Last updated: September 30, 2026
AI-referred traffic to U.S. retail sites grew 393% year over year in Q1 2026, with March alone up 269%, according to Adobe Analytics data published in April 2026 (Source: TechCrunch).
GrowthCentr builds growth systems for B2B and commerce teams whose buyers now begin their journey inside an AI interface rather than a page of blue links.
That traffic shift matters far less than the mechanical change underneath it: AI shopping agents do not browse your site the way a human does, they read your structured product data and make inclusion decisions from it before a shopper ever sees a recommendation.
This article breaks down exactly what changed in product feed specifications, freshness requirements, attribute schemas, and measurement during 2026, and what merchants need to fix to stay visible.
Key Takeaways
- ACP and UCP turned product feeds into ranking infrastructure, not just ad plumbing.
- Feed freshness moved from daily uploads to a 15-minute maximum lag window.
- Google added conversational attributes built specifically for AI Mode discovery.
- Missing GTINs, thin attributes, and stale stock now cause silent exclusion from recommendations.
- Most mid-market retailers must now maintain two protocol feeds simultaneously.

What Changed in the September 2026 Update
What Changed on September 30, 2026
This pass adds Adobe's July 2026 cart and category-level LLM-visibility readings so feed owners can prioritize which catalog gaps to close first.
- AI-referred shoppers add items to cart 28% more often than non-AI traffic (Source: Adobe Analytics via Digital Commerce 360).
- On Adobe's expanded July cohort, LLM homepage visibility scored 76% for apparel, 70% for electronics, and 59% for grocery, leaving grocery with the largest machine-readability gap in the set.
GrowthCentr's take: a 28% cart lift is worthless if grocery and general-merchandise pages stay opaque to agents. Treat category visibility scores as a backlog order: fix thin attributes where the gap is widest first. Pair this guide with GrowthCentr's ChatGPT Instant Checkout fees and requirements teardown and agentic payments statistics when deciding whether discovery or checkout ownership is the bottleneck.
Two things moved since this guide first ran: Google retired the API many feed pipelines were built on, and checkout inside Google's AI surfaces went from pilot to default. Here is what to act on.
- Google's Content API for Shopping reached sunset on August 18, 2026. Since September 1, 2026, requests without an approved extension intermittently fail with HTTP 410 Gone, and full decommissioning is planned for early 2027. Custom integrations must move to the Merchant API; merchants on platforms such as Shopify have the migration handled by their platform (Source: Google for Developers).
- Google now describes the Shopping Graph as more than 60 billion product listings, with people shopping across Google more than a billion times a day (Source: Google).
- Merchant Center's AI performance insights report your share of voice in AI shopping experiences, based on your AI impressions compared with competitors (Source: Google Merchant Center Help).
- Google began enabling native checkout in AI Mode and Gemini by default for eligible Shopify merchants in September 2026, with an opt-out under Sales channels > Agentic (Source: Search Engine Roundtable). For how in-chat checkout economics played out the first time, see GrowthCentr's breakdown of ChatGPT Instant Checkout fees and requirements.
- Adobe's July 2026 data shows AI-referral traffic to US retail sites up 62% year over year, converting 60% better than non-AI traffic, while 39% of retailer homepages in its expanded cohort were not machine-readable (Source: Digital Commerce 360). Channel-level context sits in GrowthCentr's marketing ROI statistics.
GrowthCentr's take: the Merchant API migration is the most urgent item on this list for any team running a custom feed pipeline. A feed that silently starts returning 410 errors produces exactly the stale price and stock data that agents punish.
The Agentic Commerce Market Entering H2 2026
The numbers behind agentic shopping stopped being speculative in 2026. Adobe recorded a 693% year-over-year surge in AI-driven retail traffic across the 2025 holiday season, and that momentum carried forward into a 393% year-over-year Q1 2026 increase (Source: TechCrunch).
Salesforce, drawing on 1.5 billion shoppers, reported that AI and agents influenced $67 billion in global sales during Cyber Week 2025, touching 20% of all global orders (Source: Salesforce).
The conversion picture inverted just as sharply. In March 2025, AI-referred traffic converted 38% worse than non-AI traffic. By March 2026, Adobe measured AI-referred traffic converting 42% better, an eighty-point swing in twelve months (Source: TechCrunch).
Shopify's Q1 2026 data reinforced the pattern: AI chatbot referral sessions to its storefronts grew more than 8x year over year, AI-referred orders grew nearly 13x, and those orders carried 14% higher average order values than organic search (Source: Shopify).
Forecasts diverge because definitions diverge. EMARKETER, counting only checkout that happens inside an AI platform, projected about $20.57 billion in U.S. retail ecommerce for 2026, roughly 1.5% of total U.S. ecommerce (Source: EMARKETER), then cut that 2026 forecast to $19.74 billion in July 2026 as adoption ran slower than expected (Source: EMARKETER).
