Ecommerce · Most guides on this topic are running old information

What is agentic commerce, and how do you get AI shopping agents to recommend your store?

Agentic commerce is AI agents researching, comparing and sometimes purchasing on a shopper’s behalf. The critical development is that in-chat checkout stalled — OpenAI ended Instant Checkout in March 2026 after Walmart measured it converting roughly three times worse than clicking through to walmart.com.

The durable model is now discover in AI, buy on your own site. That is good news: AI-referred US retail traffic grew 393% year-over-year in Q1 2026 and converts about 42% better than traditional search, and merchants keep the customer relationship rather than renting it.

The lesson buried in the failure is the one worth acting on. The infrastructure worked. The product data did not — scraped merchant data was wrong on stock, delivery and shipping, which are exactly the three things a shopping agent most needs right. Fixing that is the whole assignment, and it is not wasted under any protocol outcome.

By Jugl17 min readInteractive readiness model29 questions answered

Short answerFor AI overviews

The 60-second version

Agentic commerce is AI agents researching, comparing and sometimes purchasing on a shopper’s behalf. In-chat checkout stalled: OpenAI ended Instant Checkout in March 2026 after Walmart measured it converting roughly 3× worse than clicking through to walmart.com. The durable model is discover in AI, buy on your own site.

The discovery channel is growing fast. AI-referred US retail traffic grew 393% year-over-year in Q1 2026 and converts about 42% better than traditional search. 45% of consumers already use AI somewhere in the buying journey, and ChatGPT handles roughly 50 million shopping-related queries a day.

The failure was about data, not technology. Only around 12–30 Shopify merchants ever went live, and product data scraped from retailer websites was frequently inaccurate on stock availability, delivery timing and shipping costs.

What to do: fix your product data first. Then structured markup, then check you have not blocked AI crawlers, then optimise for comparison. Protocol support matters less than data quality, and the data work is not wasted under any outcome.

What agentic commerce actually is
The agentic commerce picture at a glance
What happened to in-chat checkout
So is agentic commerce dead?
Why discover-in-AI, buy-on-site is better for merchants
The protocol landscape
Score your AI-referral readiness
How to get AI shopping agents to recommend you
The mistakes to avoid
The five questions behind every agentic commerce decision
Where Jugl fits — and where it does not
Methodology and disclosure
FAQ — 21 questions answered
People also ask

01Definition

Definition

What is agentic commerce?

Agentic commerce is AI agents researching, comparing and sometimes purchasing on a shopper’s behalf. It differs from conversational commerce — chat that assists a human shopper — because it delegates decision-making to an AI system: the shopper sets the intent and constraints, and the agent handles discovery and comparison. In-chat transaction has stalled: OpenAI ended Instant Checkout in March 2026 after Walmart measured it converting roughly three times worse than a click-through, and completing a purchase inside an answer engine remains the least-adopted use case among regular users. The operating model is now discover in AI, buy on the merchant’s own site. AI-referred US retail traffic grew 393% year-over-year in Q1 2026 and converts about 42% better than traditional search.

Definition maintained by the Jugl Editorial Team. Jugl sells an AI customer agent platform and is an interested party; it is not a protocol integration or product feed manager, and does not submit catalogues to ACP or UCP.

Why this page carries dates when the rest of our library does not

Almost everything else we publish is written without a year, so it does not silently age. This page is the exception, deliberately. Most published material on agentic commerce is running information from before in-chat checkout was withdrawn, which makes it actively wrong rather than merely stale — it tells merchants to build for a transaction model that no longer exists.

Correcting that requires saying what happened and when. Treat the dated claims here as a snapshot to verify rather than a permanent fact: platform policy in this area has moved more than once and will again. The durable parts — fix your product data, be machine-readable, do not block the crawlers, answer the arriving visitor fast — hold regardless of which protocol wins.

What is durable here

Accurate stock, delivery and shipping data is the prerequisite under every protocol
Structured markup — agents query data, they do not navigate storefronts
Crawler access: blocking by accident makes you invisible for no reason
Comparison-ready product information beats persuasive copy
AI-referred visitors arrive pre-qualified with one question left
Answering that question in seconds is what converts them

What is a snapshot to verify

Which specific protocols exist and who backs them
Whether in-chat checkout stays withdrawn on any given platform
The share of AI app usage held by any single assistant
Whether Amazon continues to block AI crawlers
Exact traffic growth and conversion multipliers, which are period figures
Platform fee structures, which change without much notice

