For a year, the agentic commerce conversation has run on announcements. Platform launches, protocol unveilings, partnership press releases, blockade threats. On October 22, the conversation finally got something better: behavior.
FirstPageSage’s “Agentic Commerce Statistics: October 2026” is the first sizable behavioral dataset of the agent-shopping era. The methodology deserves attention because most of what passes for agentic commerce data does not: a panel of 4,936 US online shoppers, anonymized order and referral data from 43 ecommerce, travel and insurance brands, and 18,400 shopping prompts run across the major AI platforms, normalized against US Census ecommerce totals. The study window, July 13 through October 16, 2026, happens to contain the single most important event in the market so far: the September 8 consumer launch of Meta Muse.
The dataset kills three comfortable narratives at once. Autonomous checkout is not the near future: shoppers insist on approval by more than five to one. The platforms hosting shopping conversations are not the platforms completing purchases. And the factor that most powerfully decides which products agents buy is not product quality at all. It is whether the merchant paid to be inside the checkout partnership.
The Funnel Collapses Exactly Where Money Changes Hands
The study’s core finding is a stage-by-stage adoption funnel, and the drop-off pattern is stark. Among US online shoppers between July and October 2026:
- 61.4 percent used AI for product research and discovery, up 5.8 points since July
- 48.7 percent used AI for price and feature comparison, up 6.3 points
- 39.2 percent used AI for review summaries, up 4.1 points
- 14.6 percent let an agent build a cart, up 5.2 points
- 6.8 percent completed an agent checkout with their approval, up 3.9 points
- 3.4 percent enabled scheduled or price-triggered reorders
- 1.3 percent allowed a fully autonomous purchase with no approval at all
Read those numbers as a shape. AI is now the default starting point for shopping research: roughly six in ten shoppers begin there, and the research and comparison stages are still growing fast. But adoption collapses by half at cart-building and by ninety percent at payment. The fastest relative growth in the entire window came at the two money-touching stages: cart-building grew 55 percent in relative terms and approved checkout grew 134 percent. That is the market’s direction of travel, and its speed. But the level matters more: even after a 134 percent growth spurt, approved agent checkout is a behavior of one shopper in fourteen.
Agent-completed orders as a share of all US ecommerce tell the same story from the other side. The quarterly series runs 0.02 percent in Q4 2024, 0.09 percent in Q3 2025, 0.27 percent in Q2 2026, and 0.68 percent in Q3 2026: the largest quarterly jump ever recorded, and still under one percent of orders. AI-referred visits to retail sites, by contrast, hit 4.73 percent of US retail site traffic in Q3 2026, up from 0.43 percent two years earlier, and those visits converted at 1.6 times the rate of all other retail traffic. The research layer scaled years before the transaction layer. That asymmetry is the single most important structural fact in the market, and it has a direct consequence: whatever the agent reads during research now influences a vastly larger pool of spending than whatever the agent is allowed to buy.
The Approval Economy: Five to One
Here is the number that should reorganize every product roadmap in the sector: agent checkouts with shopper approval outnumber fully autonomous purchases by more than five to one (6.8 percent versus 1.3 percent).
The industry’s founding assumption was that autonomy is the destination and that friction is the enemy. The data says shoppers have a different model in mind: the agent as researcher, comparer, cart-builder and payment-preparer, with the human as the final gate. Shoppers overwhelmingly want a final say.
This is not timidity. It is risk pricing. The median spending limits shoppers set when allowing agent purchases reveal a calibrated, category-specific trust: $620 for travel, $310 for home goods, $240 for consumer electronics, $112 for groceries, $68 for beauty. People are not afraid of agents spending money. They are afraid of agents spending money badly on things that are hard to return, hard to compare, or dense with fake signals.
And that reframes what the approval click actually is. When a shopper approves an agent’s cart, they are approving a recommendation they almost certainly did not verify line by line. The approval moment is not scrutiny. It is a signature on evidence the agent gathered. Which makes the quality of that evidence the entire ballgame.
