Find where agents get stuck

AI agents (Claude in Chrome, ChatGPT agent, Perplexity Comet, Browser Use…) now shop on behalf of people. Many of them reach your product pages and add to cart, then stop at one step that humans get through without thinking: a custom delivery-slot picker, a cookie wall, a "verify you are human" check, a button with no accessible name.

The agent readiness funnel finds that step, tells you which agents it blocks and how much it costs, and suggests how to fix it.

Open it in Agents & humans → Agent readiness funnel. A short version also sits at the top of Agents vs humans — overview. Available on Growth, Pro and Scale (feature Agent funnels and friction points).

What it measures

The report follows two groups of sessions along the same journey:

• Action agents: sessions driven by an AI agent or automation that browses and clicks for someone.
• Humans: people browsing on their own.

Indexing bots (crawlers, AI readers) are left out: they read pages but never try to buy.

The default journey is the shopping path most stores share:

#	Step	Counted when the session…
1	Home & landing pages	views any page
2	Search	views /search*
3	Product pages	views /p/*
4	Add to cart	clicks the element named add-to-cart
5	Shipping	views /checkout/shipping
6	Delivery slot	clicks confirm-slot on /checkout/delivery-slot
7	Payment	views /checkout/payment
8	Order confirmed	sends a purchase event

Use Edit steps in the funnel builder to check another journey (sign-up, booking, quote request). In the funnel builder, whenever you compare action agents and humans, the same three views appear above the results.

Furthest step or steps in order

Agents rarely take the human route: they often land straight on a product page from an assistant's answer, skipping home and search. So the default journey counts the furthest step each session reached: a session that reached Payment also counts for every step before it. This shows where each group stops, without penalising agents for taking a shortcut.

Custom steps and saved funnels are counted in order within the time window (default 60 minutes), like a classic funnel. The report says which mode it used under the chart.

How the blocking step is chosen

For each step, the report computes the pass-through rate of each group: the share of sessions that reached the previous step and also reached this one.

The blocking step is the step with the largest gap between the human and the agent pass-through rate, as long as:

• humans are ahead by at least 15 points, and
• at least 20 sessions of each group reached the previous step, so a handful of sessions can't trigger a false alarm.

If no step passes both tests, the report says No agent wall: agents get through your journey about as well as humans.

The upside

The upside estimates what fixing the step is worth over the selected period:

1. Extra agent sessions = agents who reached the previous step × human pass-through rate − agents who actually passed.
2. Extra orders = extra agent sessions × the share of agents that go on to order once past the step (the human share when fewer than 10 agents got through).
3. Extra revenue = extra orders × the average order value of agent orders.

It is an estimate, not a promise: it assumes agents behave like today's successful agents once the step is fixed.

The three views

All three show the same numbers. Switch with the buttons above the chart; your choice is remembered.

Duel bars (default). Two bars per step, agents then humans, as a share of the sessions that started. Under each step, the pass-through rate of each group (A 24%, H 80%). The blocking step is highlighted with the Agent wall marker.

Twin streams. Each group is a stream whose width follows the number of sessions; the side ribbons are the sessions lost at each step. The agent stream visibly collapses at the wall while the human stream carries on.

Journey storyboard. One card per step with a sketch of the page, and a progress bar per group under each card. Best for sharing with designers and developers: "agents are blocked here" points at a page they know.

The diagnosis panel

Under the chart, for the blocking step:

• Headline figures: agent and human pass-through, median time on the step for each group, and the upside.
• What agents do on this step: the elements each group clicks, in clicks per session. Repeated clicks by agents on the same element (div.slot-picker 3.4 per session vs 1.4 for humans) usually mean the agent cannot tell its click worked. The Synthetic column is the share of agent clicks sent by script rather than a real pointer, which your page may ignore.
• Agents most affected: pass-through rate by agent (for example claude-in-chrome 28% of 453 sessions), so you know which agent to test with first.
• How to let agents through: up to five rule-based suggestions drawn from the above.

Fixing common blockers

What you see	Likely cause	Fix
Agents click the same control again and again	Custom widget (div with click handlers) with no visible state change an agent can read	Use native controls (<button>, radio buttons, <select>) or add role and aria-pressed/aria-selected
A button only humans click	The control has no accessible name or is drawn in a canvas	Give it visible text or an aria-label and a stable data-ah-name
High share of synthetic clicks, low pass-through	The page ignores clicks without isTrusted	Accept synthetic clicks, or move the check to the server
Agents spend much longer on the step	The agent hunts for the next action	Make the main action obvious: one primary button, clear text, no hover-only menus
Agents stop on a /verify or captcha page	A bot wall treats agents like scrapers	Let known agents through, or ask for confirmation later (at payment, where the person is present)
Agents stop at a choice humans make easily	No sensible default	Preselect the most common option so the agent can simply confirm

After a fix, compare the same report over the weeks before and after (comparison in the date picker) to see the agent pass-through rate move.

From the API and from AI assistants

• API: POST /v1/properties/{pid}/agents/friction with { "date_range": "last_28_days" }, optionally steps, funnel_id, window_seconds or mode. See Agents and bots.
• MCP: ask "Where do agents get stuck on my site?". The assistant calls the agent_friction tool and answers with the blocking step, the gap, the upside and the top fixes. See MCP tools.