02At a glance

The agentic commerce picture at a glance

At a glance

What it is: AI agents researching, comparing and sometimes purchasing on a shopper’s behalf
The current model: Discover in AI, buy on your own site
Global retail commerce agentic AI will influence by 2030: $3–5 trillion (McKinsey)
Online shoppers using AI agents by 2030: ~50%, about 25% of spend (Morgan Stanley)
AI-referred US retail traffic growth, Q1 2026: +393% year-over-year (Adobe)
AI-referred conversion vs traditional search: ~42% better (Adobe)
Consumers already using AI in the buying journey: 45% (IBM)
ChatGPT weekly active users: ~900 million
ChatGPT shopping-related queries per day: ~50 million
Walmart: in-chat checkout vs click-through conversion: 3× worse
Walmart: new-customer rate from ChatGPT vs search: ~2× better
Shopify merchants ever live on Instant Checkout: ~12–30, of millions
Copilot users more likely to purchase within 30 minutes: 53% (Microsoft)
Merchants supporting multiple protocols: ~40% more agentic traffic
First step for merchants: Fix product data — stock, delivery timing, shipping cost
Biggest structural opportunity: Amazon blocks AI crawlers, so non-Amazon brands have an opening
Where the conversion burden sits: On your side — AI sends the visitor, you close them

SOC 2 Type 2certified

HIPAAcompliant

MetaBusiness Partner

1,000+businesses

03The timeline

What happened to in-chat checkout

+393%AI-referred US retail traffic growth

~42%better conversion than traditional search

3×worse conversion for in-chat checkout

12–30Shopify merchants ever live on it

This is the part most articles still get wrong, because they are running information from before March 2026.

When	What happened
29 September 2025	OpenAI launched Instant Checkout inside ChatGPT. US shoppers could buy from Etsy sellers without leaving the chat, with “over a million” Shopify merchants promised as coming soon. It ran on the Agentic Commerce Protocol, open-sourced with Stripe.
Early March 2026	The Information broke the story that it was ending. CNBC, Modern Retail and Forrester picked it up.
24 March 2026	OpenAI confirmed publicly that the initial version of Instant Checkout did not offer the level of flexibility they aspired to provide, and that merchants could use their own checkout experiences while OpenAI focuses on product discovery.

Why it failed — three reasons, all instructive

Conversion mathsWalmart measured in-chat checkout converting roughly three times worse than a click-through to walmart.com. That is not a tuning problem; that is a structural disadvantage against a checkout the merchant controls.

AdoptionChatGPT users, per a source cited by The Information, simply were not using the chatbot to help them make purchases. Forrester’s consumer research confirms it: completing a purchase is the least-adopted use case among regular answer-engine users, while asking questions and researching products rank first and second.

Data qualityOnly around 12 to 30 Shopify merchants ever went live. Walmart made roughly 200,000 products available, but product data scraped from retailer websites was frequently inaccurate on stock availability, delivery timing and shipping costs.

That third reason is the one merchants should sit with. The infrastructure worked. The product data did not. Every recommendation on this page follows from that single observation, and it is the reason the work is worth doing regardless of what any platform does next.

04The reframe

So is agentic commerce dead?

No — and reading it that way is the expensive mistake.

The transaction moved. The discovery did not. And discovery was always the larger prize:

What the referral channel is actually doing

AI-referred US retail traffic grew 393% year-over-year in Q1 2026 and converts about 42% better than traditional search (Adobe Analytics)
45% of consumers already use AI for some part of the buying journey (IBM)
ChatGPT handles roughly 50 million shopping-related queries a day across around 900 million weekly users
Even in the failed experiment, ChatGPT drove roughly 2× the new-customer rate of search for Walmart

Read those together: AI sends fewer visitors than search, but they arrive pre-qualified, they are likelier to be new, and they convert better. The channel is working exactly as a referral channel should. It just is not a checkout.

Meanwhile the long-term projections have not moved. McKinsey still estimates agentic AI will influence $3–5 trillion in global retail commerce by 2030, and Morgan Stanley still expects roughly half of online shoppers to use AI shopping agents by then, accounting for about 25% of their spending.

05Why it is better

Why discover-in-AI, buy-on-site is better for merchants

Because you keep everything that matters.

Dimension	In-chat checkout	Discover in AI, buy on site
Customer relationship	Mediated by the platform	Yours
Login and account data	Platform-held	Yours
Loyalty programme integration	Difficult	Native
Upsell and cross-sell at checkout	Limited	Full control
Platform transaction fee	4% on top of ~2.9% + $0.30 Stripe	None
Conversion	3× worse (Walmart)	Baseline

Merchants briefly faced a future where an AI platform owned the funnel end to end and took a cut. That future receded. The current arrangement — AI does discovery and comparison, you do the transaction — is the one merchants would have chosen.

06Protocols

The protocol landscape

Protocol	Backer	Status
ACP — Agentic Commerce Protocol	OpenAI + Stripe	Open-sourced September 2025; now used for discovery, with major retailers live
UCP — Universal Commerce Protocol	Google + Shopify	Launched at NRF 2026; Microsoft adopted it in April 2026
AP2	Google	Agent payments
Trusted Agent Protocol	Visa	Payments trust layer
Agent Pay	Mastercard	Payments

Merchants supporting multiple protocols see roughly 40% more agentic traffic. But note the direction of travel: ACP was originally designed to separate discovery from transaction, with Instant Checkout as the exception. That exception is now gone, and the protocols are converging on what they were built for.

