The Agent CRO Loop: Observe, Analyze, Test, Measure
The short answer: An AI agent can run most of the conversion rate optimization loop when it is connected to analytics and testing through MCP. It reads traffic and funnel drop-off, inspects click targets and scroll depth on the leaking page, proposes a test with the evidence behind it, and, once a person approves, creates the A/B test and later reports per-variant conversions, with revenue available as the test goal. With Humblytics MCP the agent can create, update or stop a test and save funnels. Promoting the winner stays a human decision.
Most writing about AI CRO promises autonomous optimization and stops there. This guide shows the actual loop: which tool the agent calls at each step, what comes back, and where you say yes or no. It works in Claude Code, Cursor or any MCP client connected to mcp.humblytics.com/v1.
What an agent can do in CRO today, and what still needs you
The Humblytics MCP server has 45 tools. 35 only read. 10 change something:
| Write tool | What it does |
|---|---|
create_split_test |
Creates an A/B test and makes it live straight away |
update_split_test |
Renames a running test or changes its auto-stop days |
stop_split_test |
Stops a running test. A stopped test cannot be restarted. |
save_funnel |
Saves or updates a funnel definition |
delete_saved_funnel |
Deletes a saved funnel. Past results are not affected. |
create_property / delete_property |
Adds or removes a site on the account |
get_agents_checkout_link, buy_data_point_top_up, set_auto_top_up |
Billing. A top-up charges the saved card, so the agent confirms first. |
That is the whole write surface. The ad account tools only read, so the agent cannot touch campaigns or budgets.
The agent can: read traffic, pages, sources, entry and exit pages, clicks, forms, funnels and ad attribution. Suggest funnels and test ideas from your real data. Draft a test. Create it once you approve. Read results per variant. Stop a test.
You still own:
- Approving the launch. A new test goes live to real visitors the moment it is created. The
create_split_testdescription tells the agent to show you the plan first, and your MCP client asks before it runs a tool unless you turn that off. The server itself does not hold the test for approval, so keep tool approval on. - Calling the result.
get_split_testreturns sessions, conversions and conversion rate per variant. It does not return a significance verdict. Ask the agent to show its working, or check the result in the Humblytics app. - Shipping the winner. There is no promote tool. When a variant wins, you or your developer put the change on the site.
Step 1, observe: traffic, pages, entry and exit
Start wide. The agent needs to know where visitors arrive, where they leave, and which pages carry the most traffic.
| Question | Tool | What comes back |
|---|---|---|
| How is the site doing overall? | get_traffic_summary |
page views, sessions, unique visitors, bounce rate, average session duration. Bots excluded. |
| Is it going up or down? | get_traffic_trends |
page views and unique visitors per hour, day, week or month |
| Which pages matter? | get_pages_breakdown |
per page: page views, unique visitors, average scroll depth, bounce rate |
| Where do sessions start and end? | get_entry_exit_pages |
top entry and exit pages with session share |
| Where does traffic come from? | get_traffic_breakdown |
UTM, countries, devices, landing pages and referrers with session share |
Prompt:
"Summarize last month's traffic. Then list the 10 pages with the most views, with bounce rate and scroll depth, and the top exit pages. Point out any page that gets heavy traffic but loses most of it."
What you want out of this step is one or two candidate pages. A pricing page with high entries and high exits is a better place to start than a blog post with low traffic.
Step 2, analyze: funnel drop-off and click data
Now find the step where people leave, then look at what they do on that page.
Funnels. If you do not have a funnel yet, get_funnel_suggestions proposes funnels for a page from real traffic patterns, each with steps and a rationale. query_funnel then runs one. You give it ordered steps such as /, /pricing, /signup, a date range, and a mode: unbounded (steps in any order) or sequential (steps in order). It returns sessions and conversion rate per step, relative to the step before, plus overall conversion rate. Add breakdownBy such as device or utm_source to split it.
get_funnel_sankey shows the paths people take between steps, including drop-off nodes. Use it when the funnel says people leave at a step and you want to know where they go instead.
When a funnel is worth tracking every week, the agent can keep it with save_funnel.
Clicks, forms and scroll. On the leaking page:
| Question | Tool |
|---|---|
| What do people click on this page? | get_clicks_details: click targets with per-target share. This is the closest the server gets to a click heatmap. |
| Do people use the form? | get_forms_details: form submissions for the page with per-form share |
| Who is landing here? | get_page_details: page metrics plus UTM, device and country breakdowns |
| How far do they scroll? | get_pages_breakdown: average scroll depth per page |
Visual heatmaps stay in the Humblytics app. Over MCP, the agent works from the click and scroll numbers behind them.
Prompt:
"Run a sequential funnel from / to /pricing to /signup for the last 30 days, broken down by device. Find the step with the biggest drop. Then show what people click on that page and whether the main button gets a fair share of clicks."
Step 3, change: the test recommendation and your approval
get_split_test_recommendations takes a page and returns test ideas from that page's real data:
overall_analysis: a plain-language read of the pagekey_insights: specific observations behind itrecommendations[]: each withelement_selector,change_type(textorcss),control_value,variant_value,confidence_score(0 to 100) andexpected_impact
The tool tells the agent to prefer recommendations with a confidence score above 70.
Each recommendation maps straight into a test. element_selector becomes selector, change_type becomes op, variant_value becomes value. So the agent can turn an idea into a ready-to-run test without you writing code.
