No Fluff Advisory · Research · Commerce and AI agents Full paper

Plain-language edition

Who Does the Shopping?

What AI shopping tools really do today, in plain words

By Evgeny Popov · No Fluff Advisory · October 2026

5 minto read, about 1200 words
92.4Flesch reading ease, measured on this page with textstat 0.7.13 (the full paper scores lower)
1011atomic claims checked in the full paper

Current status · 10 October 2026

Published 10 October 2026. Documentary cutoff 9 October 2026; a verification addendum on 10 October added primary pages opened by an external reviewer. Review: three internal AI review passes, an external AI critical review and a primary-evidence check (OpenAI's ChatGPT), each finding adjudicated in the corrections ledger; no independent human peer review. Open: the author's statements on commissioning, funding, draft access and editorial control, and the exact wording of his affiliations (see the disclosure). Update history

What this is

This is the short, plain version of a longer research paper. It has the same findings and the same limits. The full paper has the sources, the data and the corrections.

The facts are as of 9 October 2026. A few were checked again on 10 October against company pages. AI tools helped write and check it. No outside expert has reviewed it.

The big picture

Shopping has many steps. You find things. You choose one. You pay. Then it comes, and you may send it back.

Each new way to shop took over some of these steps. Catalogues did. Websites did. Phones and social apps did. None of them made the old ways go away. Most US shopping still happens away from the web. Online is about one dollar in six. Read more

Now AI tools can help with the steps too. This paper asks which steps they really do, and who gains.

What we found

AI is used for shopping research; buying on its own is rarer. In big surveys, close to half of shoppers say they use AI somewhere when they shop. Most use it to look things up and to compare. Far fewer say they would let an AI buy for them. Surveys ask this in different ways, so their numbers differ a lot and should not be pooled. Read more

We found no fair count of AI purchases. Visa says it ran a few hundred test payments started by AI. But we found no public total of real purchases an AI made on its own. Companies may have their own numbers. Visits from AI tools to shops are still few in the data we found, though each study counts them differently. Big forecasts for 2030 count different things. They should not be added up.

Most AI tools help you find. A few let you check out. Every live tool we looked at helps you find and compare. A few let you approve a ready-made checkout inside the tool: Microsoft Copilot, Google for some shops, and Perplexity with PayPal. ChatGPT tried checkout inside the chat. In March 2026 it moved most shops back to their own site. Its help pages still say some products can be bought in the chat; we could not test that. Walmart now runs its own app inside ChatGPT. Read more

An AI buying on its own is rare. The clearest case we found is on Amazon. If you pay for Prime, you pick an item and a price. When the price drops to that level, Amazon buys it for you. That is one item, in one store, on your terms. We read about it; we did not see it. One payment rulebook already lets an AI buy within limits you set ahead of time. We found no sign it is in wide use. Read more

The rules differ on who says yes. One payment rulebook lets you set limits in advance; the others expect you to approve the order. A payment limit is usually just the cart total, not a budget you chose. Shops track returns, but the rules we read give an AI no standard way to start one. Who pays when an AI buys the wrong thing is up to existing law and card rules. Read more

Ads moved inside the answers. OpenAI, Google, Amazon and Microsoft now put paid spots inside their AI shopping answers. The companies say their free results are fair. We found no outside check of that. Shops can see what they pay for. Google now shows shops how often they appear for free in its AI answers, against rivals Google picks by a rule it does not publish. We found no such report from the others. Read more

Fees differ, and the customer may not be yours. Some AI tools say they take no cut beyond card fees, though that can change. For Google, if a shop connects directly, we could not find the fee. ChatGPT's old checkout had a "small fee", reported as 4%, but no one confirmed the rate or what it was charged on. Marketplaces charge about 6% to 15%, depending on the category, plus other charges. Which way earns a shop the most depends on its own costs, so we do not rank them. What a shop learns about the buyer depends on the tool, and we could not see it. Read more

Our own test

We could not reach any real AI shopping tool from where we did the research. So we built a small fake shop and let an AI shop in it. Read more

The AI did well in a small, easy shop: 15 good orders out of 16. A hidden note told AI shoppers to buy a costly model and skip asking. It reached 12 sessions. None bought that model, picked fast shipping or went over budget.

Where a person had to approve, the AI tried the wallet first, was told no, then asked. We had told it to buy and never told it to ask. No order went through without approval. And the shop itself only checked the amount. In our test it took the wrong item when the other check was switched off on purpose.

The lesson: what you asked for is safe only if something outside the AI checks it before you pay.

This test used one fake shop and one kind of AI. It does not tell you how any real tool behaves.

What a shop owner should do

  1. Fix your product data. Each size and colour needs its own code. Show real stock, the full price with delivery, a true delivery date and clear return rules.
  2. Keep your own checkout and your customer details where you can.
  3. Count visits from AI tools on their own, not mixed with search.
  4. Test AI checkout in one area, with a rule for when to stop. A test shows what happened. To see what you gained, keep a group of shoppers out of it to compare.
  5. Grow a new way to sell only if the extra sales it brings pay for all its costs, after the sales it just moves from elsewhere.

Read more

Three ways it could go

AI could buy for us widely by 2030. That needs trust, tools that work every time, fair fees and clear rules on mistakes. Most of that is not here yet.

AI could buy in narrow ways, inside one store, for known items. This path needs the fewest changes. It is a possibility, not a forecast.

It could stall, if a big mistake or new rules break trust.

These are possible paths, not forecasts. Read more

Limits

We could not open most company web pages. We read them through search results. The paper marks each such fact. We could not test real AI shopping tools. We could not measure whether AI makes people buy more.

About the author

Evgeny Popov wrote this for No Fluff Advisory. He works at Samba TV. He helps lead a working group at AgenticAdvertising.org, which writes rules for AI in advertising. No Fluff Advisory sells advice. AI tools from Anthropic helped with the research, the test and the writing. AI tools from OpenAI checked it at the author's request. No outside expert has reviewed it. The author has not yet said who paid for the work, or whether any company saw it before it came out.

Update history 5 dated notes
  • Research draft 0.9 assembled for review on the evidence cutoff date.
  • Review passes 1–3 applied: 93 review findings (81 sustained, 2 partly sustained, 10 rejected), each with its change in the corrections ledger.
  • Revised after an external critical review: 22 findings adjudicated (R4 rows in the corrections ledger); passage provenance added to every claim; 115 supported claims narrowed, split or capped; survey comparison table, organic-reporting map, fee-type table, displacement worksheet and sandbox audit added. Documentary cutoff unchanged (9 October 2026); retrievals and reruns on 10 October are dated as such.
  • Edition 1.0 published. Primary-evidence handoff integrated (36 pages opened by the external reviewer; R5 rows); fee routes split; survey attributions corrected; figures given a high-fidelity pass.
  • Corrected after a live final review (R6 rows in the corrections ledger): TikTok Shop's referral-fee refund rule (fee returned less a 20% administration charge, capped at $5) now applied, so its illustrative base case is $18.33 per order, not $17.40, and the share of genuinely new orders needed to add contribution before fixed costs falls to 8.2%; the sixteen executed agent sessions now show their outcome instead of 'not run'; the holdout explanation, protocol scope, traffic comparison and planted-instruction summary narrowed; source-register split, route claim references and print legibility repaired.