Some of your next customers will not click through your shop themselves. They will ask ChatGPT's agent, Claude or Perplexity to find a product and put it in the basket, and the agent will use your site the way a person does: reading the page, clicking buttons, filling in fields. If it cannot finish, it moves on to a shop where it can.
The agency MERJ has published the most detailed public look so far at where those agents fail. Starting in November 2025, it built 100 controlled test pages, each isolating one interface pattern, and asked several frontier agents to complete a task on each without scripting the steps. It then repeated the runs as models improved and moved on to real production sites. A test counted as a success if the agent reached the required end state, even after more than one attempt.
The headline: average success across all tests rose from around 30% to around 60% between November 2025 and March 2026. Agents are getting better quickly. But the same patterns still sink them, and most of those patterns are things a shop owner can fix in an afternoon.
The patterns that matter for a shop
MERJ grouped its tests into six stages, from what the agent can see to whether it can be tricked. The chart below picks the tests that map most directly onto a product page, cart and checkout. Look at the right-hand bars: even after five months of model improvements, an agent facing an icon-only button or a transparent overlay still fails more often than it succeeds.
Two results stand out. The pre-checked subscription add-on barely moved (about 26% to 31%): agents rarely notice a box that is already ticked, which is bad for the customer and, once disputes start, bad for you. And the cookie banner is still a coin toss for many runs, on a page every visitor sees first.
The fix list, in order
1. Buttons that say what they do
An icon-only basket button with no accessible name was one of the weakest results (about 34% in November, 45% in March). Agents that read the page structure rely on the button's name. Use a real <button> element, not a styled <div>, and give it a label that describes the action: "Add to basket", not "cart icon". If your theme shows only an icon, the label can be hidden visually but must be present.
One warning from MERJ's most striking test: wrong labels are worse than missing ones. On a panel where the Save button was labelled "Cancel" in the markup, ChatGPT's agent clicked Cancel first in 25 out of 25 runs. If your theme was edited by several people over the years, check that the labels are true, not just present.
2. Nothing invisible on top of your buttons
A transparent overlay that catches clicks (often left by a slider, a chat widget or a promo layer) succeeded only about 41% of the time even in March. Open your product page, try to click Add to basket right after the page loads, then again after scrolling. If anything sits on top, even for a second, fix it.
3. Cookie banner and popups that close cleanly
A full-screen cookie banner got through about 64% of the time in March; a newsletter popup that appears after a delay, about 58%. Keep the banner's buttons as plain buttons with clear names ("Accept", "Reject"), and do not open a newsletter popup in the middle of a checkout. Every interruption is a chance for the agent to lose its place.
4. No pre-checked extras
Insurance, gift wrap, a subscription: if it is ticked by default, an agent will usually buy it without noticing. Customers who discover the extra later ask for refunds. Leave every paid extra unticked.
5. Variants that can be picked in any order
Product pages where colour only becomes available after size, or where choosing one option resets another, scored about 52% in March. Make each option selectable on its own, and show clearly which combination is selected before Add to basket.
6. Errors and confirmations that stay on screen
MERJ recommends success and error states that persist rather than four-second toasts. A message like "Added to basket" or "Postcode is invalid" that fades away is easy for an agent to miss, and it then repeats the action or gives up. Show the message in the page until the customer moves on, and mark invalid fields so the agent knows which one to fix.
7. Fields that clear when you type
Typing into a pre-filled field often appends rather than replaces, so "1" in a quantity box becomes "12". In March, agents got this right about 66% of the time. Keep quantity and postcode fields empty, or select their content on focus.
Test one real journey, not a score
You will see tools offering an "agent readiness score" for your store. MERJ's warning is blunt: its Cancel button test would score close to perfect on readiness checkers, because those checks confirm that labels exist, not that they are true. A score cannot see the failure.
Instead, run one real task once a month:
- Open an agent you have access to (ChatGPT's agent mode or a similar tool).
- Give it a specific task: "On [your shop], find the blue linen shirt in size M and add it to the basket, then go to checkout and stop before payment."
- Watch what it does. Note where it hesitates, clicks the wrong thing, or declares success too early (MERJ saw agents stop on the search results page and report the job done).
- Fix the first thing that went wrong, and run the same task again.
If you sell in more than one language, run the test in each. MERJ cites research showing agents perform worse outside English, by about 9 to 18 percentage points depending on the model.
Every fix on this list also helps human customers, especially people using screen readers, keyboards or small phone screens. That is the honest pitch: you are not optimising for robots, you are removing friction that was already costing you sales.
