Back to the Lab

Research Lab · Article · Agentic selling

How to Sell to AI Agents: What Buying Machines Check First

How to sell to AI agents, program by program: OpenAI's eleven feed fields, Google's checkout gate, Copilot's checkout default, and a readiness checklist.

By Arseniy Kirykouski, Alessio Savi14 min read
How to Sell to AI Agents: What Buying Machines Check First

In brief. Selling to AI agents is program work, and the rules are published. OpenAI requires eleven feed fields, chains checkout eligibility behind search eligibility, and sends buyers to the merchant's own checkout, Google gates a UCP checkout button behind Merchant Center and select-merchant status, Copilot Checkout is on by default for eligible Shopify stores, with a per-channel toggle, and Perplexity's Instant Buy works only with merchants compatible with its checkout. Each buying step maps to a seller obligation, and the readiness checklist in this piece turns those obligations into pass or fail checks you can run this week.

  • OpenAI's product feed spec requires eleven fields for a discovery feed, nine of basic product data plus two geo fields, and a product cannot be checkout-eligible in ChatGPT unless it is search-eligible first.
  • Google runs agentic checkout for AI Mode and Gemini through Merchant Center, with a UCP checkout attribute and access for select merchants in the US, Canada, and Australia.
  • Copilot Checkout is on by default for eligible Shopify stores, with a per-channel toggle in the Shopify admin, while other merchants onboard through Stripe or a Microsoft Merchant Center account with a UCP-compliant feed.
  • Perplexity's Instant Buy works only with merchants compatible with its agentic checkout, and Perplexity says advertisers cannot pay to list their products in its related-products section.
  • A per-program readiness checklist turns each agent buying step into a pass, fail, or unknown seller obligation you can verify on the live store.

OpenAI's product feed specification lists eleven required fields for a discovery feed. Nine describe the product: item_id, title, description, url, brand, seller_name, image_url, availability, and price. Two say where it sells: target_countries and store_country. That list reads like a sales script written for a buying agent, and it is the most concrete answer any platform has published to what selling means when the buyer is software.

Search "how to sell to AI agents" and most of page one answers a different question: how to sell AI tools to human sales teams. The operator question is narrower. What exactly do ChatGPT, Gemini, Copilot, and Perplexity need from your catalog before they can retrieve a product, compare it against rivals, and complete a purchase?

Each program has published its answer in a feed spec, a help center, a blog post, and a merchant program. Almost nobody reads the four documents side by side. This guide does, then inverts them into a checklist you can run against your own store this week.

For a buying agent, the sales conversation starts in the product feed.

Selling to AI agents starts with the programs' own rules

Selling to AI agents means meeting four published programs on their terms: a product feed with exact required fields for OpenAI, Merchant Center data plus a checkout attribute for Google, a recorded on-or-off decision for Microsoft's Copilot Checkout, and a checkout that Perplexity's Instant Buy supports, so shopping agents can retrieve, compare, select, and buy your products.

These four are the buying surfaces this guide covers, each with documented seller entry points. They form the sell side of the buy side and the sell side of agentic commerce: the buying agent belongs to the platform, and the evidence it buys from belongs to you. That seller-side craft is Agentic Selling, and its raw material is the catalog.

Who sells when AI buys
The brand owns offer, answer, checkout, rules.

The programs differ on three axes that decide your order of operations: what you must publish, who is allowed in, and where the purchase happens. Get those three straight per program and "how to sell to AI agents" turns into a short list of files, flags, and forms.

Every step the agent takes is an obligation you own

A buying agent performs up to four consequential steps, depending on the platform: it retrieves candidate products, compares them on aligned attributes, selects under the shopper's constraints, and completes or hands off checkout. Read from the seller's side, each step is a requirement with a named enforcement point.

Agent stepWhat the agent doesYour obligationWhere it is enforced
RetrievalPulls candidates from feeds, indexes, and pagesPublish an ingestible feed and accessible product URLsOpenAI feed spec, Merchant Center, Perplexity's index
ComparisonAligns price, availability, and attributes across sellersFill every required field with a current, page-consistent valueRequired-field validation in each program
SelectionShows only what it may showSet eligibility flags and clear the account gatesis_eligible_search, UCP checkout attribute, program approval
CheckoutBuys directly or hands off to your storeKeep your checkout open to agents, implement UCP where offered, and keep policy URLs liveYour own site for ChatGPT, UCP, Copilot Checkout

The third column is the article. Everything below walks those obligations program by program, at the altitude where an e-commerce team actually works: field names, flags, account gates, and forms.

