Sell Your Product Inside ChatGPT: 2026 Guide

The founder's playbook for getting discovered, recommended, and bought inside the world's biggest AI assistant.

ChatGPT now handles more than 2.5 billion prompts a day for over 900 million weekly users, and roughly 50 million of those daily prompts are people deciding what to buy - TechCrunch. That is a shopping mall the size of a small country, open every hour, where the shop assistant is trusted, tireless, and answers in full sentences. For the first time since Google indexed the web, there is a genuinely new front door to commerce, and most founders have no idea whether they are standing inside it or locked out.

The obvious reading of "sell your product inside ChatGPT" is the literal one: a shopper types a question, your product appears, they tap Buy, and the money lands in your account without anyone leaving the chat. That version exists. It is called Instant Checkout, it launched on September 29, 2025, and it briefly sent Etsy's stock up 16% in a single day - CNBC. But the literal reading is also the smaller half of the story, and by mid-2026 it was the half that had already been quietly rebuilt.

Here is the problem this guide solves. Nearly every article you will find on this topic tells you to "turn on Instant Checkout" and stops there. That advice was written before the feature stumbled, before only about 30 Shopify merchants ended up live on it, before in-chat purchases converted roughly three times worse than a click to the merchant's own site, and before OpenAI itself stepped back and told merchants to use their own checkout while it refocused on discovery - Modern Retail. Following that advice today means optimizing for the wrong thing.

This guide is built from first principles instead. It separates the hype (a buy button) from the durable opportunity (being the product an AI recommends to a purchase-ready human), gives you the real mechanics of both, and shows you exactly where to put your effort in 2026. It covers the product feed and the ChatGPT Merchant program, the Agentic Commerce Protocol that underpins in-chat payments, the Apps SDK for building an interactive presence inside ChatGPT, the generative-engine-optimization tactics that actually move the needle (and the ones that do not), how the competing surfaces from Google, Perplexity, Microsoft, and Amazon compare, and a sequenced playbook you can start this week. Throughout, we point you to deeper dives such as our companion guide on how to sell to AI agents and how to get your site cited by ChatGPT and Claude.

Contents

1. What "selling inside ChatGPT" actually means in 2026
2. Inside the numbers: how much AI shopping is really happening
3. Layer one: getting your products into ChatGPT's index
4. Getting recommended: what makes ChatGPT pick you
5. Layer two: in-chat checkout and the Agentic Commerce Protocol
6. Layer three: building an app inside ChatGPT
7. The competitive map: where else AI shoppers can find you
8. The practical playbook: what to do, in what order
9. Failure modes, limits, and honest risks
10. The future outlook: when your customer is an agent

Before the deep dive, here is the whole category on one screen. If your real question is "which AI assistant should I invest in to get my product in front of buyers," this weighted scorecard answers it. Every AI shopping surface is scored on the four things a founder actually cares about, with the real data point sitting inside each cell so you can see why the score is what it is. It is ordered by final score, highest first.

The criteria are weighted for an independent founder or small brand, not a Fortune 500 retailer. Merchant openness carries the most weight (30%) because a channel you cannot actually join is worth nothing to you, no matter how large. Reach and discovery value each carry 25% because attention only matters if it can find your specific product. Economics and ownership carries 20% because the fee and, more importantly, whether you keep the customer relationship decide whether a sale is worth having. The single most important takeaway from this table is structural: the winners are the open ecosystems you can opt into for free, and the loser for an independent seller is the one closed garden, Amazon, precisely because being big is not the same as being open. We unpack every row in section 7.

1. What "selling inside ChatGPT" actually means in 2026

Start with the structural question rather than the surface one. The surface question is "how do I add a buy button in ChatGPT?" The structural question is "when intelligence becomes the interface between a shopper and the entire internet of products, where does value move?" The answer is that value moves to whoever the intelligence trusts enough to recommend. A checkout button is a convenience at the very end of that chain. The recommendation is the whole game, because a recommendation from an assistant a person already relies on is worth more than any amount of ad inventory. This is why "selling inside ChatGPT" splits cleanly into two very different activities that most people blur together.