McKinsey's broader orchestrated-retail definition reaches $3 trillion to $5 trillion globally by 2030 (Source: Digital Commerce 360). Gartner expects 90% of B2B buying to run through AI agents by 2028, pushing more than $15 trillion through agent exchanges (Source: Gartner).
For a fuller breakdown of the underlying transaction data, see GrowthCentr's agentic payments statistics for 2026.
Why Feeds Became the Ranking Layer
Traditional product feeds were built for comparison shopping engines that rendered a grid of options for a human to evaluate.
A human forgives a missing material attribute or a slightly stale stock count. An agent does not, because the agent is filtering programmatically before anything reaches the shopper.
Google's Shopping Graph now holds more than 60 billion product listings (Source: Google), with over 2 billion updated every hour (Source: PPC Land). AI Mode surpassed one billion monthly active users globally, and the average AI Mode search is triple the length of a traditional Search query (Source: Google).
Longer queries carry more constraints, and every constraint maps to a structured attribute. A shopper asking for a waterproof hiking boot in wide sizes, under a price cap, with free returns is running four filters against your feed.
If your catalog lacks a width attribute, you are invisible to that query regardless of your domain authority or ad budget.
This is the practical difference between classic SEO and AI-surface visibility, a distinction GrowthCentr covers in more depth in its guide to AI search visibility.

The 2026 Timeline: What Actually Changed
- September 2025. OpenAI and Stripe released the Agentic Commerce Protocol (ACP) under Apache 2.0, replacing bespoke merchant integrations with a shared standard for validating credentials, confirming orders, and processing payments.
- January 2026. Google unveiled the Universal Commerce Protocol (UCP) at NRF 2026, co-developed with Shopify, Etsy, Wayfair, and Target, covering discovery through post-purchase. On January 11, Google announced dozens of new Merchant Center data attributes designed for conversational interfaces including AI Mode, Gemini, and Business Agent.
- January 30, 2026. OpenAI published ACP product feed spec version 2026-01-30, adding product extensions, discount fields, and payment handler metadata. The spec versions roughly once per quarter.
- March 2026. UCP added Cart and Catalog capabilities, letting agents add multiple items from a single store and retrieve real-time variants, inventory, and pricing. Google announced simplified UCP onboarding through Merchant Center.
- April 2026. Microsoft adopted UCP in its merchant tooling and wired Shopify Catalog into Copilot. Adobe Summit shipped an MCP server for Adobe Commerce. Google extended AI Max to standard Shopping campaigns on April 30, explicitly reading feed attributes such as material, fit, and durability to generate ad copy.
- May 2026. Universal Cart began rolling out in the U.S. on May 19, requiring tighter pricing accuracy and faster inventory updates across Search, Gemini, YouTube, and Gmail. Google Marketing Live on May 20 introduced conversational attributes as a formal Merchant Center data type and expanded UCP to hotels, food delivery, and three new countries.
- August 2026. Google's Content API for Shopping reached sunset on August 18, with intermittent HTTP 410 failures from September 1 for integrations without an extension (Source: Google for Developers).
- September 2026. Google began enabling native checkout in AI Mode and Gemini by default for eligible Shopify merchants (Source: Search Engine Roundtable).
The Protocols You Now Feed
Merchants are no longer choosing one standard. They are maintaining parallel infrastructure for overlapping ecosystems.
| Protocol | Owner | Scope | Feed entry point |
|---|---|---|---|
| ACP | OpenAI and Stripe | Discovery and checkout in ChatGPT | Product feed spec, TSV/CSV/XML/JSON |
| UCP | Google and coalition | Full journey: discovery, cart, checkout, post-purchase | Merchant Center plus /.well-known/ucp capability profile |
| MCP | Open standard | Tool and catalog access for agents | Server exposing catalog endpoints |
| AP2 | Google-led | Payment authorization and trust | Mandate and credential layer |
UCP is protocol-agnostic and supports REST, MCP, and A2A, and its endorsement list has grown to include Mastercard, Visa, Walmart, Target, Best Buy, Amazon, Meta, Microsoft, Salesforce, and Stripe.
If you need the underlying mechanics, GrowthCentr has primers on the Model Context Protocol, the Agent Payments Protocol, and the Agent2Agent protocol.
Inside the ACP Product Feed Specification
The ACP feed organises product data into 15 logical groups spanning roughly 67 fields. Approximately 15 are required, around 12 are conditional or strongly recommended, and about 40 are optional.
The optional fields are where result quality is won, because they are what conversational filters actually match against.
The public ACP product feed specification accepts TSV, CSV, XML, or JSON, so merchants can reuse whichever export process they already run (Source: Agentic Commerce Protocol).
