Agentic Commerce: the 2026 reversal

Through 2026 the demo inverted. The industry stopped pretending the assistant would finish checkout inside the chat, and admitted the job is discovery. Checkout stays with the retailer’s own channel.

On 24 March 2026 OpenAI said the first Instant Checkout “did not offer the level of flexibility that we aspire to provide,” and that merchants should use their own checkout while OpenAI concentrates on product discovery (OpenAI). Walmart’s Daniel Danker called in-chat checkout “a very temporary moment in time” (CNBC).

The surface that actually printed money is Amazon. On 13 May 2026 Rufus and Alexa+ became Alexa for Shopping. Rufus served 300 million customers in 2025 and drove about $12 billion in incremental annualized sales; Q1 2026 MAU was +115% year on year, engagement +400% (About Amazon, Modern Retail).

EMARKETER cut its 2026 forecast for AI-platform-driven US retail e-commerce from $20.57B (~1.5%) to $19.74B, and reported that only 10% of AI-assisted shoppers usually finish on the AI platform; 69% bounce to the retailer’s own site or app. Retailer-built agents (Rufus/Alexa, Walmart Sparky) already take 54.1% of AI-driven sales in 2026 and stay ahead through 2030 (EMARKETER).

For a marketplace-seller PM the sentence that matters: this is not “which chat checkout API to join.” It is whether the feed, structured data, and PDP are clean enough for every assistant to cite. That theme repeats on every surface and every GEO shopping tool. The living catalog is the Agentic Commerce map. GEO vendors that crossed into shopping also sit on the GEO map.

Protocols converged. Do not pick a side.

Between September 2025 and mid-2026 the rails landed in a heap: ACP (Stripe + OpenAI, Apache 2.0), AP2 (Google, donated to FIDO in April 2026), UCP (Google + retailers, January 2026), Visa Intelligent Commerce Connect (protocol-agnostic: TAP/MPP/ACP/UCP), Mastercard Agent Pay / AP4M, and Web Bot Auth (Cloudflare → IETF, RFC 9421), which Visa, Mastercard, and Amex use for agent identity.

Visa’s Connect is explicitly multi-protocol. Stripe says UCP needs no extra work for Suite customers. PayPal Store Sync covers Copilot and Perplexity. The PM move is a protocol-agnostic payments or distribution partner (Stripe, PayPal, Adyen, Visa Connect), not a religious choice of ACP versus UCP.

Four surfaces

ChatGPT Shopping. Monetization moved to ads: US tests in February 2026, CPC + Ads Manager + Conversions API in May, then UK, Mexico, Brazil, Japan, Korea, and 31 European markets. Instant Checkout is still a handful of merchants — Shopify’s Harley Finkelstein said about a dozen Shopify shops were live in March 2026 (Forbes). Ranking factors OpenAI lists: availability, price, seller type, quality, primary seller, Instant Checkout. The feed spec is Google Merchant Center-compatible; ad-eligible SKUs need is_ads_eligible: true. The “2% / 4% take rate” numbers remain unconfirmed rumor. Peec’s own research says ChatGPT Shopping is a live pull of roughly the top 40 Google Shopping organic results, then a re-rank — which is why a clean Google feed is the highest-leverage single action.

Alexa for Shopping. Conversational compare, one-year price history, target-price holds, cross-site Buy for Me. The only confirmed seller lever on recommendations is advertising. Sponsored Products/Brands Prompts went GA on 25 March 2026; existing campaigns auto-enroll with no opt-out. Organic guidance in Seller Central is: if Rufus is wrong, fix the PDP and file a ticket with ASIN and screenshots. Thumbs up/down. No visibility dashboard. That vacuum is the opening for Azoma, GetMint, and eCommerce Insights.

Shopify Agentic Storefronts. Default-on distribution for SMBs. Catalog standardizes the feed; Shopify claims 2× conversion in AI chats; ChatGPT, Google AI Mode, Gemini, Copilot, Meta. Developer docs already describe paid Global Catalog placements with commission on attributed purchases — the second confirmed agentic ad slot, after Amazon Prompts.

The rest. Microsoft Copilot Checkout (US, 8 January 2026, Stripe/PayPal/Shopify, auto-enroll). Google AI Mode / Gemini (UCP, Google Pay, retailer stays seller of record, Universal Cart, Merchant Center share-of-voice on AI surfaces). Perplexity Shopping (Buy with Pro, free Merchant Program; March 2026 court block on Comet using users’ Amazon accounts). Walmart Sparky (Walmart’s in-ChatGPT app; UCP co-developer; Seller Center still only has a listing-rewrite tool unrelated to Sparky visibility). Meta testing Buy Now on IG/FB.

GEO tools: monitor versus write-back

Profound (13 Nov 2025) and Peec (17 Jun 2026) cover ChatGPT Shopping only. GetMint is the only one of that original trio with Alexa for Shopping in the $80 starter tier. The vendors actually built around retailer agents and write-back into Amazon or Shopify are newer: Azoma (widest surface list found, Seller/Vendor Central write-back, $4M pre-A with eBay Ventures × Techstars), Wildcard (YC W25, Shopify publish, Rufus in beta, “not for SaaS”), eCommerce Insights (citation ledger, Agent-Readability Score, PDP diffs to Shopify). Syndigo and Salsify are incumbent PXM platforms adding ACP feeds and AEO copy — they will eat “agentic commerce” as a connector, not as a startup category.

Product split worth remembering: dashboards (Peec, Profound, Semrush) versus write-back (Wildcard, Azoma, eCommerce Insights). The earlier GEO observation — the category is racing from reports toward execution — is now visible in shopping.

Market size: use the number that can be revised down

Same category, two to three orders of magnitude. Plan against EMARKETER-style figures. The measured number, not a model, is Adobe: July 2026 AI referral traffic to US retail sites +62% YoY; those visits converted 60% better and 53% more revenue per visit — while average LLM-visibility scores fell from 75% to 61% (Digital Commerce 360). Conversion is up. Machine-readability is not. That gap is the product.

What to do

1. Feed quality first. Google Merchant Center done well hits Google AI Mode and, if Peec is right, ChatGPT Shopping together.
2. If you build or buy “AI shopping visibility” for marketplace sellers, Amazon is less crowded than ChatGPT. Benchmark GetMint, Azoma, eCommerce Insights.
3. Distribute through aggregators (Shopify Agentic Storefronts, Stripe Suite, PayPal Store Sync, Visa Connect).
4. Agentic ad slots already auto-enroll. Fold Amazon Prompts and Shopify Global Catalog into existing ACOS/ROAS, do not stand up a separate religion.
5. Legal: Amazon Buy for Me scraped third-party retailers into unauthorized listings (brands had to email opt-out). A court has already blocked Perplexity’s agent from Amazon accounts. If you transact on someone’s behalf, those two go in the risk list.
6. Ignore trillion-dollar slides. Treat this as a high-ROI incremental channel, not the whole roadmap.

Cards, protocols, and vendors: Agentic Commerce map.