Shopping agents rarely fail at checkout. They fail at reading your product. CartTrace-100k, charted.

Most of the engineering effort in agentic commerce this year has gone into the last step. Checkout protocols, payment tokens, agent wallets, fraud rules for bots. If an AI is going to spend someone's money, the thinking goes, the hard part is the money.

A study published on October 28 suggests the hard part is somewhere else.

General Machines released CartTrace-100k, a record of 101,416 simulated shopping journeys. Each journey gave an AI agent a shopper's request, a store to shop in and a goal: find the right thing and get it into a cart ready to buy. The researchers then traced every step and labeled where it went wrong.

The answer, in one line: agents almost never fail at paying. They fail at reading the product and at understanding the shopper.

How the study worked, and what it cannot tell you

Before the numbers, the setup, because it shapes how far you can take them.

• One agent. Every journey used the same open-weight model, Qwen3.8-27B. Commercial agents like ChatGPT Agent, Muse or Alexa for Shopping may behave differently, better or worse.
• Stand-in stores. The 1,044 stores were built for the test across 25 product categories, not live merchant sites. Payments used test tokens.
• No bot walls. The test stores did not block the agent. On the real web they often do. General Machines notes that about one live merchant home page in five was unreadable to a declared agent.
• A reference point. The researchers also ran a scripted reference agent that knows each task, which gives a sense of what the stores allow when nothing is misread.

So this is a lab result about one agent's habits, not a measurement of the market. It is still the largest public trace of where a shopping agent stumbles, and the pattern is consistent enough to act on.

The headline: checkout works, reading does not

Across all journeys the agent succeeded 68.7% of the time. The scripted reference agent reached 82.4%.

Then the researchers broke failures down by stage. Each journey could fail at more than one stage, so the figures do not add up to 100.

Show the data

Stage	Share of journeys
Reading product facts	37.5%
Clarifying the shopper's needs	28.9%
Cart	1%
Checkout	1%
Payment	1%

Source: General Machines, CartTrace-100k: Where Shopping Agents Fail, October 28, 2026. Cart, checkout and payment are each reported as under 1%; shown here at 1% as an upper bound. One open-weight agent, 1,044 stand-in stores.

Reading product facts broke in more than a third of journeys. In a grocery aisle, think of mistaking a 12-pack for a single, picking the wrong flavor or missing an allergen line. Clarifying needs broke in more than a quarter: the agent did not ask about something that mattered, or assumed an answer.

Cart, checkout and payment, the parts the industry has spent the year building, each failed less than 1% of the time.

That lines up with what we saw in Forter's order data last week. Agents that drive a store's normal checkout pages are now placing most agent orders, and they manage the checkout fine. The question is whether they put the right thing in the cart first.

Many agents never open your product page

The single most useful finding for a brand is this one. In 37.6% of journeys, the agent added an item to the cart straight from the search results without ever opening a product page.

Behavior	Share of journeys
Added to cart from search results without opening a product page	37.6%
Handed over the cart without reading it back	50.7%
Gave up on a journey that could have been completed	7.9%
Gave up on substitute requests	27.4%

Source: General Machines, CartTrace-100k, October 28, 2026.

Think about what that means for a snack or supplement brand. Your product page might have a perfect comparison table, a clear allergen panel and a nutrition label in text. For more than a third of these trips, the agent never saw any of it. It saw the search result tile: a title, maybe a price, maybe an image, maybe a short line of text.

The same logic applies to search engines and marketplaces, where an agent may pick from a list of results. Whatever your title and snippet say is, for a large share of decisions, all the agent knows.

Half of journeys also ended with the agent handing the cart to the shopper without reading back what was in it. A wrong pack size would go through unchecked.

Asking one question changes everything

The clarifying stage had the most dramatic single result. The researchers ran journeys where two of the shopper's needs were held back from the original request, the way a real person might say "get me some oat milk" without mentioning they need it unsweetened and shelf-stable.