Libraries and Cash Registers: The Platform Asymmetry
The second uncomfortable finding: the platforms hosting shopping conversations are not the platforms completing purchases.
| Platform | Share of AI shopping sessions | Share of agent-completed purchases | Average order |
|---|---|---|---|
| ChatGPT | 41.3% | 9.4% | $87 |
| Google Gemini / AI Mode | 27.6% | 23.1% | $74 |
| Amazon Alexa for Shopping | 14.2% | 31.8% | $52 |
| Meta Muse | 6.9% | 29.7% | $118 |
| Claude | 5.8% | 2.6% | $141 |
| Microsoft Copilot | 4.2% | 3.4% | $93 |
ChatGPT hosts 41.3 percent of AI-assisted shopping sessions but completes fewer than one in ten agent purchases, because since March 2026 it has deliberately pushed checkout back to merchants’ own sites. OpenAI’s stated reason, when it retreated from Instant Checkout: “the initial version of Instant Checkout did not offer the level of flexibility that we aspire to provide, so we’re allowing merchants to use their own checkout experiences while we focus our efforts on product discovery” (OpenAI, March 24, 2026, via C9 Group’s platform guide).
The completion leaders are the platforms with integrated rails: Amazon’s Alexa for Shopping and Meta Muse together account for 61.5 percent of agent-completed purchases from just 21.1 percent of sessions. And Muse is the accelerant behind the Q3 jump: 71 percent of the quarter’s increase in agent-completed orders traced to Muse, which launched six weeks before the window closed and completes purchases through Stripe Link, Shop Pay and PayPal.
The order values carry their own signal. Claude agents complete the largest orders at $141 on average, consistent with research-intensive and business purchases. Muse averages $118. Alexa averages $52, the profile of a replenishment engine: groceries, household staples, reorder loops. Different platforms are becoming different kinds of buyers, which means product trust decisions are increasingly made by different kinds of readers.
The Partnership Multiplier: Distribution Gets Bought
The dataset’s most consequential table quantifies what formal checkout partnerships are worth, by comparing purchase completion rates for merchants inside each partnership against comparable merchants outside it:
- Meta Muse retail partners (Walmart, Gap, Sephora, Wayfair, Best Buy): 3.8x
- Muse and Shopify (Shop Pay): 3.4x
- Muse and Stripe (Link): 2.9x
- Google Universal Commerce Protocol: 2.6x
- Amazon Buy for Me: 2.3x
- Muse and PayPal: 2.2x
- Walmart Sparky inside ChatGPT: 1.9x
- Claude Commerce Agents: 1.6x
- OpenAI product feeds (data only, no checkout): 1.4x
A merchant an agent can buy from in one step completes agent purchases at up to nearly four times the rate of an equivalent merchant the agent can only recommend. Partnerships that merely improve product data without enabling checkout produce the smallest lift.
Now project forward. FirstPageSage’s own forecast expects AI platforms to start charging merchants for checkout modules in the first half of 2027, with in-experience ads expanding through 2027. When placement in the agent’s consideration set becomes a paid product, the recommendation layer inherits the economics of sponsored listings. The platforms will sell access to the moment of approval. Nobody is planning to sell verification of what earns that approval.
Amazon, for its part, runs the mirror strategy: it buys from outside merchants through Buy for Me, having onboarded more than 400,000 outside shops and 100 million-plus products into Shop Direct in March 2026, while blocking outside agents such as Muse from its own store. US stores can be included in that program without asking, and must email to leave. The walled garden now reaches outward with its own agent while sealing itself from everyone else’s: a structure optimized for Amazon’s margin, not for the shopper’s evidence quality.