Delivery splits into two modes:
- File upload for daily full snapshots
- API for incremental updates and promotions.
Conditional fields catch most first-time submissions. Seller privacy policy and seller terms of service are required when checkout is enabled. Availability date is required for pre-orders.
Dimension and weight units are required whenever dimensions or weights are supplied. Pickup SLA is required when a pickup method is declared. Base measure is required when unit pricing measure is supplied.
Geo fields handle multi-region merchants, with target countries and store country required as ISO 3166-1 alpha-2 codes.
Merchants evaluating in-chat purchase flows should also review current ChatGPT Instant Checkout fees and requirements, since transaction economics differ meaningfully from standard paid channels.

Google's Eight Conversational Attributes
Several attributes added during 2026 carry an explicit note in Google's documentation stating they are primarily intended for conversational experiences such as AI Mode. These are not retrofitted Shopping fields.
| Attribute | Function | Key limits |
|---|---|---|
| Product highlight | Short selling-benefit fragments | 1 to 150 characters, 4 to 6 recommended, 100 max |
| Product detail | Technical specs grouped into sections | Section name, attribute name, attribute value |
| Variant option | Variant dimensions outside the standard six | Name and value up to 250 characters, 30 max |
| Item group title | Shared generic title across variants | 150 characters, must exclude variant specifics |
| Related product | Six relationship types including substitute and accessory | 30 entries max, separate row each |
| Question and answer | Pre-written FAQ content in the feed | Character and pair limits per Google's attribute spec |
| Document link | Public PDF manuals and specs | HTTPS, 50 MB max, five per product |
| Popularity rank | Sales rank relative to your own inventory | 0 to 100, one decimal, no percent sign |
Google lists question and answer, document link, related product, item group title, variant option, and popularity rank as its conversational attributes, submitted through a supplemental data source or your main feed (Source: Google Merchant Center Help). Formatting breaks are the most common failure. Commas inside product highlights or Q&A values split fields when submitted via CSV, so tab-separated files or escaped values are the safer path.
Variant option must be submitted alongside item group ID and item group title, and all variants sharing an item group ID must use the same name sub-attributes.
The New Freshness Standard
The legacy model of a nightly CSV upload is now a liability. Google's Shopping Graph refreshes billions of listings hourly, and AI surfaces prefer live data. The ChatGPT feed accepts updates every 15 minutes (Source: Agentic Commerce Protocol).
The operational standard that emerged during 2026 is a maximum 15-minute lag on price and inventory state, with real-time API sync for high-velocity SKUs. The reason is reputational rather than cosmetic.
When an agent recommends a product that turns out to be unavailable at checkout, the transaction fails and the assistant records that failure.
Platforms track merchant reliability, and merchants with repeated out-of-stock mismatches get down-weighted in subsequent recommendation rounds, even after restocking.
Practical implementation paths include Merchant API push updates (the Content API for Shopping reached sunset on August 18, 2026), webhook-based feeds through a platform partner, and scheduled fetches at 15-minute intervals as a floor rather than a target (Source: Google for Developers).
Identifiers, Matching, and Silent Exclusion
GTIN remains one of the strongest matching signals in the ecosystem. A wrong, missing, or invented GTIN drops a listing out of the competitive product clusters where AI surfaces aggregate offers from multiple retailers.
For private-label or handmade products, the correct move is an explicit identifier_exists: false flag rather than a blank field.
Mirakl notes that 42% of customers will abandon a purchase due to insufficient product information, and that poor image quality has a similar impact (Source: Mirakl).
Agents inherit those same evaluation criteria and apply them at machine scale, without the benefit of a shopper who might squint at a product page and guess.
The failure mode is silence. There is no error message, no rejected ad, and no manual action notice. The product simply is not considered.

What This Costs: Team, Tooling, and Compensation
Feed work moved from a quarterly maintenance task to a standing function, and compensation data reflects that.
| Role | 2026 U.S. base salary benchmark | Source and date |
|---|---|---|
| VP or Director, AI and agentic commerce | $170,000 to $260,000 base | eCommerce Placement, 2026 guide |
| Director of eCommerce | $130,000 to $185,000 base | eCommerce Placement, 2026 guide |
| Product Manager, eCommerce | $110,000 to $160,000 base | eCommerce Placement, 2026 guide |
| Retail Media Manager | $90,000 to $130,000 base | eCommerce Placement, 2026 guide |
AI and agentic commerce leadership is currently the highest-paid emerging function in ecommerce, AI fluency carries a reported 10% to 15% salary premium, and Head of Amazon compensation rose more than 15% year over year (Source: eCommerce Placement). For most mid-market teams the realistic build is one feed owner, a data engineering allocation for real-time sync, and a feed management platform, rather than a full new department.
