-tested automatically.
Fix list ranked by points gained for the effort
Fix Who usually does it Effort Gain Keep the cookie banner clear of the checkout button Web developer Hours +8 Name the 12 icon buttons and links Web developer Hours +5 Allow verified PerplexityBot through bot protection Security or CDN team Hours +3 Make size options real radio buttons Web developer Days +2 Re-tested Last week's fix, “Put prices in the product page HTML”, passed its re-test on Tuesday: +4 points.
Sample screen: fix what matters first, for a fictional store. -
4
Prove
Put a dollar figure on it
Revenue from people an AI assistant sent, and orders agents placed themselves, from your store's order data, with what each fix is worth.
Revenue from AI, 30 days$12,480+18%From AI referrals$9,860142 ordersOrders agents placed$2,62019 ordersRevenue by assistant
- ChatGPT$7,240
- Perplexity$2,610
- Claude$1,640
- Copilot$990
Fix the checkout banner and agents that reach the cart can finish: at this month's rate, that's worth about $1,900 a month.
Sample screen: put a dollar figure on it, for a fictional store.
Sample data for a fictional store. Click through the full interactive demo.
Observe
See the visitors your analytics can't
Most AI agents never run JavaScript, so Google Analytics and other tag-based tools don't see them. We read server and CDN data instead.
Every AI agent request, as it happens
Who visited, what they read, what your site sent back and how it changed over time. Dig into any day, see how fresh the pages agents read are, compare periods and export to CSV.
What you see
- Requests by agent, company and purpose
- Live requests, bandwidth and errors
- Pages agents read most, and what they get back
- New agents flagged as they arrive
Know which agents are real
Scrapers pretend to be Googlebot or ChatGPT to slip past bot protection. We check every request that claims a known agent against the IP ranges and DNS its operator publishes, and hand you ready-made rules to block impostors.
What you get
- Every known AI agent, grouped by the company behind it
- Who's missing, and whether robots.txt lets each one in
- Impostor watch across all your sites
- Block rules for your CDN, ready to paste
Follow agents through your funnel, and into AI answers
Journeys group each assistant's requests into visits, show which pages they read and where they end, and on a store follow them from product page to cart, checkout and order, so you see where agents give up. AI visibility (beta) shows whether assistants name you when customers ask, and who they name instead.
Questions it answers
- How far do agents get before they give up?
- Which page loses the most agent sessions?
- Do AI assistants recommend us, or a competitor?
- Which of our pages do they cite?
Test
Prove agents can finish the job
Ghost Agents are real AI agents that work through your key journeys from a plain-English goal, on a schedule, with a replay of every step.
Plain-English tests, real AI agents
Write the goal the way a customer would say it, like “find a size 10 trail shoe under $150 and get to checkout”, or start from a template. Each check runs several times, so you get a pass rate rather than one lucky result.
Every run gives you
- A step-by-step replay with screenshots
- The agent's reasoning at each step
- A pass or fail against conditions you set
- Daily checks, or hourly on Growth and above
Hear about it before your customers' agents do
A new bot protection rule or a release can cut agents off without anyone noticing. Alerts tell the right people by email, push or Slack. And Ghost Agents are built to be safe on a live site.
Guardrails
- Stop at the payment form and never enter card details
- Refuse buttons that place orders, pay or delete accounts
- Stay on your site, within step and time limits
- Sign every request (Web Bot Auth) so you can verify them
Improve
Know exactly what to fix, and in what order
AgentScore grades how well agents can get in, read and use your site, then turns every finding into work your team can plan and finish.
One score out of 100, with the reasons
AgentScore checks Access, Readability, Navigability and Task completion the way agents see your site: robots.txt and bot protection, content and prices in the HTML, named buttons and fields, pop-ups, and on a store the cart and checkout. It works out what kind of site it's looking at, so a software company, hotel or clinic is only scored on what applies.
What it checks
- Whether AI assistants and crawlers can get in
- Whether content and prices are readable without JavaScript
- Whether buttons, links and forms have names agents can read
- Whether agents can press your sign-up or add-to-cart button, and reach checkout
A to-do list your team will actually finish
Readiness findings and failing tests become one ranked list, with who usually fixes each one and how long it takes. Claude explains each fix in plain English. Choose which agents may use your site by what they're for, and we write the robots.txt and draft your llms.txt.
What you get
- Fixes ranked by score gained for the effort
- An owner and effort estimate for each fix
- Every shipped fix re-tested, so you know it worked
- robots.txt and llms.txt written for you
Prove
Put a dollar figure on AI
Leadership wants to know whether AI agents help the business. Give them the number, every week.
Orders from AI, counted
Connect your store's orders and see revenue from people an AI answer sent, and from orders AI agents placed themselves. Show whether being recommended by ChatGPT, Perplexity and other assistants is paying off, and whether fixing agent problems was worth it.
On Growth and above
- Revenue from AI referrals
- Agent-influenced orders
- Trends by assistant and agent
- The value of each fix you ship
Keep everyone up to date without logging in
A weekly report says, in a few plain sentences, what happened with AI agents on each site and the one thing to do next. Mission Control shows every site's health, readiness and journeys side by side, and how you compare with other sites AgentScore has scanned.
Built for leaders
- The weekly AI report by email every Monday
- Every site's health at a glance
- Your readiness against the benchmark
- Team activity log and CSV export on Team and above
Works with your stack
Nothing to install on your pages
Send logs from the CDN or server you already use, as Cloudflare Logpush, a Vercel log drain or any JSON log feed, connect your store's orders, and get alerts where your team already works.
- Cloudflare Logpush
- Vercel log drains
- Any JSON log feed
- Fastly, Akamai, CloudFront and Netlify, as JSON for now
- Shopify orders
- Slack
- Email and push alerts
FAQ
Platform questions
Do I need to add a script to my site?
No. Most AI agents never run JavaScript, so a tag would miss them. Ghost Agent Lab reads the access logs your CDN or server already keeps. Connect Cloudflare Logpush, a Vercel log drain, or any JSON log feed from your CDN or server.
Does it block bots?
No. Bot protection decides who to block. Ghost Agent Lab shows you which agents are real and what they're worth, and helps the ones you want succeed. It gives you ready-made rules if you decide to block impostors, and works alongside the bot protection you already have.
What does a Ghost Agent do on my site?
It works through one journey from a plain-English goal, like a customer would, and records every step. It stops at the payment form, never enters card details, refuses buttons that place orders or delete accounts, and signs every request so you can verify it. More about Ghost Agents.
How long does setup take?
Minutes for the first agents. Add your site, send your logs from your CDN or server, and agent traffic appears as it arrives. AgentScore scans the site straight away, and you can start a Ghost Agent test from a template.
See every agent. Fix every journey.
Every plan includes the whole platform and starts with a 30-day trial.