The first activity is discovery: getting ChatGPT to surface, describe, and recommend your product when someone asks what to buy. This is free, it is enormous, and it is where the durable opportunity lives. The second activity is transaction: letting the purchase itself complete inside the chat window through Instant Checkout. This is optional, it carries a fee, and in 2026 it is the smaller and more contested half. Conflating the two is the single most common mistake founders make, because the tactics, the economics, and even the odds of success are entirely different for each. We spend most of this guide on discovery for exactly this reason, and we treat checkout as a decision you make on top of a working discovery strategy, not as the strategy itself.

To understand why the split matters so much, you have to know what actually happened. OpenAI launched Instant Checkout on September 29, 2025, letting US shoppers buy from Etsy sellers directly in the chat, and promised that "over 1 million" Shopify merchants including Glossier, SKIMS, Spanx, and Vuori would follow - Stripe. The market reacted as if the future had arrived. Then reality set in. By February 2026, roughly 30 Shopify merchants were actually live, Walmart found in-chat purchases converted at one-third the rate of a click-out, and Etsy admitted it "did not end up seeing a large volume of sales" even as it found ChatGPT valuable for discovery - Search Engine Land.

On about March 24, 2026, OpenAI did something telling. Roughly six months after launch, it revamped the shopping experience, letting merchants use their own checkout while it refocused on product discovery, with OpenAI stating plainly that "the initial version of Instant Checkout did not offer the level of flexibility that we aspire to provide" - CNBC. Walmart had already pulled out and replaced the integration by embedding its own Sparky assistant as an app inside ChatGPT - Grocery Dive. The lesson is not that in-chat checkout failed. It is that the discovery layer proved far more valuable than the checkout layer, so the platform reorganized around it. A founder reading the news as "ChatGPT commerce flopped" learned the wrong lesson. The right one is that ChatGPT became a recommendation engine at planetary scale, and the retailers who understood that (Target, Sephora, Nordstrom, Lowe's, Best Buy, Home Depot, Wayfair all integrated for discovery) are the ones positioned to win - Retail Dive.

There is a third, less-discussed activity worth naming now because it reframes the whole opportunity: building an app inside ChatGPT. At DevDay on October 6, 2025, OpenAI opened ChatGPT to developers with the Apps SDK, letting businesses run interactive experiences (a booking flow, a product configurator, an ordering interface) directly in the conversation - VentureBeat. So the honest, first-principles model of "selling inside ChatGPT" in 2026 is three layers stacked on one foundation, and the foundation is always discovery.

The rest of this guide walks each layer in order of leverage. Why this matters is that your time is finite and the platform's own behavior tells you where to spend it: on being findable and recommendable first, on transaction convenience second. How to apply it starts with a mindset shift. Stop asking "how do I sell in ChatGPT" and start asking "when a purchase-ready person describes their problem to an AI, is my product the answer it gives?" Everything downstream, feed quality, reviews, checkout, apps, is in service of that one question.

2. Inside the numbers: how much AI shopping is really happening

Before you invest a single hour, you should size the prize honestly, including the parts that argue against the hype. The case for paying attention starts with raw scale. ChatGPT's weekly active users climbed from about 100 million in late 2023 to 300 million by December 2024, 700 million by August 2025, and 900 million by late February 2026, with the mobile app crossing a billion monthly users in June 2026, the fastest app in history to that mark - TechCrunch. No consumer surface has ever grown attention this fast, which is why every payments and retail giant scrambled to plug in within months.

That scale only matters to a merchant if the users are shopping, and increasingly they are. OpenAI's own usage research shows the share of messages classified as "seeking information" climbing as people offload research to the assistant, and roughly 50 million shopping-related queries a day now flow through ChatGPT, about 2% of its total volume - Dataslayer. Two percent sounds small until you multiply it by 2.5 billion daily prompts. The absolute number is tens of millions of purchase-adjacent conversations every single day, each one a moment where a product either gets recommended or does not exist.