The OpenAI product feed spec requires eleven fields in every discovery feed, nine of basic product data plus target_countries and store_country, and "the standard OpenAI-format upload currently targets the US." One of them gives the game away: seller_name exists so an agent can present an offer and say who stands behind it.

Eligibility is explicit, and it is chained. In the spec, is_eligible_search set to true "enables search eligibility; false disables it and checkout eligibility." For is_eligible_checkout, "true opts in only when search eligibility is also true and checkout is enabled for the integration." The spec adds that "eligibility does not guarantee display." A product can be searchable yet unbuyable. A buyable product is always searchable first. Checkout also "requires a separately enabled integration", its rows need live seller_privacy_policy and seller_tos URLs, and a separate is_ads_eligible flag governs sponsored placement.

Search eligibility is the tollgate.

Formats are where secondhand advice goes wrong most often. The popular belief, repeated across agency explainers, is that any structured export will do and OpenAI will cope. The spec is stricter. It is OpenAI's own schema, written for JSONL records or CSV and TSV files: booleans go in as true or false, identifiers stay strings "to preserve leading zeros", and "product and image URLs must be publicly accessible." A Merchant Center export uses Google's attribute names, such as link and image_link, so it is not an OpenAI feed until it is mapped and converted.

The full merchant path no longer centers on a standalone checkout inside ChatGPT. On March 24, 2026, OpenAI wrote that "the initial version of Instant Checkout did not offer the level of flexibility that we aspire to provide", let merchants use their own checkout, and extended the Agentic Commerce Protocol (ACP) to product discovery. OpenAI's commerce documentation is now a six-step partner playbook, from submitting product feeds to accepting ChatGPT Wallet payments and claiming the business, and once the shopper decides, "the journey moves to browser use or a plugin." Since November 10, 2026, Business Hub, in beta, lets verified businesses add self-service product feeds; since November 13, ChatGPT Wallet, rolling out to US Pro subscribers, lets their agents pay through the merchant's existing checkout flow, where accepting ChatGPT Wallet purchases depends on browser access and fraud-rule recognition. On Shopify, ChatGPT buyers complete the purchase on the store's checkout, opened in a ChatGPT in-app browser. The feed makes you a candidate either way; the operational path from claiming your store through connecting its feed and publishing an agent is its own playbook, built on the same catalog truth.

One jacket through the feed, field by field

Take an illustrative product from Fernline, an invented outdoor brand: the Marlowe Field Jacket in waxed olive cotton, $248. Its feed row is eleven decisions, and each wrong cell carries a specific cost at the moment an agent shops the category.

FieldFernline's valueWhat a wrong value costs
item_idFL-MFJ-OLV-MDuplicated IDs collapse two variants into one listing
titleMarlowe Field Jacket, Waxed Cotton, OliveA vague title loses the "waxed jacket" query
description10 oz waxed cotton, corduroy collar, four pocketsMissing attributes drop it from comparisons
urlfernline.com/products/marlowe-field-jacketA dead link ends the journey at retrieval
image_urlcdn.fernline.com/marlowe-olive.jpgA broken image leaves the product card bare
brandFernlineBrand-name queries miss the product entirely
price248.00 USDA price that contradicts the page discredits the feed
availabilityin_stockRecommending a sold-out jacket burns the shopper once
seller_nameFernlineThe agent cannot say who is selling
target_countries["US"]Left empty, it does not mean worldwide, and the row lacks a required value
store_countryUSAny market beyond the US waits for OpenAI to confirm the integration

Now the flags. Fernline sets is_eligible_search to true, which makes the jacket a candidate in ChatGPT search. Checkout needs is_eligible_checkout set to true as well, honored only while search is true and checkout is enabled for the integration, plus a privacy policy and terms of service resolving at the seller_privacy_policy and seller_tos URLs in the feed. The flag alone "does not complete checkout onboarding"; OpenAI enables the integration separately. Sequence is the design principle: discovery comes first, and checkout follows.

Google keeps Merchant Center as the hub and gates the button

Google's route runs through infrastructure most merchants already operate. The Universal Commerce Protocol, which Google defines as "a new open standard for agentic commerce", lets an integrated merchant "implement a checkout button on eligible product listings in AI Mode in Google Search and on Gemini". The data side stays where it was: "Merchant Center will continue to be the central hub to prepare your product data". Google announced UCP at NRF in January 2026, co-developed with Shopify and a group of large retailers, under the Apache 2.0 license.