Two other developments worth knowing

Microsoft Copilot Checkout launched in the US in January 2026 with Shopify, PayPal, Stripe and Etsy. Microsoft reports Copilot users are 53% more likely to purchase within thirty minutes, and — importantly — merchants stay the merchant of record, owning the transaction, customer data and relationship. That is the structural difference from the model OpenAI withdrew.

Amazon has blocked ChatGPT-User and OAI-SearchBot in its robots.txt, meaning Amazon listings cannot surface in ChatGPT shopping results. That is a structural first-mover advantage for non-Amazon brands in AI product discovery — the largest catalogue in US ecommerce has voluntarily removed itself from the fastest-growing discovery surface, and it will not last forever.

07The model

Score your AI-referral readiness

Eight inputs. Four decide whether you get recommended; the last one decides whether you convert the visitor once you do. Outputs are illustrative estimates from your inputs, not a forecast.

How ready are you for AI-referred traffic?

Product data, markup, crawler access and reply speed — and what each is worth

Sessions a month20,000

Total site sessions across every source. Use your analytics figure — it is the only input here you already have exactly.

Share arriving from AI referrals3%

Traffic from ChatGPT, Gemini, Perplexity, Copilot and AI Overviews. Attribution is genuinely poor here — a shopper who researched in an AI and arrived direct looks like direct traffic.

Your baseline conversion rate2.4%

Your site-wide conversion rate from traditional search and direct traffic. AI-referred visitors convert about 42% better than traditional search.

Average order value$85

Use the average for the products people actually research before buying, which is often higher than your blended basket.

Product data accuracy60%

Stock availability, delivery timing and shipping costs — the three things shopping agents most need right, and the three that were wrong in the first wave of agentic checkout.

Structured markup coverage40%

JSON-LD and Schema.org labelling for price, availability, shipping parameters and product attributes. Agents query structured data; they do not navigate your storefront.

AI crawler accessAllowed

Check robots.txt for ChatGPT-User, OAI-SearchBot and equivalents. Amazon blocks them deliberately as a strategy. Blocking them by accident makes you invisible for no reason.

Reply speed on arriving questions40%

The share of pre-purchase questions you answer within minutes rather than the next morning. AI sends a pre-qualified visitor with one question left.

AI-referral readiness57/100across four weighted inputs

AI-referred sessions600converting at 3.41%

Revenue captured$1,269/mofrom AI-referred visitors

Lost to slow replies$469/mo6 visitors

Upside at full readiness$316/moif data and speed were fixed

57/100. Product data is where this starts, not protocol supportAt 57/100 the constraint is upstream of any protocol decision. Inaccurate stock, delivery and shipping data was a primary reason the first wave of agentic checkout failed — the infrastructure worked and the product data did not. An agent that recommends an out-of-stock product does not merely lose that sale; it stops recommending you. Fix data accuracy first, then markup, then check you have not blocked the crawlers by accident. None of that work is wasted under any protocol outcome.

The conversion half is the half you controlAI sends a pre-qualified visitor with one question left. The free conversation audit reads a real week of your own messages and reports how many pre-purchase questions waited more than an hour.

Get the free auditNo card required

08The method

How to get AI shopping agents to recommend you

The work is unglamorous and it is mostly about data. Six steps.

Fix your product data firstThis is the lesson of the Instant Checkout failure. Scraped merchant data was wrong on stock availability, delivery timing and shipping costs — the three things a shopping agent most needs right. An agent that recommends an out-of-stock product does not just lose that sale; it stops recommending you. Audit for complete structured attributes, accurate real-world descriptions, current pricing and real-time inventory.

Make it machine-readableUse JSON-LD and Schema.org markup to label price, availability, shipping parameters and product attributes explicitly, rather than burying them in prose. Standardise product identifiers. AI agents query structured data; they do not navigate your storefront interface.

Test your current visibilityAsk ChatGPT, Google AI Mode and Perplexity to find products in your category. Do you appear? This takes ten minutes and most merchants have never done it. Vary the phrasing the way a shopper would — by problem, with a budget, with a use case.

Check your protocol coverageConfirm whether your ecommerce platform supports UCP or has API coverage for catalogue, cart, checkout and orders. Shopify merchants are largely covered by default; others should ask.

Do not block the crawlersCheck robots.txt for ChatGPT-User, OAI-SearchBot and equivalents. Amazon blocks them deliberately as a strategic choice. If you are blocking them by accident, you are invisible for no reason.

Optimise for comparison, not just discoveryThe replacement experiences are explicitly built around comparison — shoppers describe items with budget, preferences and constraints, then compare results. Your product data needs to support being compared: clear differentiators, honest specs, sizing, compatibility.

