7 Spec-Verified Steps to Get Your Products Cited by AI Shopping Agents

TL;DR

ChatGPT's Agentic Commerce Protocol (ACP) feed spec defines 66 fields with 16 required. Google Merchant Center's product data specification defines 68 attributes, also with 16 required. Schema.org's Product type documents 70 properties and Offer documents 51. Those two 16-required counts look identical until you compare which 16: ACP unconditionally requires checkout-eligibility flags and geo-targeting fields Google's spec never mentions, while Google unconditionally requires apparel attributes like age_group, gender, and color that ACP doesn't touch. Adobe Analytics data, stitched across four separate company releases, shows AI-referred retail traffic went from converting 91% worse than other traffic in July 2024 to 54% better in May 2026. This page pulls every number directly from OpenAI, Google, Shopify, schema.org, and Adobe's own primary documentation, audits nine competing guides against those sources, and gives merchants a seven-item checklist where every item cites the actual spec, not someone's blog post about the spec.

Last verified: August 1, 2026

I spent an evening doing something no competing guide bothered to do: I fetched the actual OpenAI feed spec, the actual Google Merchant Center spec, and the actual schema.org Product page, and counted the fields myself. Not summarized them. Counted them, one by one, in three separate browser tabs, until I had numbers I could defend.

This is not another "GEO checklist" assembled from other GEO checklists. Most published guides on "optimize for ChatGPT shopping" name a spec without linking to it, or link to it without reading past the intro paragraph. This page covers the feed layer. For the payment and protocol layer underneath it, I broke down ACP, AP2, x402, MPP, and UCP side by side on the newsletter.

What's Inside

• How many fields does OpenAI's ACP feed spec require, and how does that compare to Google Merchant Center?
• If ACP and Merchant Center both require 16 fields, why can't one feed satisfy both?
• What does the schema.org Product and Offer vocabulary cover that platform-specific feeds don't?
• Has AI-referred shopping traffic gotten better at converting, and since when?
• What's the shortest checklist that's still fully sourced to official documentation?
• Why won't any major AI shopping platform publish its actual merchant fee?
• What do the four Q&As at the bottom of this page cover for voice assistants?

The Spec Comparison Missing From All Nine Guides I Audited

I dare you to find one published GEO guide that puts OpenAI's spec and Google's spec side by side with real numbers. I looked through nine while researching this page and came up empty. Here's the table, pulled directly from primary sources:

Schema.org doesn't mark anything as required. The consuming platform decides what it needs, which is why the ACP and Merchant Center rows matter more for anyone shipping a feed this quarter.

OpenAI's ACP feed spec and Google's Merchant Center spec both land at 16 required fields, and that's where the resemblance ends. ACP's 16 include two flags most merchants have never declared before, is_eligible_search and is_eligible_checkout, plus geo-targeting fields (target_countries, store_country) Google's spec treats as optional or skips entirely (OpenAI's ACP feed spec). Google's 16 pull in category-conditional apparel attributes instead: age_group, gender, and color become mandatory the moment you're selling apparel into Brazil, France, Germany, Japan, the UK, or the US (Google's product data specification).

Same headline number. Completely different homework.

A feed built to satisfy Google Merchant Center won't automatically clear ACP's checkout-eligibility bar, and vice versa for Google's apparel rules. I built this comparison by reading both documents in full myself, since neither company has a reason to describe the other's spec. Selling apparel across US and EU storefronts feeding both ChatGPT and Google Shopping from the same data? Budget separate QA passes. One clean feed doesn't mean two clean integrations.

Nine Guides Audited, One Consolidated Table Missing

Nine pages currently rank or show up for queries like "optimize products for ChatGPT shopping" and "GEO for ecommerce." I read all nine in full. None of them assembles the table above.

Alhena.ai's ChatGPT Shopping Optimization guide comes closest, naming real ACP field identifiers (title, image_url, price, q_and_a, is_ads_eligible) and quoting OpenAI directly on ranking being "organic and unsponsored" (Alhena.ai). Its "96.1% of US LLM referral traffic" figure is attributed only to its own cohort research, and it stops short of the full field count or a Merchant Center comparison.

Search Engine Land's October 2025 piece names the Agentic Commerce Protocol with no working links in the fetched content (Search Engine Land). It still describes Custom_variant1-3_category/_option as current, fields marked deprecated in the version I fetched, superseded by variant_dict.

Search Engine Journal's ACP/UCP explainer gets the September 29, 2025 launch date right and credits Shopify with enabling over a million merchants on day one (Search Engine Journal). It links to no spec directly, and its note that WooCommerce and Wix merchants need a Stripe waitlist is dated since Shopify's June 2026 Catalog release.

