Written 15 October 2026.
Most shoppers will let an AI agent find, compare and even fill a basket. Most will not let it pay on its own. So the contest that matters right now is getting recommended and getting into the basket, and that is won with complete, accurate, current product data.
What the surveys say
Four recent surveys, different wording, same direction.
| Survey | Who was asked | Headline finding |
|---|---|---|
| Forrester (fieldwork April) | Online adults, US, UK, Canada | Three-quarters uncomfortable letting an agent complete a purchase and pay autonomously, even with preset spending limits and rules |
| Checkout.com, Agentic Commerce 2026 (9 June 2026) | Consumers in six markets | Would let an agent spend £177 per purchase on average without extra approval. 24% will never delegate purchases. 27% trust no organisation to run a shopping agent |
| Checkout.com, Peak Season 2025 (8 Dec 2025) | 4,000 consumers | 48% comfortable letting an agent make a purchase, 52% not. US respondents would let one spend up to $223, UK up to £204.53 |
| ACI Worldwide / YouGov (16 Sept 2026) | 3,300+ fashion and sportswear shoppers, US and UK, aged 18–65 | 7% would let an AI assistant buy without approval. 53% uncomfortable with AI buying on their behalf |
The pattern:
- Recommend, yes. Pay, no. Forrester found many people happy for agents to recommend and compare, some happy for them to fill a basket, and the vast majority stopping short of payment. The few comfortable with autonomous payment limit it to routine, low-cost or strictly rule-bound purchases.
- The spend ceiling is low. Roughly £150–£200 per purchase across the Checkout.com studies. Treat that as a ballpark, not a number. The two studies asked different questions a year apart.
- Routine categories go first. Checkout.com’s 2026 study puts willingness to delegate at 41% for groceries and 31% for household supplies, against 15% for financial services.
- People want brakes, not features. Top conditions in the 2026 Checkout.com study: spending caps (30%), instant revocation (29%), easy cancellation (28%). Forrester’s four recurring concerns: loss of control, errors and liability, distrust of the agent’s decisions, data security and privacy.
- Blame goes to the platform. Forrester found consumers expect the AI platform, not the retailer, to be held responsible when an agent gets a purchase wrong.
Limits. Wording moves these numbers a lot. “Comfortable letting an agent make a purchase” (48% in Checkout.com’s 2025 study) is not the same question as “pay autonomously, even with rules” (75% uncomfortable in Forrester’s). Three of the four studies were commissioned by payments companies with a stake in the answer. Read ranges, not single figures.
How agents pick what to recommend
Agents do not browse like people. They pull from structured product records first, then rank. OpenAI is the only major platform that publishes its criteria in any detail. Its help centre says ChatGPT considers:
- structured metadata from first- and third-party providers, such as price and description, plus other third-party content
- what the model already knows before it searches
- safety standards and product policies
Weighting shifts with context. State a budget and price matters more. Memory and custom instructions count too.
Once a shopper clicks a product, merchants offering it are ranked on availability, price, quality, and whether they are the maker or primary seller. Review summaries are generated from reviews on public websites and are not verified by OpenAI. OpenAI also says there can be a delay before price and shipping changes show up, which is what its direct product feed is for.
Google launched the Universal Commerce Protocol (UCP) on 11 January 2026, co-developed with Shopify, Etsy, Wayfair, Target and Walmart. It powers checkout for eligible US retailers inside AI Mode and the Gemini app, with the retailer staying seller of record. Google also announced dozens of new Merchant Center attributes for conversational shopping, covering things like answers to common product questions, compatible accessories and substitutes.
| What the agent checks | Where it is set | What you control |
|---|---|---|
| A record exists and is complete | OpenAI’s feed spec requires nine fields: item ID, title, description, URL, brand, seller name, image, availability, price | Get into Google Merchant Center and, if eligible, OpenAI’s direct feed. Shopify merchants are already in ChatGPT via Shopify Catalog |
| Product identity | Stable IDs, one row per variant, GTIN or MPN | Assign valid GTINs. OpenAI’s spec says not to invent an MPN to cover a missing GTIN |
| Price and stock | Required fields. Dates in the feed do not schedule changes | Update the feed when price or stock changes |
| Attributes that match the question | Material, colour, size, dimensions, category | Put specs in fields, not only in prose |
| Reviews | Review count and star rating fields, plus public reviews | Collect product reviews; keep count and rating from the same population |
| Returns and shipping | Return acceptance, return window, shipping price | Fill them. In OpenAI’s spec a missing shipping price means unknown, not free |
| Who sells it | Maker or primary seller is a named ranking factor | Sell your own products direct and name the seller properly |
| Checkout | OpenAI: a separately enabled integration. Google: UCP, eligible US retailers | Apply where eligible |
The UK point. OpenAI’s standard feed upload currently targets the US, and Google’s agentic checkout launched for US retailers. For most UK merchants today, the realistic win is the recommendation and the click, not the in-chat sale.
What to do, in order
- Get a complete feed into the surfaces. Merchant Center first. Apply for OpenAI’s direct feed if you qualify. Shopify Catalog if you are on Shopify.
- Make the feed and the page agree. Same price, stock, brand and variant on both.
- Fill the attributes people actually ask about. Material, dimensions, fit, compatibility. An agent answering “waterproof trail shoes under £100 for wide feet” needs those as data.
- Structure returns and shipping. Blank fields read as unknown.
- Earn reviews and third-party coverage. OpenAI builds review summaries from public websites, not just your own page.
- Treat freshness as ranking. Availability and price are named ranking factors. A stale “in stock” loses to a competitor’s accurate one.
Once everyone’s feed is clean, the edge moves to price and to being the primary seller. A feed cannot fake either.
What this can’t tell us
- OpenAI names its ranking factors but not their weights. Google publishes the protocol, not the ranking.
- Eligibility does not mean display. OpenAI’s own spec says so.
- Surveys measure stated comfort, not behaviour.
- This is a synthesis of published sources. I ran no tests of my own.
Called on 15 October 2026
The next consumer survey on agent payments from Forrester, Checkout.com or ACI Worldwide published before 30 June 2027 will still find a majority of respondents unwilling to let an agent pay without approving each purchase.
Right if: that survey’s autonomous-payment question shows under 50% comfortable. Wrong if: 50% or more. Checked on 30 June 2027 either way.
Method
Figures were taken from the primary page for each source, opened on 15 October 2026. The ACI figures come from Digital Transactions’ report of the ACI/YouGov study. Where two sources disagree, both are shown rather than averaged.
Sources
- Forrester, Consumers Aren’t Ready To Delegate Payments To AI Agents
- Checkout.com, press release, 9 June 2026
- Checkout.com, The $200 checkout, 8 December 2025
- Digital Transactions, An ACI Survey Indicates Some Hesitation About Agentic Payments, 16 September 2026
- OpenAI Help Center, Shopping with ChatGPT Search
- OpenAI Developers, product feed specification
- Google, New tech and tools for retailers to succeed in an agentic shopping era, 11 January 2026