09Mistakes

The mistakes to avoid

Treating ChatGPT as the only surface that mattersIts share of US daily AI app users has fallen substantially while Gemini roughly doubled and Claude tripled its US share in a single month. Optimise for the category, not one app.

Building around in-chat checkout assumptionsThat model is over. Build for referral — a pre-qualified visitor arriving on your site with one question left.

Waiting for the protocols to settleProduct data quality is the prerequisite under every protocol. Do it now; it is not wasted under any outcome, and it improves your own search and your own agent at the same time.

Undercounting AI’s influenceAttribution is genuinely poor here. A shopper who researched in ChatGPT and arrived via direct traffic looks like direct traffic, which means your measured AI share is a floor rather than an estimate.

Assuming B2B is exemptForrester predicts 20% of B2B sellers will face agent-led quote negotiations. Sales-rep-gated catalogues lose future discoverability — if a buyer’s agent cannot read your specifications, it will recommend a supplier whose it can.

10Direct answers

The five questions behind every agentic commerce decision

What is agentic commerce, in practice?

Short answer

AI agents researching, comparing and sometimes purchasing on a shopper's behalf. Unlike conversational commerce, which assists a human shopper, agentic commerce delegates decision-making to the system. In practice most current activity is research and comparison, with the purchase happening on the merchant's own site.

Example

A shopper describes what they want with a budget and constraints, the agent compares options across retailers, and the shopper clicks through to buy. That click-through is your channel, and it converts about 42% better than traditional search.

Key takeawayBuild for referral rather than for in-chat transaction. The discovery half is where the volume is, and it is the half that survived.

Did in-chat checkout die, and does that matter?

Short answer

OpenAI ended Instant Checkout in March 2026 after Walmart measured it converting roughly 3× worse than a click-through. It matters mainly as a signal: the transaction moved back to merchants, and discovery — always the larger prize — kept growing.

Example

Only around 12 to 30 Shopify merchants ever went live, out of millions. The experiment was smaller than the coverage suggested, which is worth knowing before you conclude that a market disappeared.

Key takeawayMerchants briefly faced a future where a platform owned the funnel and took 4% on top of card fees. That future receded, and the current arrangement is the better one.

What is the single most important thing to fix?

Short answer

Your product data — specifically stock availability, delivery timing and shipping costs. Inaccurate scraped data on exactly those three fields was a primary reason the first wave of agentic checkout failed, and it is the prerequisite under every protocol.

Example

An agent that recommends an out-of-stock product does not merely lose that sale. It stops recommending you, because the feedback loop punishes unreliable sources — and you will not be told that it stopped.

Key takeawayDo the data work before the protocol work. It is not wasted under any outcome, and it improves your own site search and your own agent at the same time.

Should I support multiple protocols?

Short answer

Merchants supporting multiple see roughly 40% more agentic traffic, so yes where your platform makes it straightforward. But protocol support matters less than data quality — a perfectly connected catalogue with wrong inventory data will still fail, under every standard.

Example

Shopify merchants are largely covered by default under UCP. Others should ask their platform about API coverage for catalogue, cart, checkout and orders before assuming either way.

Key takeawaySequence it: data accuracy, then markup, then crawler access, then protocol coverage. The order reflects what actually determines whether you get recommended.

What happens when the AI-referred visitor actually arrives?

Short answer

They arrive pre-qualified, knowing roughly what they want, with one or two questions left — will it fit, is it in stock in my size, can it arrive by Friday. Answered in seconds they convert at the better rate. Left until tomorrow morning, the advantage evaporates.

Example

That is the operational meaning of discover-in-AI, buy-on-site: the platform sends a good visitor and hands you the conversion burden. Everything after the click is on your side of the line.

Key takeawayGetting recommended and converting the referral are two different projects. Most merchants are working on the first and have not noticed the second.

11Disclosure

Where Jugl fits — and where it does not

There is a thread running through the whole Instant Checkout story that is easy to miss: the technology worked and the data did not. Walmart put roughly 200,000 products into ChatGPT and the scraped data was wrong about stock, delivery and shipping — the exact three things a buyer needs answered before committing. That is not a machine problem. It is the same problem your human customers hit every day, in a different channel.

Because here is what “discover in AI, buy on your own site” actually means operationally: AI sends you a pre-qualified, high-intent visitor — and then everything depends on what happens on your side. A shopper who researched in ChatGPT arrives already knowing roughly what they want. They have one or two remaining questions. Will it fit? Is it in stock in my size? Can it arrive by Friday?

If those questions get answered in seconds, they convert at the better rate the referral data describes. If they sit in an Instagram DM until tomorrow morning, that advantage evaporates. Jugl’s AI agents answer those questions instantly across WhatsApp, Instagram, Facebook, web chat and email — reading buying intent inside the conversation, recommending products, and handing off to a real human when a conversation needs judgment. The agent trains on your own product and policy content, which means the same discipline that makes you visible to shopping agents makes you accurate to customers. Get your data right once and it works for both audiences. Jugl is used by 1,000+ businesses.