Entry to agentic checkout on Google has three checkable parts: the checkout eligibility product attribute (native_commerce with checkout_eligibility) on your offers, a Google Pay & Wallet Console account, and Google's permission, since checkout is "available for select merchants at this time", with eligibility listed for the US, Canada, and Australia. The button went live first for US shoppers buying from Etsy and Wayfair, reached Canadian shoppers on November 4, 2026, and since October 15 Google's Universal Cart has been rolling out in the US with checkout powered by UCP. Google says that starting next year it will begin broadening access by making "simplified UCP integration available directly within Merchant Center."

The gating extends past checkout. Google's Business Agent prerequisites ask for a US e-commerce store, a verified Merchant Center account, at least 50 approved free listings, and a claimed brand profile. Approval rate and offer count have quietly become agent-entry criteria, which means the unglamorous work of tuning a Google Shopping feed for AI surfaces now buys entry at two doors.

For most brands the Google door opens on Google's schedule. The feed behind it is yours to fix today.

Microsoft enrolls Shopify merchants before they apply

Microsoft's Copilot Checkout announcement of January 8, 2026 contains the most consequential default in agentic commerce. From the post: "If you're on Shopify, you'll automatically be able to sell in Copilot Checkout. No application. No integration." Enrollment is automatic "following an opt-out window", and Microsoft pitches the buying moment as a purchase completed in place, with no redirect and the merchant staying merchant of record.

Outside Shopify, Microsoft's merchant page now names Stripe's Agentic Commerce Suite or a Microsoft Merchant Center account with "a UCP-compliant feed", and at this time limits checkout to English-language merchants selling to US buyers, on Copilot.com and the Copilot mobile app. Shopify's help pages add that selling in Copilot's direct checkout carries no fees beyond standard payment processing.

An opt-out program flips the operator's job: participation is the default state, and the real decision is whether your catalog, exactly as it stands, is ready to be sold inside a Copilot conversation. Audit first, then choose, then write the choice down.

Silence is a yes here. As of November 2026 the switch lives in the Shopify admin: under Sales channels > Agentic, a merchant turns Copilot's direct checkout on or off, and Microsoft describes UCP as the layer that "enables transactions" in Copilot.

Perplexity checks out with compatible merchants and keeps ads separate

Perplexity's Instant Buy lets US users buy directly on Perplexity, but only "with merchants who are compatible with our agentic checkout experience." One documented route is PayPal's: since November 25, 2025, its store sync makes a merchant's catalog discoverable in Perplexity's results, and its agent ready service accepts payment through the store's existing PayPal setup, with the retailer as merchant of record. Perplexity's 2024 merchant program, described then as free for merchants, still has a live interest form, but no current Perplexity page confirms its terms, so treat that program as unconfirmed. Advertisers "cannot pay to list their products" in Perplexity's related-products section, which keeps organic inclusion and paid placement legible as separate decisions.

Perplexity publishes fewer field-level specifics than OpenAI or Google. Its help center says that its answer engine "is more likely to recommend merchants who provide deeper product details such as availability, reviews, pricing, and specifications." The calculation is short: reuse the governed feed you built for the stricter specs, ask your payment provider whether it already connects you, and watch which products surface.

Which program do you start with?

Side by side, the four programs form an uneven chessboard: different entries, different defaults, different gates.

Four doors into the structured substrate
The same identifiers and offers, sent to every program.
ProgramEntry mechanicsParticipation defaultGating todayWhere the purchase happens
OpenAI (ChatGPT)Product feed per the public spec; self-service upload in Business Hub once the business is verifiedOpt in; Shopify catalogs already integratedCheckout chained behind the search flag, policy URLs, and a separately enabled integrationOn your own site or app (on Shopify, in a ChatGPT in-app browser)
Google (AI Mode, Gemini)Merchant Center data, checkout attribute, Google Pay & Wallet ConsoleOpt in"Select merchants"; US, Canada, AustraliaCheckout button on eligible listings, and Universal Cart in the US
Microsoft (Copilot)Automatic for Shopify; Stripe's Agentic Commerce Suite or a Merchant Center account with a UCP-compliant feedOn by default on Shopify, with a per-channel toggleEnglish-language merchants selling to US buyers, at this time; Copilot.com and the mobile appInside the Copilot conversation
PerplexityPayPal store sync, or the 2024 merchant interest formOpt inInstant Buy only with compatible merchants; US usersInstant Buy inside the answer for eligible merchants

If you run on Shopify, part of the decision is already made. Shopify's help documentation lists agentic storefront channels for ChatGPT, Google AI Mode and Gemini, Microsoft Copilot, and Meta as "active by default for eligible stores", and Shopify's agent documentation is explicit about the scale: one endpoint lets agents search every Shopify merchant, build carts, and hand checkout back to the merchant. On Shopify the work is verification: which channels are on, which flags are set, which defaults you accept.

