Where AI Shopping Works, and Where It Breaks

A mom asked ChatGPT to look through her past Target orders and fill her pickup cart. It added 83 items, "all of which is exactly the brands I would usually order." A Reddit shopper asked Amazon's Rufus whether a seat cover would fit their car: "Amazon knows what car I drive. I asked if the seat cover would fit. After a lot of thinking it told me it fit a different car from the same manufacturer." Both are AI shopping. The first worked because the assistant used what it knew about the shopper. The second had the same kind of information and got the answer wrong.

We wanted to know which of those two stories is more typical, so we went to where shoppers talk about it unprompted.

Do AI shopping assistants actually work?

Sometimes, and it depends on the task. Of 183 cases shoppers described in October 2026, 95 worked, 31 were mixed and 57 failed; grocery lists and meal plans worked most consistently (18 of 23).

In October 2026 we read 432 posts from TikTok and Reddit and the 5,102 comments under them. We set aside marketing: AI-tip accounts, prompt and course sellers, and brands talking about getting recommended by ChatGPT. Comments under those posts stayed in, since that is often where people say what actually happened when they tried it. Then we pulled out every case where someone used AI for a shopping task and said how it went. That left 183 cases and 226 general opinions.

People use AI to shop for ordinary things. The most common tasks were researching and comparing products (41 cases), discount codes (24), grocery lists and meal plans (23), finding a better price (19), beauty and skincare (18), filling a grocery cart (12) and outfits (11). Six cases had an agent place the order itself. ChatGPT is named in most cases; Amazon's Rufus, Claude, Gemini and Copilot come up too.

These counts are not a success rate. People post wins more than losses, and some searches were written to find wins. What the cases do show, consistently, is why it worked and how it broke.

Most of the conversation about AI shopping is about the catalog: does the assistant find the product, is the price right, does checkout go through. Those questions matter, and shoppers do report breakage there. But they are not what shoppers credit when it works.

What makes AI shopping recommendations work?

Knowing the shopper. In cases that worked or partly worked, people most often credit the assistant with building a plan around their budget and household (33 cases) and fitting the needs they stated (31).

In the cases that worked or partly worked, people credit the assistant with using details about them. The reasons cited most are building a list or plan (33 cases), fitting their needs and constraints (31), and explaining what matters for their situation (20). Finding a lower price (13) and the exact item (9) come after.

Grocery lists and meal plans worked in 18 of 23 cases. The input is a budget, a household and the brands they already buy; the output is a list they can use.

Fit shows up outside groceries too. Creator @mariesjourneyforward traced her daughter's foot, measured it and gave ChatGPT the brand's size chart: "It compared the measurements and told me exactly which size to order. And it got it right the first time."

The Target cart in the opening worked for the same reason. Creator @motherhoodlane asked ChatGPT whether it could check her previous purchase history for the brands she usually buys. It could, and the cart came back with her brands, ready for her to review.

Why do AI shopping assistants fail?

Two reasons recur: wrong facts about products, such as made-up details (10 cases) and wrong or missed prices (10), and not following what the shopper asked, such as generic answers (15) and ignored requirements (9).

The failures fall into two groups.

The first is facts about products. Among cases that failed or partly failed, ten involve made-up products or details, and ten involve a wrong price or a cheaper option it missed.

The seat cover in the opening is one of these. Creator @sarahbabula, a chemist, pasted a serum's ingredient list into ChatGPT and walked through what it got wrong: "I just copy and pasted the ingredient list and asked what was in it. They got it completely wrong."

The second group is not following what the person asked. Fifteen failures are generic answers that could have gone to anyone, and nine ignore a requirement the shopper stated.

Another Reddit commenter, after months of discussing a TV with ChatGPT: "It can also be a total Yes man, and lead you into purchasing the entirely wrong thing."

Agents add a third way to go wrong: carrying out the order. Creator @automaticallyadhd had ChatGPT order her groceries through Instacart, "and it went pretty well other than ordering 18 pounds of bananas." Her own verdict: "I just didn't review it closely enough." Creator @bri.mccarty, who runs a childcare center, had an AI agent fill her Aldi cart. It added cereal and goldfish crackers the center doesn't serve, until she gave it ground rules.

Would shoppers let AI buy for them?

Mostly not yet. Twelve opinions say they would not let AI order or pay for them, and one asks whether it could.

Mostly not yet. Twelve opinions say they would not let AI order or pay for them; one asks whether it could. In most agent cases that worked, the person reviewed the cart before paying.

Opinions on AI shopping overall are split rather than settled: 38 call it useful, 38 say it is useful for some steps but not the final call, 48 say it is not useful, and 50 raise trust concerns. The most-liked comment on ChatGPT's own shopping launch video asks: "How do we know you aren't omitting certain products or favoritizing certain ones?"

What this means for teams building AI shopping

Catalog accuracy is necessary: live prices, items in stock, specs that are real. Shoppers notice when those are wrong, and they say so. But the cases that worked were not about the catalog. They worked because the assistant fit the answer to a specific person.

That is harder to measure. A price can be checked against a feed. Whether a recommendation was right for this person can only be checked against what people like them actually bought, and why they chose it. That is the ground truth most shopping evals don't have. Ario builds it from verified purchase histories across retailers, connected with the shoppers' consent, and from those same shoppers explaining the decisions behind their purchases, which also makes their answers more reliable. See it in Ario CoreLens, or browse sample purchase data.

Shoppers already say when AI shopping works: when it knows them. The test for any shopping assistant is not whether it can find a product. It is whether it can answer what this person should buy, and why.