Two competing protocols, ACP from OpenAI and Stripe, and UCP from Google, went live in the past eight months and are quietly routing a fast-growing slice of high-intent shoppers away from traditional search and directly to checkout. Adobe Digital Insights clocked a 393% year-over-year jump in AI-driven visits to U.S. retail sites in Q1 2026. Those visitors converted 42% better than non-AI traffic. Most brands are invisible to them because their product data isn't structured for machines, only for humans.

The new plumbing In September 2025, OpenAI and Stripe launched the Agentic Commerce Protocol, or ACP. It lets users buy products inside a ChatGPT conversation, no redirect, no product page visit, no cart. The user types a request. The agent pulls options from merchant feeds, surfaces results, and processes payment via Stripe. The merchant stays the seller of record. The entire session happens inside the AI interface.

Google responded in January 2026 with the Universal Commerce Protocol, or UCP, backed by Walmart, Target, Shopify, and more than 20 additional retailers. UCP routes high-intent queries in Google Search's AI Mode and the Gemini app directly to merchant checkouts. Google also added dozens of new Merchant Center data attributes at launch specifically for conversational commerce: answers to common product questions, compatible accessories, substitutes, enhanced return and delivery terms.

Two protocols, two distribution channels. McKinsey projects this channel reaches between $3 trillion and $5 trillion globally by 2030. Morgan Stanley puts nearly half of all online shoppers using AI agents by 2030, at roughly 25% of their total spending. Those are long-range projections and the real question for operators is what's happening right now.

The Q1 2026 Adobe data is what's happening right now.

AI-driven retail site visits up 393% year-over-year. In March 2026, AI-referred shoppers converted 42% better than non-AI traffic, a new record. That number matters more when you see where it came from: in March 2025, AI traffic converted 38% worse than traditional channels. A swing from minus 38% to plus 42% in twelve months is not a trend line, it's a channel becoming viable in real time. The traffic is real, the intent is high, and the conversion potential is there. The gap is on the merchant side.

Why your catalog is invisible AI agents don't browse product pages the way humans do. They don't look at lifestyle photography, read narrative descriptions, or respond to urgency copy. They query structured product feeds, hit inventory APIs, check shipping estimators, and evaluate metadata completeness against the buyer's stated constraints.

When a customer tells ChatGPT "I need a waterproof hiking boot under $150, size 10, that ships by Thursday," the agent parses that into structured constraints: category, price ceiling, size availability, delivery window. It fires API calls against your catalog, your inventory system, and your shipping estimator. If any field is missing, stale, or inconsistently formatted, the agent either excludes you from the candidate set or fails mid-transaction.

Your storefront may exist. If it's not machine-readable in a transactional context, it's not involved in that purchase.

The research firm MetaRouter put a number on the gap: McKinsey data shows AI-generated product recommendations achieve 4.4 times the conversion rate of traditional search. But a study of 973 e-commerce sites with $20 billion in combined revenue found that when ChatGPT traffic actually arrives, it converts 86% worse than affiliate traffic. Affiliate is one of the highest-converting channels in e-commerce, so that gap is significant. That's not a demand problem. Seventy percent of consumers say they're at least somewhat comfortable with an AI agent making purchases on their behalf, according to research published by Commercetools in April 2026. The friction is almost entirely on the infrastructure side.

The specific fields agents are currently evaluating that most product listings don't have:

Real-time inventory at SKU level, ideally synced in under 15 minutes. Stale availability data produces failed transactions and trains agent platforms to trust you less over time.

Return policy in structured fields, not prose. "Free returns within 30 days" buried in a paragraph doesn't parse. A machine-readable field with policy type, window, cost, and conditions does.

Use-case context beyond feature lists. Current product descriptions are written for keyword matching. They list specs, dimensions, and materials. Agents construct context from descriptions to match user intent. "Waterproof ripstop nylon upper with Vibram outsole" is complete for a human who already knows what they're looking for. "Designed for all-day hikes in wet conditions, compatible with MicroSpikes for winter use, fits wide feet without toe box compression" gives the agent connective tissue to surface your product for conversational queries it otherwise can't evaluate.