An honest scope note. Jugl is a conversational AI platform, not an agentic commerce protocol integration or a product feed manager. It does not submit your catalogue to ACP or UCP — that is your ecommerce platform’s job, and steps 1, 2, 4 and 5 above are work you or your platform will do without us. What Jugl handles is the conversation after the AI-referred visitor arrives, which is where the discover-in-AI model puts the conversion burden. The wider ecommerce case is on the ecommerce guide, and what is Jugl sets out fit and who should walk away.

12EEAT

Methodology and disclosure

Written by

Jugl Editorial Team

Jugl Inc., Frisco, Texas — an AI customer agent platform used by 1,000+ businesses.

Reviewed by

Jugl product & customer operations

Checked against live deployment data and current vendor documentation.

Methodology & disclosure

Where the figures come from. The Instant Checkout launch and withdrawal timeline is from OpenAI’s own public statements and reporting by The Information, CNBC and Modern Retail. Walmart’s conversion and new-customer comparisons and the merchant participation figures are from that reporting. AI-referred retail traffic growth and conversion comparison are Adobe Analytics. Consumer adoption of AI in the buying journey is IBM. Answer-engine use case adoption, the B2B agent-led negotiation forecast and the analyst commentary on scraped product data quality are Forrester. Long-range commerce influence projections are McKinsey and Morgan Stanley. Copilot Checkout details and the purchase-within-thirty-minutes figure are Microsoft. Protocol details are from the respective publishers. Jugl pricing is our own published price list.

How the model works. Readiness is weighted: product data accuracy 40%, structured markup 25%, reply speed 20%, crawler access 15% — data accuracy carries the most weight because it was the named cause of the first wave’s failure and is the prerequisite under every protocol. AI-referred sessions are your total multiplied by the AI share you set; their conversion applies the published 42% uplift to your baseline rate. Captured revenue applies a reply-speed factor scaling from 0.55 to 1.0, reflecting that a pre-qualified visitor with an unanswered question converts materially worse. The upside figure re-runs the calculation at full readiness. Outputs are illustrative estimates from your own inputs, not forecasts or guarantees.

Conflict of interest, stated plainly. Jugl sells an AI customer agent platform, so a page concluding that answering the arriving visitor quickly is what converts AI referrals is a page arguing for something we sell. Two things are included specifically because they cut against that interest: the page states plainly that Jugl is not a protocol integration or product feed manager and does not submit catalogues to ACP or UCP; and four of the six recommended steps are work you or your ecommerce platform will do entirely without us.

Why this page carries dates. Unlike the rest of our library, which is written without year stamps so it does not silently age, this page states what happened and when — because most published material on agentic commerce is running pre-withdrawal information and is therefore actively wrong rather than merely stale. Treat the dated claims as a snapshot to verify against current platform documentation rather than as permanent facts. The durable recommendations — fix product data, be machine-readable, do not block crawlers, answer the arriving visitor fast — hold regardless of which protocol prevails.

13FAQ

Agentic commerce: 21 questions answered

What is agentic commerce?

Agentic commerce is AI agents researching, comparing and sometimes purchasing on a shopper's behalf. It differs from conversational commerce — chat that assists a human shopper — because agentic commerce delegates decision-making to an AI system: the shopper sets the intent, and the agent handles discovery and comparison. The critical development is that in-chat checkout stalled. OpenAI ended Instant Checkout in March 2026 after Walmart measured it converting roughly three times worse than clicking through to walmart.com. The durable model is now discover in AI, buy on your own site — which is good news for merchants, because AI-referred US retail traffic grew 393% year-over-year in Q1 2026 and converts about 42% better than traditional search, while merchants keep the customer relationship.

What actually happened to in-chat checkout?

This is the part most articles still get wrong, because they are running information from before March 2026. On 29 September 2025, OpenAI launched Instant Checkout inside ChatGPT: US shoppers could buy from Etsy sellers without leaving the chat, with "over a million" Shopify merchants promised as coming soon, running on the Agentic Commerce Protocol open-sourced with Stripe. In March 2026 it ended. The Information broke the story in early March; CNBC, Modern Retail and Forrester picked it up. On 24 March OpenAI confirmed it publicly, stating that the initial version of Instant Checkout did not offer the level of flexibility they aspired to provide, and that they would allow merchants to use their own checkout experiences while focusing on product discovery.

Why did in-chat checkout fail?

Three reasons, all instructive. Conversion maths: Walmart measured in-chat checkout converting roughly three times worse than a click-through to walmart.com. Adoption: ChatGPT users, per a source cited by The Information, simply were not using the chatbot to actually help them make purchases — and Forrester's consumer research confirms it, with completing a purchase the least-adopted use case among regular answer-engine users while asking questions and researching products rank first and second. And data quality: only around 12 to 30 Shopify merchants ever went live, and while Walmart made roughly 200,000 products available, product data scraped from retailer websites was frequently inaccurate on stock availability, delivery timing and shipping costs. That third reason is the one merchants should sit with — the infrastructure worked and the product data did not.