Everywhere else, sequence by gate. Build the OpenAI feed first, because the spec is public and its format rules force catalog discipline, then add it in Business Hub once your business is verified. Check Perplexity's route the same week: PayPal's store sync if you take PayPal, the interest form if you do not. Fix Merchant Center while Google's select-merchant gate decides your timing, and for Copilot Checkout add a UCP-compliant feed in Microsoft Merchant Center or go through Stripe. Every feed above carries your product URLs, so the same work also serves an agentic storefront on the brand's own domain, where people and buying agents read one catalog truth.

Run the readiness checklist before you spend anywhere else

The checklist folds all four programs into one artifact. Run it against your live store, one row at a time, and write pass, fail, or unknown in the result column with a date next to it. Unknown is a finding. If you want an outside read first, the free report reads your catalog from the outside, the way a buying agent does.

#CheckProgramHow to verifyResult (pass / fail / unknown)
1A feed exists in OpenAI's schema, as JSONL, CSV, or TSVOpenAIOpen the exported file, check field names and delimiter
2All eleven required fields populated on every rowOpenAISample 25 SKUs against the spec's required list
3is_eligible_search set deliberately per productOpenAIRead the flag on your 20 best sellers
4No row has checkout true while search is falseOpenAIFilter the feed on the two flags
5seller_privacy_policy and seller_tos URLs loadOpenAIOpen both URLs straight from the feed
6Merchant Center account verified and offers approvedGoogleMerchant Center diagnostics
7Business Agent prerequisites met (US store, 50+ approved free listings, claimed brand profile)GoogleCount approved listings, confirm verification and the brand profile
8Checkout attribute and Google Pay & Wallet Console in placeGoogleConfirm the attribute and the console account
9Copilot direct-checkout setting known and the decision recordedMicrosoftShopify: Sales channels > Agentic; others: Merchant Center feed or Stripe status
10Perplexity route checked (PayPal store sync or interest form), or declined on recordPerplexityStatus with PayPal or the form
11Product URLs load without login walls or geo blocksAllFetch 10 product URLs from a clean session
12Page and feed agree on price and availabilityAllCompare 10 SKUs on both surfaces

When rows 2 and 12 fail together, the usual root cause is organizational: the feed and the product pages are owned by different teams on different schedules. Fix ownership before format.

Where the brand agent fits

Veliu is an agentic commerce lab building the Brand Agent, an agent that sells for one brand: it studies how your category gets shopped across ChatGPT, Gemini, Copilot, and Perplexity, readies the store so every required field carries a current, defensible value, and sells to people and to the AI shopping agents that arrive, in the brand's voice, with checkout on the brand's rails. We set out the thesis in our founding paper.

The sale, on the brand’s rails
The brand agent asks, verifies, answers, then closes the sale on the brand’s checkout.

What this means for you

  • Field names are now revenue work. The agent's buying behavior is documented, so catalog quality has a written rubric: eleven fields for OpenAI, an attribute plus a console for Google, a decision for Microsoft, a compatible checkout for Perplexity.
  • Defaults decide your exposure. Microsoft enrolls Shopify merchants automatically and Shopify activates agentic channels for eligible stores, while Google waits for you to qualify. Map where you already are before choosing where to apply.
  • Eligibility flags are strategy. is_eligible_search decides which products ChatGPT may show, and is_eligible_checkout waits on an integration OpenAI enables separately; set them per product, deliberately, and re-verify after every catalog sync.
  • Agreement beats volume. A feed that contradicts its own product pages loses comparisons the catalog could have won, on every surface at once.
  • Gates move, feeds compound. Select-merchant programs widen on the platform's schedule, and the feed you clean now is the entry ticket each time a gate opens.

The concrete next step: export your current product feed this afternoon, check 25 rows against OpenAI's eleven required fields, and fill the OpenAI block of the readiness checklist before the week ends.

Authors: Arseniy Kirykouski, Member of Technical Staff at Veliu, and Alessio Savi, Member of Technical Staff at Veliu

Methodology note: Veliu Editorial Team reviewed the linked primary documentation from OpenAI, Google, Microsoft Advertising, Perplexity, PayPal, and Shopify, first on August 6, 2026, and re-read every linked source and Shopify's Help Center pages on November 17, 2026. Program availability changes quickly: gates described as select-merchant or US-first can widen or close without notice, and this review did not test a live merchant account on any of the four programs.

The next paper, when it is written.

One email per paper, and nothing else.

Subscribe

More from the blogResearch Lab