Answers to common questions as structured data. Google's new Merchant Center attributes specifically ask for this. If you're not populating them, you're invisible on UCP surfaces.

Where the traditional playbook runs short Standard SEO is a content-first acquisition model. Blog posts, buying guides, category pages, editorial content. That model still works for human-driven search and it isn't going away.

Agents don't need editorial content. They need structured data. A detailed buying guide about your product category will not improve your ranking in a ChatGPT shopping result. A complete, accurate, real-time product feed will.

For most marketing teams, this is a resource allocation question that hasn't been asked yet. If your budget runs heavily toward content production and keyword-based SEO, you have a growing mismatch with where high-intent traffic is originating. The infrastructure budget, the one that historically felt like IT overhead, is now a direct revenue driver.

There's also a measurement problem compounding the issue. AI agent traffic doesn't appear cleanly in Google Analytics. Sessions initiated by agents look like direct traffic or get flagged as bots because of their behavioral signatures: rapid sequential queries, no mouse movement, no page scrolling. If your analytics stack is filtering this traffic, you can't measure it, can't attribute revenue to it, and can't make the case for investing in the infrastructure that captures it.

The window for first-mover advantage on this channel is probably 12 to 18 months. Brands that get agent-ready in 2026 will collect early data and early customer relationships from a channel that's scaling. Brands that wait will be rebuilding their product data infrastructure in 2027 while competitors are already optimizing against real performance data.

The four-layer agentic commerce audit You don't need to rebuild your stack to become agent-ready. Most mid-market brands are weak in one or two of the four layers below, not all four. Finding which ones is the actual work.

Layer 1: Product data quality

This is where most catalogs break down, and it's the most fixable.

Inventory accuracy first. Run a spot check on 20 SKUs: pull the availability your product feed reports, then check the actual warehouse count. If those numbers don't match within a 15-minute window, your inventory sync is producing stale data that agents will penalize over repeated failed queries.

Return policy structure second. Go to your top 10 SKUs and look at how return policy information is currently stored. If it's in a freeform description field or only on a standalone page, it's not machine-readable in a transactional context. You need a structured field: policy type (free/paid), return window in days, return cost, restocking fee if applicable. Most PIMs and product data tools support this as a custom attribute you can add without a development sprint.

Use-case enrichment third, and this is where the work is. Pull your top 20 SKUs by revenue and write three to five conversational use-case statements for each one. "Best for" framing, compatible products, compatible activities, who it's designed for, common use contexts. This is not marketing copy. It's context scaffolding for an agent that's trying to match your product to a buyer's stated intent. A skilled copywriter can do 20 SKUs in a day. A content manager using a structured enrichment template can do 50 in a week.

The tools worth using for this layer: Akeneo PIM for centralized data management and feed distribution, Plytix for product content distribution to agent channels, and Schema App for structured data markup. For brands under $10M revenue where dedicated PIM tooling isn't yet justified, a well-maintained Google Merchant Center feed with enriched attributes and custom fields is a viable starting point.

Layer 2: Protocol compatibility

ACP, the OpenAI and Stripe protocol, connects through your Stripe account. You map your catalog to ACP's schema and your products become purchasable inside ChatGPT without redirect. If you're already on Stripe and your catalog is reasonably structured, ACP compatibility is a configuration project, not an engineering project.

UCP, the Google protocol, runs through Google Merchant Center and Merchant Center Next. The January 2026 attribute additions are the key: if you're already running Google Shopping campaigns, you have the plumbing. What you probably lack is the enriched attribute layer described in Layer 1. Add those attributes to your Merchant Center feed and you're UCP-compatible on Google surfaces.

Most brands will need to support both protocols. ACP captures conversational, exploratory shopping. UCP captures high-intent search. The underlying data requirement is the same: clean, complete, structured product information. Build the data layer once and feed it to both.

Layer 3: Checkout API compatibility

Agents complete purchases via API calls, not browser sessions. Your checkout needs to be callable as a service. Shopify merchants have a clear structural advantage: Shopify's commerce APIs are already compatible with both ACP and UCP, and Shopify is a named partner in Google's UCP coalition.