This is the approval point. Before it calls create_split_test, the agent should show you:
- The page, for example
/pricing - The change, with the current value and the new one
- The goal
- How long it runs (
auto_stop_days) - The evidence: the drop-off, the click share, the confidence score
create_split_test takes:
| Field | Options |
|---|---|
name |
A descriptive name |
page |
The page path |
type |
nocode for on-page changes, redirect to send a share of visitors to a different URL |
goal |
click_through, form_submission, bounce_rate, session_time, destination_page, external_destination or revenue |
auto_stop_days |
Optional. Stops the test after this many days. Without it, the test runs until stopped. |
variants |
For nocode: changes with a selector, an operation (text, css, html, image, delete or duplicate) and a value. For redirect: a URL. |
A Control variant is created for you. The test goes live as soon as the call succeeds.
Pick the goal that matches the money. If you sell online and have Stripe or Foxy connected, revenue scores variants on revenue per visitor. A variant can win on clicks and lose on revenue, and the goal decides which one you see. More on that in A/B testing for revenue per visitor.
Prompt:
"Get test recommendations for /pricing. Show me the top three with their confidence scores and expected impact. Draft a no-code test for the best one with revenue as the goal and a 14-day auto-stop. Do not create it until I say yes."
Step 4, measure: results and revenue
list_split_tests lists tests, filtered by active or complete. get_split_test returns one test in full: status, goal, auto-stop days, and per variant the label, sessions, conversions, conversion rate (as a decimal, so 0.166 means 16.6%) and the changes applied.
Because the tool returns raw counts, ask the agent to compare variants and say how confident it is, and to show the numbers it used. If a test has few conversions per variant, the honest answer is "keep running".
Two other write actions belong here:
update_split_testchanges a running test's name or auto-stop days. Use it to extend a test that needs more data.stop_split_testends a test and records an optional reason. A stopped test cannot be restarted, so treat this as final.
For tests on pages that paid traffic lands on, close the loop with revenue. get_ads_attribution joins Meta or Google Ads spend with sessions and Stripe or Foxy revenue per campaign. It shows whether the campaigns that feed the tested page are paying back. It needs an ad source and a revenue source connected, and returns empty revenue until both are.
Prompt:
"Show results for every active test: sessions, conversions and conversion rate per variant. Tell me which have enough data to call and which need more time. Do not stop anything."
Use the skills instead of writing prompts
The server ships 12 marketing skills as MCP prompts. Five use your live Humblytics data and map to this loop:
| Skill | Where it fits |
|---|---|
cro-optimizer |
Steps 2 and 3. Finds funnel drop-offs and writes ranked test ideas. |
heatmap-analyst |
Step 2. Finds ignored buttons and friction from element clicks and exit signals. |
ab-test-generator |
Step 3. Turns page and click data into test ideas, then sets up no-code tests. |
funnel-reporter |
Steps 1 and 4. Weekly or monthly reports from traffic through to revenue. |
revenue-attributor |
Step 4. Shows which campaigns pay back and where to move budget. |
The shared CRO benchmarks and frameworks are exposed as MCP resources, so the agent can compare your numbers against them. Browse all skills at humblytics.com/skills.
Ten prompts to copy
- "Summarize traffic for the last 30 days and compare it with the 30 days before."
- "Which pages get the most entries and the most exits? List the ones that are both."
- "Which pages get traffic but no form submissions?"
- "Suggest a funnel for /pricing and run it for the last 30 days."
- "Run my signup funnel broken down by device. Where does mobile drop off more than desktop?"
- "Show the Sankey paths from /pricing. Where do people go instead of /signup?"
- "On /pricing, what do people click? Does the main button get a fair share?"
- "Get test recommendations for /pricing. Show the ones above confidence 70 with the evidence. Do not create anything."
- "Create the test we agreed on: no-code, revenue goal, 14-day auto-stop. Show me the final payload before you call the tool."
- "Report every active test with per-variant sessions, conversions and conversion rate, and tell me which ones are ready to call."
Prompts 1 to 8 only read. Prompt 9 creates a live test. Prompt 10 reads.
FAQ
Can an AI agent run A/B tests on its own?
It can create, update and stop tests through Humblytics MCP. It should not launch one without you. A new test goes live to real visitors immediately, so the agent is told to present the plan first, and your MCP client asks before it runs the tool unless you have turned approval off.
What can the agent change on my site?
A/B tests, saved funnels, sites and billing. 10 of the 45 tools write: create_split_test, update_split_test, stop_split_test, save_funnel, delete_saved_funnel, create_property, delete_property and three billing tools. The other 35 read. Ad account tools are read-only.
Does the agent decide the winner?
No. get_split_test returns per-variant sessions, conversions and conversion rate, not a verdict. The agent can compare the numbers and explain them. You decide, and you put the winning change on the site.
Can the agent read heatmaps?
It reads the data behind them. get_clicks_details returns click targets with their share of clicks on a page, and get_pages_breakdown returns average scroll depth. Visual heatmaps are in the Humblytics app.
Can I optimize a test for revenue instead of clicks?
Yes. Set goal to revenue when you create the test. It scores variants on revenue per visitor. Connect Stripe or Foxy first so there is revenue to measure.
Which agents does this work with?
Claude Code, Cursor and any client that supports MCP. Remote clients connect to mcp.humblytics.com/v1 over streamable HTTP. Developers can also run the server locally over stdio.
How is this different from an AI CRO tool?
Here the agent is your own: Claude Code, Cursor or another MCP client, working in the place you already use it, against your live Humblytics data, with every step visible as a tool call you can read. See our AI CRO guide and list of AI CRO tools for the wider category.