The second half of the argument is intent quality, and this is where AI shopping stops looking like a novelty. An analysis of more than 50 million ChatGPT prompts by Profound classified 9.5% as commercial and 6.1% as transactional, versus 14.5% and 0.6% for traditional search - Profound. Read that carefully: transactional intent is roughly nine times higher in ChatGPT than in a Google search box. People do not just browse in the assistant, they arrive with a decision to make. A separate study from Measure Protocol found more than one in five ChatGPT conversations show some commercial intent, with 7.1% showing strong purchase signals - BusinessWire.

The traffic this produces is exploding off a small base. Adobe Analytics, watching over a trillion visits to US retail sites, reported AI-referred traffic up 693% year over year during the 2025 holiday season and up 138% year over year in May 2026, which is more than a 1,300% increase since Adobe began tracking in October 2024 - Digital Commerce 360. The deceleration in the growth rate is not a warning sign, it is what maturation looks like: the base is now large enough that triple-digit growth is a harder feat. The direction is unmistakable, and it points at your own website, because most of that AI-referred traffic clicks through to buy on the merchant's site rather than in the chat.

Now the honest counter-narrative, because a guide that only sells the upside is a brochure. The quality of AI-referred traffic is genuinely contested. Adobe reports that AI visitors convert 31% to 54% better than non-AI visitors, spend more time on site, and view more pages - Digital Transactions. But a working paper analyzing 973 e-commerce sites and 20 billion dollars of combined revenue found the opposite: ChatGPT referrals were only about 0.2% of sessions and actually converted worse than organic search, beating only paid social - Search Engine Land. Both can be true. AI referral is a high-quality but still tiny channel whose measured conversion swings wildly with the type of product and the way the study defines a session. The correct posture is to treat AI discovery as a fast-growing option worth positioning for, not a channel that will replace your existing traffic this quarter.

Finally, the market-size forecasts. Salesforce estimated AI agents influenced 67 billion dollars of 2025 Cyber Week sales and about 20% of the full holiday season, worth roughly 262 billion dollars - CX Today. Looking to 2030, the forecasts diverge wildly: Juniper sees 1.5 trillion dollars globally, Bain sees 300 to 500 billion in the US, McKinsey models up to 1 trillion in the US and 3 to 5 trillion globally, while eMarketer's conservative figure is around 144 billion because it only counts purchases completed inside the AI platform - Bain. That 35x spread is not analyst incompetence, it is the definitional fog of a new category: the numbers depend entirely on whether "agentic commerce" means the checkout happens in the chat or merely that an AI influenced the purchase somewhere along the way.

It helps to understand why OpenAI wants this business at all, because its incentives predict its moves. OpenAI reached a 500 billion dollar valuation in an October 2025 secondary sale, becoming the most valuable private company in the world - CNBC. Its annualized revenue run rate then topped 40 billion dollars by August 2026, roughly double where it ended 2025 - Bloomberg via Yahoo Finance. Subscriptions alone cannot carry those numbers, and inference is expensive, so OpenAI needs revenue that scales with usage. Commerce is the obvious candidate: a small fee on transactions plus, increasingly, advertising. That is the structural reason to expect the shopping surface to grow more commercial over time, and the reason to get established in the organic, merit-based version of it now, while it is still free and before the economics tilt toward pay-to-play.

Why this matters is that the numbers justify effort but not panic. The prize is real, the intent is unusually high, and the traffic is compounding, yet the channel is still a rounding error in your total sales today and its conversion quality is unproven at your specific price point. How to apply it is to size your investment to that reality: get positioned cheaply now, measure obsessively, and scale spend only when your own analytics (not Adobe's averages) show AI referrals converting.

3. Layer one: getting your products into ChatGPT's index

Everything starts with being present in the data ChatGPT reads, because an assistant cannot recommend a product it has never seen. There are two on-ramps, and which one you use depends almost entirely on where you already sell. The first on-ramp is automatic and covers most small merchants without any work at all. The second is a deliberate feed submission for everyone else. Understanding both, and knowing which applies to you, is the difference between appearing in millions of shopping conversations and being invisible in all of them.