So is agentic commerce dead?

No, and reading it that way is the expensive mistake. The transaction moved; the discovery did not — and discovery was always the larger prize. AI-referred US retail traffic grew 393% year-over-year in Q1 2026 and converts about 42% better than traditional search. 45% of consumers already use AI for some part of the buying journey. ChatGPT handles roughly 50 million shopping-related queries a day across around 900 million weekly users. Even in the failed experiment, ChatGPT drove roughly twice the new-customer rate of search for Walmart. Read those together: AI sends fewer visitors than search, but they arrive pre-qualified, they are likelier to be new, and they convert better. The channel is working exactly as a referral channel should — it just is not a checkout.

Have the long-term projections changed?

No. McKinsey still estimates agentic AI will influence $3–5 trillion in global retail commerce by 2030, and Morgan Stanley still expects roughly half of online shoppers to use AI shopping agents by then, accounting for about 25% of their spending. What changed is the mechanism rather than the magnitude: influence over the purchase decision rather than execution of the transaction. That distinction matters commercially because influence is a referral channel you can optimise for with work you control, while execution would have been a platform relationship you negotiate. The first is a better position for a merchant to be in, and it is the one that turned out to be durable.

Why is discover-in-AI, buy-on-site better for merchants?

Because you keep everything that matters. Under in-chat checkout, the customer relationship was mediated by the platform, login and account data were platform-held, loyalty programme integration was difficult, upsell and cross-sell at checkout were limited, and OpenAI charged 4% on top of roughly 2.9% plus $0.30 in Stripe fees — while converting three times worse in Walmart's measurement. Under discover-in-AI, buy-on-site, the relationship is yours, the data is yours, loyalty is native, you have full control of checkout merchandising, and there is no platform transaction fee. Merchants briefly faced a future where an AI platform owned the funnel end to end and took a cut. That future receded, and the current arrangement is the one merchants would have chosen.

What is the current protocol landscape?

Five competing standards launched over roughly a year, and the fragmentation is real. ACP, the Agentic Commerce Protocol from OpenAI and Stripe, was open-sourced in September 2025 and is now used for discovery, with major retailers live. UCP, the Universal Commerce Protocol from Google and Shopify, launched at NRF 2026 with Microsoft adopting it in April 2026. AP2 is Google's agent payments standard, Visa has a Trusted Agent Protocol as a payments trust layer, and Mastercard has Agent Pay. Merchants supporting multiple protocols see roughly 40% more agentic traffic. Note the direction of travel: ACP was originally designed to separate discovery from transaction, with Instant Checkout as the exception — that exception is now gone and the protocols are converging on what they were built for.

What about Microsoft Copilot Checkout?

Copilot Checkout launched in the US in January 2026 with Shopify, PayPal, Stripe and Etsy. Microsoft reports Copilot users are 53% more likely to purchase within thirty minutes, and — importantly — merchants stay the merchant of record, owning the transaction, customer data and relationship. That last point is the structural difference from the model OpenAI withdrew, and it is why this one is worth watching rather than dismissing. It is also a useful reminder not to treat one surface as the whole category: ChatGPT's share of US daily AI app users has fallen while Gemini roughly doubled and Claude tripled its US share in a single month. Optimise for the category, not one app.

Why does Amazon blocking AI crawlers matter?

Amazon has blocked ChatGPT-User and OAI-SearchBot in its robots.txt, which means Amazon listings cannot surface in ChatGPT shopping results. That is a structural first-mover advantage for non-Amazon brands in AI product discovery — the largest catalogue in US ecommerce has voluntarily removed itself from the fastest-growing discovery surface. It will not last forever, and the strategic implication is about timing rather than permanence: the merchants who fix their product data and markup now are competing for AI recommendations in a market where the dominant player is absent. Check your own robots.txt while you are thinking about it, because blocking those crawlers by accident is common and makes you invisible for no reason.

How do I get AI shopping agents to recommend my store?

Six steps, and the work is unglamorous and mostly about data. Fix your product data first — accurate stock availability, delivery timing and shipping costs, which are the three things a shopping agent most needs right. Make it machine-readable with JSON-LD and Schema.org markup labelling price, availability, shipping parameters and product attributes explicitly rather than burying them in prose. Test your current visibility by asking ChatGPT, Google AI Mode and Perplexity to find products in your category. Check your protocol coverage with your ecommerce platform. Do not block the crawlers by accident — check robots.txt for ChatGPT-User and OAI-SearchBot. And optimise for comparison rather than just discovery, with clear differentiators, honest specs, sizing and compatibility.

Why is product data the first step rather than protocol support?