Measurement Breaks Before Attribution Does
AI agents do not execute client-side JavaScript. They make direct API calls to merchant systems. Any measurement stack resting on pixels develops blind spots that widen as agent traffic scales, and AI referrals are frequently misattributed inside standard analytics configurations.
Three fixes matter most. Move to server-side conversion tracking so agent-originated orders register. Implement cross-surface identity resolution so agent purchases attach to customer profiles for loyalty and personalization. Add a post-purchase email opt-in on protocol-mediated orders to preserve first-party data, since session-based retargeting and cart abandonment flows will not capture these buyers.
Security posture deserves equal attention as authorization moves from humans to agents, which GrowthCentr covers in its breakdown of agentic payments security risks.

A 30-Day Feed Readiness Sequence
- Days 1 to 5. Audit identifier coverage. Fix GTIN accuracy, set
identifier_existscorrectly, and validate item group IDs across every variant family. - Days 6 to 12. Fill conversational attributes. Add product highlights, structured product details, variant options, and Q&A pairs. Switch feed files to TSV or escape delimiters properly.
- Days 13 to 18. Upgrade freshness. Move price and inventory to Merchant API push or a 15-minute fetch cadence, prioritising high-velocity SKUs.
- Days 19 to 24. Add protocol layers. Publish a UCP capability profile, apply
native_commerceto eligible listings, and format an ACP-compliant feed for OpenAI ingestion. - Days 25 to 30. Rebuild measurement. Deploy server-side tracking, tag agent-referred sessions, and set a baseline for agent-sourced revenue.
Risks and Constraints Worth Planning For
The ecosystem is not settled. Amazon added OpenAI and other AI crawlers to its robots.txt restrictions in August 2025, then joined the UCP Tech Council in April 2026, so access to its catalog is increasingly on its own terms (Source: PPC Land). Walled gardens are getting taller even as open protocols expand.
Consumer adoption also lags the infrastructure: EMARKETER's July 2026 forecast cut reflected shoppers who prefer AI for research over purchasing (Source: EMARKETER). Gartner separately predicts that over 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls (Source: Gartner). Feed investment holds value regardless of which protocols win, because clean structured data is the input to every version of this future. Teams evaluating settlement rails should also review GrowthCentr's comparison of stablecoins versus card rails for AI agent payments.

Conclusion
GrowthCentr helps commerce and B2B teams turn structured data into durable visibility across AI surfaces where discovery now begins.
2026 moved product feeds from an advertising utility to the primary ranking layer of agentic commerce.
ACP formalised a 67-field spec with quarterly versioning, UCP extended Merchant Center into cart and checkout, Google shipped attributes built specifically for conversational queries, and freshness expectations compressed from 24 hours to 15 minutes.
The merchants gaining ground are not the ones spending more on bids. They are the ones whose catalogs are complete, accurate, machine-readable, and current enough for an agent to trust.
Read Next
- How to Sell on ChatGPT in 2026: What Changed After OpenAI Dropped Instant Checkout
- Agentic Payments Statistics 2026: Market Size, Adoption, and Growth Data
- Stablecoins vs Card Rails for AI Agent Payments: Which Settlement Layer to Choose
FAQs
1. What is product feed optimization for AI shopping agents in 2026?
Product feed optimization for AI shopping agents in 2026 is the practice of structuring, enriching, and continuously refreshing catalog data so protocol-driven agents can parse, trust, and recommend your products across surfaces like ChatGPT, Google AI Mode, Gemini, and Copilot.
2. What changed in product feed requirements in 2026?
What changed in product feed requirements in 2026 is the shift from static daily feeds to protocol-based specifications, with ACP publishing a versioned 67-field spec, UCP extending Merchant Center into cart and checkout, and Google adding conversational attributes built for AI Mode.
3. How often should a product feed update for AI shopping agents?
A product feed should update for AI shopping agents within a maximum 15-minute lag window for price and inventory, with real-time API sync for high-velocity SKUs, because stale availability data causes failed transactions and lower reliability scoring.
4. Which product feed attributes matter most for AI Mode visibility?
The product feed attributes that matter most for AI Mode visibility are product highlight, product detail, variant option, item group title, related product, question and answer, document link, and popularity rank, alongside accurate GTINs and complete images.
5. Do merchants need both ACP and UCP feeds in 2026?
Merchants need both ACP and UCP feeds in 2026 if they sell across ChatGPT and Google surfaces, since ACP powers ChatGPT shopping experiences while UCP powers Google AI Mode and Gemini, and most mid-market to enterprise retailers appear in both.
Disclaimer: This content is provided for informational purposes only and does not constitute legal, financial, or compliance advice. Protocol versions, governance arrangements, and partner counts cited here reflect publicly announced milestones as of August 2026 and are moving quickly. Adoption figures come from vendor and foundation announcements with differing methodologies and should be treated as directional signals rather than guaranteed outcomes.