If you sell on Shopify, you are likely already in. Shopify built a direct pipe: eligible US-selling stores are synced into ChatGPT through Shopify Catalog, and the company states bluntly that "you don't need to do anything for your products to be discovered by ChatGPT" - Shopify Help Center. The eligibility bar is basic hygiene rather than a high wall: sell to US customers, keep products eligible for Shopify Catalog, publish complete terms of service, privacy, and return policies, and accept the agentic-storefront terms. Etsy sellers are ingested the same way. This means the single fastest path to appearing in ChatGPT is not a special integration at all, it is running a clean, compliant store on a platform that already feeds the assistant. For a deeper look at the connective tissue between your store and these surfaces, see our guide to the top integrations for your online business.

If you are not on Shopify or Etsy, you submit a product feed directly through OpenAI's merchant program, and this is where the mechanics get specific. You apply at the ChatGPT merchant portal, pass verification of your business identity and policies, and then provide a structured feed that OpenAI ingests as its source of truth for your products - OpenAI Developers. The feed is a plain file (tab-delimited text or CSV are both accepted, gzip supported), and critically it can refresh as often as every 15 minutes, versus roughly 24 hours for a traditional Google Shopping feed - OpenAI Developers. That refresh cadence is a real advantage: your prices and stock levels can stay accurate to the quarter-hour, which matters enormously when an assistant is quoting them to a buyer in real time.

The required fields are worth knowing precisely, because a feed that is missing them simply will not surface. Getting these right is table stakes, not optimization. The mandatory fields the spec requires are the ones that let ChatGPT display an accurate, trustworthy product card.

• item_id and title - a unique ID per variant and a title up to 150 characters
• description and brand - plain-text description up to 5,000 characters, brand up to 70
• price and currency - the amount with an ISO currency code
• url and image_url - the product page and a usable image
• availability and eligibility flags - in stock or out of stock, plus is_eligible_search and is_eligible_checkout

Those flags deserve a note, because they are your control switch. is_eligible_search governs whether a product appears in ChatGPT's discovery at all, and is_eligible_checkout governs whether it can be bought in-chat - OpenAI Developers. You can be discoverable without enabling in-chat purchase, which is exactly the posture many merchants now prefer after the checkout wobble. Beyond the required fields, the spec accepts a long list of optional attributes that feed the ranking and trust signals: star_rating, review_count, actual review text, popularity_score, return_rate, GTINs, variant dictionaries, additional images, and even 3D models. OpenAI's own documentation is explicit that these extras "improve ranking, relevance, and user trust," which tells you the model rewards richer data - OpenAI Developers.

There is a second, quieter path into ChatGPT that has nothing to do with feeds: the open web. ChatGPT's search capability reads structured data from your product pages the way a search engine does, so Schema.org markup (Product, Offer, AggregateRating) on your own site is a discovery signal even without a formal feed - Search Engine Land. This is governed by crawler access. OpenAI runs distinct bots you can control in robots.txt: OAI-SearchBot surfaces your site in ChatGPT search and citations, GPTBot gathers training data, and ChatGPT-User fetches pages on a user's behalf - OpenAI Developers. The practical trap here is subtle and common: a firewall or rate-limit rule that returns "429 Too Many Requests" to these crawlers gets your site quietly classified as unreliable and dropped, so blocking is not always intentional. We cover the full crawler-access checklist in our guide to the top technical SEO skills for 2026.

Why this matters is that presence is binary and cheap: you are either in the index or you are not, and getting in costs little more than good store hygiene plus a clean feed. How to apply it is a fifteen-minute audit. Confirm your Shopify or Etsy store is eligible and synced, or apply to the merchant program with a complete feed, then check that OAI-SearchBot is not being blocked or throttled by your server. Do this before you spend a minute on anything more advanced, because none of the later tactics work if you are not in the data.

4. Getting recommended: what makes ChatGPT pick you

Being in the index makes you eligible. Being recommended is a different and harder thing, and it is where most of the durable advantage sits. This is the discipline that has picked up the awkward name generative engine optimization, and the good news for founders is that the tactics that actually work are largely the same brand-building fundamentals that have always worked, just pointed at a new reader. The bad news is that several of the tactics being sold as GEO silver bullets have no evidence behind them, and telling the two apart saves you months. Approach this section as a skeptic, because the field is young and full of confident claims.