For brands on custom stacks or legacy platforms, the question is whether your checkout can accept machine-initiated orders without human navigation. If it can't, you have two paths. A lightweight API wrapper in front of your existing checkout is a 30 to 90 day project for a competent developer and gets you agent-compatible without a full migration. A headless architecture migration is more thorough but takes 6 to 12 months. Get the wrapper working first.

Layer 4: Measurement

This is the most neglected layer, and the one that makes CFO approval for the other three nearly impossible to get.

AI agent traffic doesn't appear cleanly in standard analytics. You need server-side event tracking that captures agent authentication signals. Emerging authentication frameworks are pushing cryptographic verification via Signature-Input and Signature-Agent headers, which lets you distinguish legitimate agent traffic from malicious bots and route the data correctly. MetaRouter has written specifically about this infrastructure requirement, and it's becoming a standard part of serious agentic commerce implementations.

As a practical starting point before that infrastructure is in place: create a segment in your analytics platform for sessions with agent-like behavioral signatures (high query velocity, no scroll depth, no mouse movement, unusual sequential page behavior). It won't be precise, but it gives you a baseline to improve from and a number to put in front of leadership when the investment conversation comes up.

The AI citations rate is the metric you're ultimately building toward. It measures whether a shopping assistant retrieved, referenced, or recommended your inventory during an agentic session. If the agent didn't surface your product, your brand wasn't part of that transaction. That metric doesn't exist in most analytics dashboards yet. Building toward it now puts you ahead of the measurement curve when the channel reaches the scale where everyone starts paying attention.

What's actually working: How Papa John's deployed agentic ordering across three surfaces in under 90 days Papa John's is not a DTC brand, but the deployment structure is directly applicable to any operator thinking about what agent-readiness looks like in practice.

In January 2026, Google announced that Papa John's was among the first retailers live on Gemini Enterprise for Customer Experience, Google's agentic commerce platform. The deployment covers natural language ordering across mobile, kiosks, and in-car systems, with real-time menu synchronization and AI-driven upsell logic built into the conversation flow.

The operational detail that matters here: Papa John's didn't rebuild its ordering stack. Google's Food Ordering agent sits on top of existing menu and inventory data, translating natural language requests into structured order inputs that the existing system already understands. The lift was on the data side, specifically clean, real-time menu availability and structured item data, not on the checkout or fulfillment side.

The specific features announced include voice and text ordering across mobile apps, kiosks, and in-car systems, an "Intelligent Deal Wizard" that automatically applies the best available promotions without requiring customers to hunt for promo codes, and a no-tap reordering flow for returning loyalty customers. The system is designed to handle complex multi-person orders with natural language and real-time modifications, a use case that traditionally required human intervention.

Why it's relevant here: the deployment was possible because Papa John's already maintained clean, consistently structured menu data across digital channels. That existing data quality, built for a different reason, turned out to be the prerequisite for agentic compatibility. It's the same pattern that will play out in product commerce: brands that already maintain rigorous, structured product data for marketplaces and shopping feeds will be faster and cheaper to make agent-ready than brands starting from scratch.

How to apply this at $1M to $20M:

The prerequisite audit is the same regardless of brand size. Pull your product data from whatever system maintains it today, whether that's your Shopify backend, a spreadsheet, or a PIM, and run it against the four fields agents currently require: real-time inventory, structured return policy, use-case context, and answers to common questions. Document which fields exist, which are incomplete, and which are missing entirely.

Start with your top 20 SKUs by revenue. Enrich those first. Don't try to boil the ocean across a full catalog before you have a proof of concept. Get 20 SKUs fully agent-ready, submit them to ACP via Stripe and to UCP via Merchant Center, and run the baseline measurement for 60 days. You'll have data before you commit to the full catalog enrichment project.

The brands that move on this in Q2 and Q3 of 2026 will have 12 months of learning before this channel reaches mass-market attention. That learning window is the actual competitive asset.