The Category Paradox: Where Trust Is Scarcest, Delegation Is Lowest
The category data hides the sharpest insight in the study. Consumer electronics is the most AI-researched category of all: 68.1 percent of electronics shoppers used AI to research. It is also among the least delegated: only 5.2 percent let an agent complete the purchase, with a $240 median spending cap. Home goods: 55.9 percent research, 4.1 percent checkout. The categories people delegate most readily are groceries (11.9 percent) and event tickets (10.7 percent): low-cost, low-ambiguity, low-fraud-surface purchases.
The pattern is exact. Delegation is inversely proportional to the density of product claims. Where a purchase decision depends on comparing specs, prices across sellers, and hundreds of reviews of unknown authenticity, shoppers keep the final click for themselves. Where verification is trivial, they hand it over.
That is the market telling you, in behavioral data, where the product trust problem lives: in precisely the categories with the highest order values and the most fake-review exposure. A $52 grocery basket does not need an evidence layer. A $240 electronics purchase drowning in incentivized five-star ratings does.
What Nobody Measured: The Reading List
For all its rigor, the FirstPageSage study contains a silence that defines the next phase of this market. It measures what shoppers and agents did. It does not, and cannot, measure what agents read: whose reviews, filtered how, weighted by what.
That gap matters because the review-summary stage is now a mainstream behavior: 39.2 percent of US shoppers use AI to distill reviews, and it sits directly upstream of every approved checkout in the funnel. Meanwhile the review pools being distilled remain polluted at industrial scale. Google blocked or removed more than 292 million policy-violating Maps reviews in 2025; Trustpilot removed 7.8 million fake reviews in the same year (CompaniesHistory compilation, October 2026). BrightLocal’s 2026 survey of 1,002 US adults found that 97 percent read reviews for local businesses, that 45 percent now use ChatGPT or other AI tools for local recommendations, and that the bar keeps rising: 31 percent will only use businesses rated 4.5 stars or higher, up from 17 percent a year earlier, with 47 percent refusing businesses with fewer than 20 reviews.
Put those two datasets side by side and the mechanism is plain. Consumers are tightening their quality thresholds and offloading the reading to machines, while the underlying review corpus remains a fraud-contested medium. Higher standards plus machine mediation plus unverified inputs equals a system where fabricated consensus can be laundered into clean-sounding summaries and delivered at the exact moment a shopper is predisposed to approve. A review-fraud economy once had to fool one human skimming stars. Now it only has to fool one ranking layer, and every downstream approval inherits the error.
The five-to-one approval preference shows shoppers sense this. What they lack is anything worth approving against. An approval with no independent evidence under it is a rubber stamp with a fingerprint.
The Counterweight: Evidence Before the Approval Click
Every structural trend in this dataset points the same way. Research volume is enormous and compounding. Checkout is gated on human approval. Partnerships and, soon, paid checkout modules will bias the consideration set. The one input no platform is selling, filtering or even measuring is whether the product deserves the recommendation.
That is the slot an independent evidence layer occupies. Not another store, not another agent: a trust layer consulted before the approval click. GoBuy’s contribution to that stack is concrete: fake reviews filtered out and authentic ones weighted up, a Smart Score from 0 to 100 built on review quality rather than review quantity, a deliberately short output of only the top 7 products per category instead of a thousand listing slots, a GoBuy Verified badge reserved for products holding a score of 80 or more across 90 days, and an MCP endpoint at gobuy.ai/api/mcp so any shopping agent, from a personal assistant to a merchant’s store agent, can consult independent trust data before it builds the cart. A Chrome extension carries the same trust panel onto Amazon product pages, where the raw review corpus lives.
The agent-shopping era’s first behavioral dataset settled one argument: humans are keeping the final click. The next argument is what the human sees in that final instant. If it is only the agent’s summary of a polluted corpus, the approval economy will industrialize bad purchases at scale. If it is a verified score over filtered evidence, the approval click becomes what shoppers already believe it is: a judgment.
The checkout war is over. The evidence war starts now. Start with the score: gobuy.ai, or wire the trust layer into your agent at gobuy.ai/agent-docs.