Start with what the data actually supports. The strongest study on the question, an Ahrefs analysis of 75,000 brands published in December 2025, found that the factors correlating most with ChatGPT visibility are not on-page technical tweaks at all. They are off-site brand signals: YouTube mentions (0.737 correlation), branded web mentions (0.664), and branded anchor text (0.511), all far ahead of raw backlinks at around 0.2 - Ahrefs. Ahrefs is careful to note that correlation is not causation, but the pattern is consistent and it rhymes with how the models work: an assistant recommends what it has seen discussed, reviewed, and referenced across the web, not what has the cleverest meta tags. The implication is that getting people to talk about your product, which we cover in depth in our guide on how to get people to talk about your product, is a GEO tactic even though it looks like plain marketing.

Where ChatGPT gets its information matters as much as how much of it exists, and the concentration is striking. Research by 5W found that Wikipedia and Reddit together drive more than 25% of all US ChatGPT citations, with no other single domain exceeding 3% - PR Newswire. A larger Semrush study of 100 million citations found the picture shifts over time as OpenAI tunes its sources, with sites like Forbes, Medium, and PR distribution rising after a mid-2025 change - Semrush. The practical reading is not "spam Reddit," which backfires, but "make sure your product is genuinely, positively present in the places the model trusts": a real Reddit footprint, an accurate Wikipedia presence where warranted, and credible third-party reviews.

The academic backbone for content tactics comes from the original GEO paper out of Princeton and IIT-Delhi, presented at SIGKDD 2024, which tested what actually raises visibility inside generative answers. Adding relevant statistics, credible quotations, and citations to reliable sources lifted visibility by up to roughly 40% in their experiments - arXiv. This is why a product page or comparison article that states concrete numbers and cites its sources outperforms one full of vague superlatives: the model is looking for verifiable, quotable substance, and it rewards content that reads like evidence rather than advertising. It is the same reason this guide is built from cited data rather than adjectives.

Freshness is the lever most merchants underuse, and it is unusually powerful for ChatGPT specifically. Analysis by ConvertMate found that 76.4% of ChatGPT citations come from content updated within the last 30 days, the strongest recency bias of any AI engine, a finding Ahrefs corroborated across 17 million citations - Contently. This has a direct operational consequence: a product page or buying guide that was last touched a year ago is at a structural disadvantage to one refreshed last week, regardless of quality. Keeping your key pages, prices, reviews, and comparison content genuinely current is one of the highest-leverage and most-neglected things you can do, and it happens to be exactly the kind of relentless, low-glamour upkeep that an always-on operation handles better than a busy founder.

One more finding should shape your expectations about how often you can even appear. A Profound study of roughly two million prompts found that only about 9% of prompts trigger ChatGPT's shopping module at all, while 79% never trigger it, and the pattern is close to all-or-nothing: a given query type either reliably shows product cards or almost never does - Profound. The practical consequence is that your category matters as much as your optimization. If shoppers in your space phrase their needs in ways that reliably trigger the shopping surface (open-ended "best X for Y" questions tend to, narrow brand-name lookups often do not), discovery is a live channel for you. If they do not, you are competing to be mentioned in plain prose rather than shown in a card, a different and harder game that leans entirely on brand reputation. Knowing which bucket your category falls into tells you whether to pour effort into feed richness or into the brand-mention work that gets you named in the answer itself.

Now the myths, stated plainly so you do not waste effort. Two widely-promoted tactics have weak or no evidence for editorial citations. A controlled Ahrefs test of 1,885 pages that added JSON-LD schema found no statistically significant lift in AI citations, roughly +2% and indistinguishable from zero - Stan Ventures. And llms.txt, the proposed file for telling LLMs what your site is about, sits at only about 10% adoption with no major provider confirming they use it - SE Ranking. The crucial nuance: schema is still required for shopping-feed eligibility (the model needs your Product data to show a card) and helps rich results elsewhere, so keep it. Just do not expect it to make ChatGPT quote your blog. The distinction between "table-stakes structured data" and "citation magic" is the kind of thing our companion guide on how to get your site cited by ChatGPT and Claude unpacks in full.