Because it is the lesson of the Instant Checkout failure and the prerequisite under every protocol. Scraped merchant data was wrong on stock availability, delivery timing and shipping costs — the three things a shopping agent most needs to be right about. An agent that recommends an out-of-stock product does not just lose that sale; it stops recommending you, because the feedback loop punishes unreliable sources. Audit for complete structured attributes, accurate real-world descriptions, current pricing, and real-time inventory. The commercial argument for doing this first is that the work is not wasted under any protocol outcome: whichever standard wins, or if none does, accurate structured product data is what makes you recommendable.

What does machine-readable actually mean here?

Using JSON-LD and Schema.org markup to label price, availability, shipping parameters and product attributes explicitly, rather than leaving them in prose for something to infer. AI agents query structured APIs and structured data; they do not navigate your storefront interface the way a human does. Standardise product identifiers so the same product is recognisably the same across surfaces. The practical test is whether a machine reading your page can answer "is this in stock in medium, what does shipping cost to this postcode, and when would it arrive" without guessing. If any of those three requires interpretation, that is where you will be misrepresented — and the misrepresentation will be confident.

How do I test my current AI visibility?

Ask ChatGPT, Google AI Mode and Perplexity to find products in your category and see whether you appear. This takes ten minutes and most merchants have never done it. Vary the phrasing the way a shopper would — by problem rather than by product name, with a budget constraint, with a use case. Note which competitors surface consistently and look at what their product pages do differently; it is usually markup completeness and specificity rather than brand strength. Repeat it monthly, because the surfaces change ranking behaviour frequently and the composition of AI app usage is shifting. Ten minutes a month is a reasonable price for knowing whether your discovery channel exists.

Should I optimise for discovery or for comparison?

Both, but comparison is the one merchants under-invest in. OpenAI's replacement experience is explicitly built around comparison — shoppers upload images or describe items with budget, preferences and constraints, then compare visual results. Your product data needs to support being compared, which means clear differentiators, honest specifications, sizing and compatibility. A product page written to persuade a human reader often performs badly here, because persuasive copy is imprecise by design. The version that gets recommended states the attributes plainly and does not overstate them, since an agent that recommends something inaccurately described stops recommending that source.

What are the mistakes to avoid?

Five. Treating ChatGPT as the only surface that matters — its share of US daily AI app users has fallen substantially while Gemini roughly doubled and Claude tripled its US share in a single month, so optimise for the category rather than one app. Building around in-chat checkout assumptions, since that model is over and you should build for referral. Waiting for the protocols to settle, when product data quality is the prerequisite under every protocol and doing it now is not wasted under any outcome. Undercounting AI's influence, because attribution is genuinely poor — a shopper who researched in ChatGPT and arrived via direct traffic looks like direct traffic. And assuming B2B is exempt, when Forrester predicts 20% of B2B sellers will face agent-led quote negotiations.

How should I measure AI-referred traffic?

Carefully, and with the knowledge that you are undercounting. Segment referral traffic by AI source where the referrer is visible, but recognise that a shopper who researched in ChatGPT and then typed your domain looks like direct traffic — which means your measured AI share is a floor rather than an estimate. Two proxies help: growth in branded search alongside flat traditional referral, and a rise in direct traffic with above-average conversion. Track conversion rate for identified AI-referred sessions separately, because the published figure is roughly 42% better than traditional search and you want to know whether you are seeing that. If you are not, the problem is usually on your side rather than theirs.

Does this apply to B2B?

Yes, and assuming otherwise is one of the named mistakes. Forrester predicts 20% of B2B sellers will face agent-led quote negotiations, and sales-rep-gated catalogues lose future discoverability by design — if a buyer's agent cannot see pricing, specifications or availability without a form fill, the agent will recommend a supplier it can read. The tension is real for businesses whose pricing is genuinely negotiated, and the practical middle ground is publishing enough structured specification and availability data to be discoverable while keeping commercial terms behind the conversation. The discoverability decision and the pricing decision are separable, and treating them as one is what makes catalogues invisible.

What does an AI-referred visitor actually want when they arrive?

One or two things confirmed. A shopper who researched in ChatGPT arrives already knowing roughly what they want — that is why they convert about 42% better than traditional search traffic. They have a remaining question: will it fit, is it in stock in my size, can it arrive by Friday. If those get answered in seconds, they convert at the better rate the data describes. If the question sits in an Instagram DM until tomorrow morning, that advantage evaporates entirely and you have paid for a pre-qualified visitor and lost them at the last step. This is the operational meaning of discover-in-AI, buy-on-site: the platform sends a good visitor and hands you the conversion burden.

What is the difference between agentic and conversational commerce?

Conversational commerce is chat that assists a human shopper — the person is deciding, and the system is helping. Agentic commerce delegates decision-making to an AI system: the shopper sets the intent and constraints, and the agent handles discovery and comparison on their behalf. In practice most current activity is the first with a layer of the second: consumers use AI to research and compare, then decide and buy themselves. The commercial implication is that both matter and they operate at different points — agentic discovery brings the visitor, and conversational commerce on your own channels is what closes them. Our ecommerce guide covers the second half.