LumenGEO's e-commerce GEO strategy offers a genuinely distinct three-layer framework: on-site schema, review amplification, original-data content (LumenGEO). It cites zero official platform specs, and its JSON-LD example is generic best practice, not sourced to schema.org.

AthenaHQ's ranking guide carries citation markers whose URLs don't resolve, with no OpenAI spec links or field counts anywhere (AthenaHQ). Its "42% of consumers use generative AI for shopping ideas" claim has no traceable source.

Hello Retail's GEO guide is the most statistically dense of the nine, naming real sources for nearly every number: Adobe, Gartner, SparkToro, BrightEdge, Pew Research (Hello Retail). Its Q1 2026 393% traffic-growth figure checks out against Adobe's own data, but it's a behavioral-research roundup, not a spec audit.

WRKNG Digital's feed optimization piece conflates the two specs it audits (WRKNG Digital). It states GTIN and MPN are required for ChatGPT Shopping eligibility. In the actual ACP spec, both are optional. Only brand is required.

Precis's three-level feed playbook makes the most rigorous empirical claim of the nine, citing Peec AI research that "100% of products appearing in ChatGPT Shopping could be explained by the top 40 organic results in Google Shopping" (Precis). This is contested. It contradicts OpenAI's own documentation, which describes ChatGPT ingesting merchant-submitted ACP feeds directly [UNCERTAIN, unresolved between a third-party research firm and the platform's own docs].

BigCommerce's Ecommerce GEO PDF names five schema.org types without field counts or a schema.org link (BigCommerce). Its cited Gartner prediction of a 25% search-volume drop is named but not linked.

Every one of these nine either speculates about requirements, cites behavioral research without touching the specs, or names a spec without quantifying it.

The 22-Month Conversion Swing, Stitched Into One Line

I measured this one across four separate Adobe reports because a single snapshot doesn't tell the real story.

AI-referred retail traffic converted 91% worse than non-AI traffic in July 2024. By May 2026, it converted 54% better. That's a swing of roughly 145 percentage points in 22 months, computed from Adobe Analytics' own successive releases, not from a single vendor's cherry-picked quarter. The swing makes sense once you watch what an agent actually does at checkout. I ran the same purchase as a human and as an agent, step by step, and the difference explains most of that conversion gap. The data points come from Adobe's Q2 2025 report, Adobe's May 2025 report, Adobe's January 2026 holiday recap, and Adobe's Q3 2026 PDF.

Here's the full trendline:

That reversal happened on a dataset covering more than 1 trillion visits to US retail sites and over 100 million tracked SKUs since October 2024 (Adobe's Q3 2026 report). Hello Retail's guide cites the Q1 2026 393% traffic-growth figure accurately, matching Adobe's own numbers, but citing one quarter isn't the same as stitching the full arc. AI traffic in May 2026 was also up 138% year over year and revenue per visit was up 53% against non-AI traffic that same month, itself a full reversal from twelve months earlier when non-AI traffic was worth 128% more per visit.

This is Adobe's own customer base, instrumented through Adobe Experience Cloud. Treat it as representative of that cohort, not all of US retail.

The Seven-Item Checklist, Every Line Cited to a Spec

Skip the blog opinions. Straight from the primary documents:

1. Set both ACP eligibility flags explicitly. is_eligible_search and is_eligible_checkout are Required fields. A missing flag, not a missing description, is the more common reason products get excluded (OpenAI's ACP feed spec).
2. Add seller_privacy_policy and seller_tos URLs if checkout-eligible. Conditionally Required in the same spec. Omitting them blocks Instant Checkout specifically, not search visibility (OpenAI's ACP feed spec).
3. Populate all 16 Google Merchant Center required attributes, including category-conditional ones. Apparel sellers into Brazil, France, Germany, Japan, the UK, or the US carry more mandatory fields than any other category (Google's product data specification).
4. Use variant_dict, not the deprecated Custom_variant1-3_category/_option fields, for ChatGPT-facing variants. The current spec marks the legacy fields deprecated, though several 2025-dated guides still describe them as live (OpenAI's ACP feed spec).
5. Add structured q_and_a and reviews objects, not just a star rating. Both are Recommended fields and one of only two documented paths into ChatGPT's product card, the other being third-party web corroboration (OpenAI's ACP feed spec).
6. On Shopify, treat your UCP profile's trust tier as its own gate, separate from feed completeness. Shopify ties checkout-completion capability directly to trust tier (Shopify's developer docs).
7. Check your Merchant Center attribute completeness score if you're in the US, Canada, Australia, India, or New Zealand, where the rollout began May 27, 2026. This is Google's own diagnostic for AI Mode, AI Overviews, and Gemini discoverability (support.google.com/merchants/answer/17117204).

