The buyer reading an AI answer is increasingly not a person but a piece of software acting for one. That shift is real, but it is arriving slower and stranger than the announcements suggested. OpenAI quietly withdrew ChatGPT's Instant Checkout in March 2026, with fewer than 30 of Shopify's millions of merchants ever going live, and refocused on product discovery through its Agentic Commerce Protocol, which retailers including Target, Sephora, Nordstrom, Lowe's, Best Buy, The Home Depot, and Wayfair have adopted. Meanwhile Microsoft launched Copilot Checkout in the US, Walmart's own agent Sparky is lifting spend among users by about 35%, Google is bringing agentic tasks to Gemini in Chrome, and Perplexity's Computer agent now carries memory across sessions.
For brands, the practical question is not when agents will buy everything, but what changes when an agent, rather than a person, is the one reading about you. Here is where agentic commerce actually stands, how agents choose, what breaks for brands, and what to do now.
Where agentic commerce actually stands
The demand signal is unambiguous. Shopify reported that AI-driven traffic to its merchants' stores grew eightfold in the first quarter of 2026, with orders from AI-powered searches up nearly 13 times year on year. Walmart's Sparky users spend about 35% more on average than non-users. Amazon now lets merchants sync product feeds into Shop Direct so agents can discover, and in some cases complete purchases from, external retailer sites.
The supply side is messier. Forrester's assessment, quoted widely this spring, was that nobody has figured agentic commerce out and everyone has fear of missing out. Instant Checkout, announced as native cart-and-payment inside ChatGPT across millions of merchants, was pulled within months because, in OpenAI's words, it did not offer the flexibility the company aspired to provide; the unofficial reading is that AI-native checkout proved far harder than the demo. What replaced it is more durable and more relevant to visibility: discovery. Merchants share product feeds and promotions through the Agentic Commerce Protocol so their offerings are represented accurately inside ChatGPT, then complete the sale in their own checkout or a ChatGPT app.
The pattern to expect is agents doing the research and the shortlist, and humans or merchant-owned checkouts doing the transaction, for a while yet. For services and B2B the same pattern applies without any checkout at all: vendor research, booking, comparison, and policy lookup are already agent tasks.
How an agent decides
An agent choosing for a user reads three kinds of material, and the first two are structural.
Structured product and service data. Feeds shared through protocols like ACP, product schema on pages, and clean machine-readable facts about price, availability, specifications, and policies. An agent comparing options cannot use a slogan. It uses fields.
Live pages it can actually fetch. Agents fetch on demand. OpenAI's ChatGPT-User agent, for instance, retrieves a page when a user asks the assistant to look at it, and because the visit is user-initiated, robots.txt may not always apply. Other agents respect crawl rules strictly, and edge-level bot blocking, including default protections from CDNs, can turn a brand invisible to an agent without anyone deciding it should be.
The same third-party sources chat answers use. Reviews, comparison lists, community threads, and reference sources feed agents exactly as they feed answers, and our breakdown of where AI gets its answers applies unchanged. An agent that finds three inconsistent prices and two contradictory descriptions of your product does what a model always does with contradiction: it guesses, hedges, or moves on to a competitor.
What breaks for brands
Four failure modes account for most of the risk.
Blocked at the door. A robots policy or bot-management rule written in 2024 to keep content out of model training now also blocks the agents that would recommend and buy. Training crawlers, search crawlers, and on-demand agents use separate identifiers, and treating them as one switch is the most common self-inflicted wound we see, the same one that now also breaks ad approvals inside ChatGPT.
Facts that disagree with each other. Agents act on brand facts, and wrong facts get acted on. A stale price on a directory or a discontinued product still described on a review site becomes the agent's version of you. The causes and fixes are the ones in our guide to what happens when AI gets your brand wrong, with the stakes raised because an agent will not pause to ask.
Nothing structured to read. A brand whose product or service facts live only in marketing prose gives an agent nothing to compare. Structured data is not a ranking trick here; it is the interface.
Measuring the wrong thing. Agent-driven discovery leaves an even fainter trail than chat answers do. If a shopping agent shortlists three brands and the person then searches for one of them, your analytics sees a branded visit and nothing else, the blind spot described in our AI referral traffic analysis. The only place the shortlist is visible is inside the answer itself.
What to do now
The work is unglamorous and mostly already on your list if you have been serious about AI visibility.
Audit which AI agents can reach your site, per agent, and fix the rules you did not mean to write. Diploria's AI Access tool shows the live robots policy for each major AI agent and what it can and cannot fetch.
Publish accurate, structured facts for whatever you sell: product schema, service details, pricing where you can, policies, and availability, kept current with visible dates. If you are a retailer, get into the discovery feeds your buyers' assistants use.
Make the web agree with itself about you. Diploria's Brand Consistency Check keeps a ledger of your verified facts and compares what each engine says against it, which is exactly the check an agent will run implicitly.
Add shopping- and selection-intent prompts to what you track. "Which should I buy," "best option for," "is X worth it," and "book me" questions are the agent's raw material, and Question Radar surfaces the ones your buyers are asking that you are not yet monitoring. Then measure the outcome that matters: whether you are the option the answer recommends, per engine, over time. The free AI visibility check shows where you stand today.
For services and B2B
None of this is retail-only. Google has described agentic capabilities coming to Gemini in Chrome for tasks like booking appointments and ordering groceries, Perplexity's Computer agent researches companies and builds outputs directly from a query, and B2B buyers already delegate vendor research to assistants. The agent evaluating a law firm, an accounting practice, or a software vendor reads the same structured facts, the same reviews, and the same third-party sources as the one comparing headphones. The brands that win that evaluation are the ones that made themselves easy for a machine to read and hard to describe wrongly.