Where should a merchant start if resources are limited?

Product data accuracy, then markup, then reply speed on arriving questions — in that order, and each is worth doing alone. Data accuracy is first because it determines whether you get recommended at all and because it improves everything downstream including your own search and your own agent. Markup is second because it is a bounded technical task with a clear finish line. Reply speed is third only because it is the one most merchants already have some capability for. Protocol support comes last, not because it does not matter but because it is largely determined by your ecommerce platform and it fails anyway if the data underneath is wrong.

How does Jugl fit into agentic commerce?

There is a thread running through the whole Instant Checkout story that is easy to miss: the technology worked and the data did not. Walmart put roughly 200,000 products into ChatGPT and the scraped data was wrong about stock, delivery and shipping — the exact three things a buyer needs answered before committing. That is not a machine problem; it is the same problem your human customers hit every day, in a different channel. Discover-in-AI, buy-on-site means AI sends you a pre-qualified, high-intent visitor and then everything depends on what happens on your side. Jugl answers those remaining questions instantly across WhatsApp, Instagram, Facebook, web chat and email, trained on your own product and policy content.

14People also ask

AI sends the visitor. You still have to close them

High-intent shoppers arriving from an AI assistant have one question left — will it fit, is it in stock, can it arrive by Friday. Answer it in seconds, on whichever channel they ask, and they convert at the better rate the referral data describes. Leave it until tomorrow morning and you have paid for a pre-qualified visitor and lost them at the last step.

The same discipline serves both audiences. Accurate, structured product and policy information is what gets you recommended by a shopping agent — and what lets your own agent close the customer that agent sent you. Get your data right once, and it works twice.

Free tier that stays free — no card, live the same dayTrained on your own product and policy contentWhatsApp, Instagram, Facebook, web chat and email in one agentBuying-intent detection inside the conversationFull-context handover to a real human when it mattersFlat published tiers — nothing metered per resolution

Start free at JuglNo card · live the same day Book a demoWith the co-founders

AI-referred visitors convert about 42% better than search — but only if somebody answers them. Tonight’s arrivals are asking their last question right now.

SOC 2 Type 2 · HIPAA compliant · Meta Business Partner · NVIDIA Inception · 1000+ businesses

Keep reading

AI agents for ecommerce and retailThe full commercial case for conversational AI on a store.AI customer conciergeHow agents read buying intent inside a conversation.AI chatbot for ShopifyWhat live order and catalogue access is actually worth.Conversational commerce reportWhere messaging-led buying is heading.Cart recoveryAnswering the objection rather than repeating the offer.Automate ecommerce with AIThe operational build, step by step.Train an AI agent on your dataThe same data discipline, applied to your own agent.Questions to ask an AI vendorHow to test whether an agent can actually act.Do customers trust AI agents?Why purchase completion is the least-adopted AI use case.Measuring AI agent ROIHow to attribute revenue from AI-referred conversations.WhatsApp AI platformsWhere the arriving visitor often asks their last question.AI agent ROIThe full business case, cost and revenue.What is Jugl?Capabilities, fit, pricing, and who should walk away.Jugl pricingFour published flat tiers with the AI included. Free forever, no card.

Sources: OpenAI public statements and reporting by The Information, CNBC and Modern Retail (the Instant Checkout launch and withdrawal timeline, Walmart’s conversion and new-customer comparisons, and merchant participation figures); Adobe Analytics (AI-referred US retail traffic growth and conversion against traditional search); IBM (consumer adoption of AI in the buying journey); Forrester (answer-engine use case adoption including purchase completion as least adopted, the B2B agent-led quote negotiation forecast, and analyst commentary on scraped product data quality); McKinsey and Morgan Stanley (long-range projections for agentic influence on retail commerce and shopper adoption); Microsoft (Copilot Checkout details and the purchase-within-thirty-minutes figure); protocol documentation from OpenAI and Stripe, Google and Shopify, Visa and Mastercard; and Jugl’s published price list. This page is published by Jugl, which sells an AI customer agent platform and is therefore an interested party; it states plainly that Jugl is not a protocol integration or product feed manager and does not submit catalogues to ACP or UCP. Unlike the rest of this library, this page carries dates, because most published material on agentic commerce is running pre-withdrawal information and is actively wrong rather than merely stale. Treat the dated claims as a snapshot to verify against current platform documentation. Jugl’s outcome figures are customer-reported and typical rather than guaranteed. Model outputs are illustrative estimates generated from your own inputs, not forecasts or guarantees. Meta, WhatsApp, Messenger, Instagram and Facebook are trademarks of Meta Platforms, Inc.; Jugl is a Meta Business Partner and this page is published by Jugl and is not endorsed by or affiliated with Meta Platforms, Inc. All other product names are trademarks of their respective owners.

Start free at Jugl · No card required · Permanent free tier