Show the data

Agent behavior	Journey success rate
Agent asked	62.1%
Agent did not ask	12.4%

Source: General Machines, CartTrace-100k: Where Shopping Agents Fail, October 28, 2026. Sandbox journeys where two of the shopper's needs were withheld from the opening request.

When the agent asked, it succeeded 62.1% of the time. When it did not ask, 12.4%. A brand cannot make an agent ask questions. But a brand can make the answers easy to find once it does, by writing the distinguishing facts (sweetened or not, shelf-stable or chilled, fragrance-free or not) in plain words where they can be read.

Substitutes are where agents struggle most

The researchers grouped requests into types. "Express" requests named the product. "Discover" requests described a need without a product name. "Substitute" requests named a product that was out of stock, where a similar one would do.

Show the data

Request type	Success rate
Express (named item)	69.6%
Discover (described need)	55.6%
Substitute (alternative)	52.3%

Source: General Machines, CartTrace-100k: Where Shopping Agents Fail, October 28, 2026. One open-weight agent, 1,044 stand-in stores.

Substitutes are a grocery and household problem more than most. Out-of-stocks happen weekly, and shoppers ask for "something like" a product all the time. The agent gave up on 27.4% of substitute tasks. If your product would be a good substitute for a bigger brand, say so in plain text: what it replaces, how it compares, and what is different.

Where CPG categories land

Category success rates sat in a fairly narrow band, but the order is still useful.

Category	Success rate
Jewelry (highest)	74.9%
Haircare and fragrance	70.4%
Baby and kids	68.9%
Pet	68.6%
Skincare and makeup	68.5%
Food (non-alcoholic)	67.6%
Non-prescription health	66.6%
Personal care	66.3%
Books (lowest)	63.9%

Source: General Machines, CartTrace-100k, October 28, 2026.

Most CPG categories cluster between 66% and 69%. Food, personal care and non-prescription health sit in the lower half. These are categories where small facts decide the purchase: dietary needs, ingredients, dosage, scent, size.

Weak stores make agents worse

The study also compared stores. Success ranged from 54.2% at the fifth-percentile store to 82.3% at the ninety-fifth. In weak stores, a fitting product appeared in the first search results in 80.9% of journeys against 85.3% in the top quarter of stores, and searches came back empty 12.3% of the time against 7.5%. Site search quality, a topic most brands treat as a conversion detail, decides whether an agent finds you at all.

One technical note for developers: cart journeys over WebMCP used a median of 50,000 tokens against 157,000 over MCP. Lighter interfaces are cheaper for agents to use, which may shape which stores they prefer over time. We would treat that as an early signal from one setup.

What to change on your site and in your feeds

1. Write your titles for an agent that will not click. Put pack count, size, flavor or scent, and the key dietary fact into the product title. "Oat Milk" loses to "Oat Milk, Unsweetened, Shelf-Stable, 32 fl oz, Pack of 6".
2. Make the first lines of your description do the comparing. Search snippets and feeds often show only the opening text. Lead with the facts that separate you from the next result, not with brand story.
3. Name your substitutes. If your product replaces a better-known one, say so plainly and say how it differs. Agents struggle most with substitute requests.
4. Test your own site search. Run your top 20 shopper requests through your search box. Count empty results and results where the first products are wrong. Agents depend on this more than people do.
5. Keep facts in text, not only images. Nutrition panels, allergen lines and ingredient lists in images are invisible to many agents. Tip 37 covers Shopify's knowledge base for store-level facts, and the new Catalog eligibility view in Shopify's Fall '26 Edition is worth checking at the same time.
6. Decide who can get in. If one live home page in five is unreadable to a declared agent, some of that is deliberate bot protection. Tip 33 covers how to decide which agents you let shop.

The broader picture fits McKinsey's consumer survey out the same day: shoppers value AI most for narrowing choices. CartTrace shows how narrowing goes wrong when the facts are hard to read. If you want to see how your products come across when an assistant reads them, check a few shopper questions by hand each month, or use a tool; XENA Intelligence tracks how products show up in assistant answers, with the results in SearchPanel.