The Fee Number None of the Five Platforms Publishes

I went looking for one number across five platforms and came back mostly empty-handed.

Across OpenAI's ACP/Instant Checkout, Shopify UCP, Google UCP, the Perplexity Merchant Program, and the Microsoft Copilot Merchant Program, only one comparable transaction fee is publicly disclosed anywhere, and it belongs to Etsy: 6.5% standard. OpenAI's own help center says merchants "pay a small fee on completed purchases" and stops there, with no percentage attached (OpenAI help center). Perplexity, Google, Shopify, and Microsoft publish no fee percentage at all.

That absence is a fact, not a guess, worth saying plainly instead of repeating some vendor's estimated rate as confirmed [UNCERTAIN: exact fee percentages across OpenAI, Shopify, Google, and Microsoft remain undisclosed as of this verification date].

The Biggest Catalog Has the Thinnest Public Spec

Here's a contrast that surprised me while researching this page. Alibaba connected Qwen to the full Taobao and Tmall catalog on May 10, 2026, covering more than 4 billion products and 300 million monthly active users, per Reuters and eMarketer. No public developer documentation equivalent to OpenAI's ACP or Shopify's UCP docs exists for it. Even a detailed third-party technical writeup states directly that "Alibaba has not released full technical documentation" (dev.to). The largest agentic commerce catalog surveyed here has the least verifiable merchant-facing spec of any platform [UNCERTAIN, could change if Alibaba publishes developer docs after this date].

Microsoft's Copilot Merchant Program, launched April 18, 2025, has a similar problem at smaller scale. Merchants share "key product specifications" through an onboarding site Microsoft's blog links to but never details (Microsoft Copilot blog) [UNCERTAIN: Microsoft's public documentation on this program remains thin].

Frequently Asked Questions

What is the Agentic Commerce Protocol (ACP)?

ACP is OpenAI's product feed specification for making items discoverable and purchasable inside ChatGPT, launched September 29, 2025 and built with Stripe (OpenAI). I counted 66 total fields, 16 of them required, straight from the spec before writing a line of this page. Ranking is "organic and unsponsored," per OpenAI, based on availability, price, quality, and whether the merchant is the primary seller.

Does my product feed need to match Google Merchant Center and OpenAI's ACP separately?

Yes. Both specs require 16 fields, but the actual 16 differ enough that a single feed rarely clears both bars automatically. ACP requires checkout-eligibility flags and geo-targeting fields Google never asks for, and Google requires apparel attributes ACP doesn't touch. That comparison came from reading both specs side by side, the one thing every audited guide skipped.

How much does it cost to sell through ChatGPT, Perplexity, or Copilot's shopping features?

I went looking for one clean fee percentage across five platforms and came back mostly empty-handed. None of them discloses an exact number. OpenAI states merchants "pay a small fee on completed purchases" with no figure attached, and Perplexity, Google, Shopify, and Microsoft publish no comparable figure either. Etsy's public 6.5% standard fee is the closest reference point, not any AI platform's actual rate.

Has AI-referred shopping traffic gotten better at converting?

It has, and the swing surprised me enough to double-check it across four separate reports before trusting it. Stitching those Adobe Analytics releases together shows AI-referred retail traffic going from converting 91% worse than non-AI traffic in July 2024 to converting 54% better in May 2026, a reversal of roughly 145 percentage points in under two years.

Do I need schema.org markup if I already have an ACP or Merchant Center feed?

Still yes, and this is the distinction that tripped me up the first time I mapped these layers out. Platform feeds like ACP and Merchant Center get submitted directly to that platform, while schema.org's Product (70 properties) and Offer (51 properties) markup lives on your own site and gets read by crawlers and AI engines that never see your submitted feed at all. Google's structured data docs confirm it reads Product markup for price, availability, and review signals, though Google doesn't publish a field count for schema.org the way it does for Merchant Center.

Where This Leaves Us

Every number on this page traces to a document I fetched myself: OpenAI's feed spec, Google's product data spec, schema.org's Product and Offer pages, Shopify's developer docs, and four dated Adobe Analytics releases. No secondhand summaries, no guessing at field counts from a screenshot.

If you've built a feed against one of these specs and found a field behaving differently than documented, I'd genuinely like to hear about it. Specs drift between blog posts, and the fastest way this page stays accurate is readers telling me what broke on their end 🤗