Why this matters is that the recommendation layer is where competitors cannot simply copy you overnight, because brand reputation and third-party presence compound slowly and defensibly. How to apply it is to prioritize by evidence: invest first in genuine brand mentions and reviews across YouTube, Reddit, and comparison sites, keep your key pages fresh on a real cadence, write product content thick with specifics and sources, and treat schema and llms.txt as hygiene rather than growth levers. The founders who win discovery are the ones building a real reputation the model can find, not the ones chasing the tactic of the week.

5. Layer two: in-chat checkout and the Agentic Commerce Protocol

With discovery working, you can decide whether to add in-chat purchase on top. This is the layer that got the headlines and the layer that stumbled, so approach it as a deliberate business decision rather than a default. The mechanism is genuinely elegant, the economics are real, and the fit is narrow. Understanding all three lets you make the call for your own catalog instead of following a trend that already reversed once.

The technical foundation is the Agentic Commerce Protocol (ACP), an open standard co-developed by OpenAI and Stripe and released under the Apache 2.0 license, which means any platform or payment provider can implement it - Stripe. The design goal was to let a shopper complete a purchase inside the chat while the merchant stays the merchant of record, keeping control of pricing, the customer relationship, tax, fulfillment, and returns, and keeping the right to accept or decline each order. This is the opposite of a marketplace that owns your customer. You are not handing the relationship to OpenAI, you are letting its interface trigger a sale that is fully yours. That distinction is the entire reason serious merchants were willing to try it.

The payment magic is the Shared Payment Token (SPT), a Stripe primitive that lets an application like ChatGPT initiate a charge without ever seeing the buyer's card details. The token is scoped to a specific merchant and a specific amount, limited by time, revocable at any moment, and monitored via webhooks - Stripe. Mechanically, the shopper confirms in the chat, the agent issues an SPT and passes it to the seller, and the seller redeems it by creating a normal Stripe PaymentIntent, complete with Stripe Radar fraud protection. Stripe extended SPT support beyond its own rails to Visa Intelligent Commerce, Mastercard Agent Pay, and buy-now-pay-later methods like Affirm and Klarna, and importantly the payment layer works across the US, Canada, and a list of European countries even though the ChatGPT buying experience itself launched US-only.

Then there is the money. Shoppers pay nothing extra, but merchants pay OpenAI a fee on completed in-chat purchases. The rate is not officially published, but according to reporting on the confidential contracts, Shopify merchants pay a 4% fee on ChatGPT-completed sales, on top of their normal payment processing of roughly 2.9% plus 30 cents, with a reported 30-day trial window; Etsy sellers were not charged at launch - PYMNTS. That 4% is the crux of the decision. For a high-margin impulse buy, paying 4% to capture a sale that would otherwise never happen is obviously worth it. For a low-margin or considered purchase, 4% on top of processing can erase the profit, especially when the same shopper would happily click through to your own site and cost you nothing. This is precisely the calculation our guide to the best payment platforms for your business walks through in detail.

Now the honest verdict on fit, because the data already told us how this played out. In-chat checkout underperformed for most merchants: Walmart measured conversion at one-third of a click-out and called the experience unsatisfying, largely because shoppers did not want to split a single-item purchase away from the rest of their cart - Search Engine Land. The single-item limitation was a real constraint, and it maps to where in-chat checkout does work: impulse, gift, and single-SKU purchases where the friction of leaving the chat genuinely costs the sale. It works poorly for multi-item baskets, considered purchases, and anything where the shopper wants to compare or configure. Since OpenAI's March 2026 pivot, you can also simply let the shopper click through to your own checkout while still being fully present in discovery, which for many catalogs is the better trade.

Why this matters is that in-chat checkout is a tool with a narrow, real use case, not a mandate, and treating it as optional rather than essential is what the platform's own behavior now recommends. How to apply it is to run the margin math per product line: enable in-chat checkout on high-margin, impulse-friendly items where friction kills sales, keep everything else on discovery-plus-click-through, and revisit as OpenAI expands multi-item carts and international support. Do not enable it everywhere by default, and do not skip it out of caution on the products where it fits.

6. Layer three: building an app inside ChatGPT

The third way to sell inside ChatGPT is the most ambitious and the least understood: you can build an actual app that runs inside the conversation. This is not a chatbot bolted onto your site and it is not a listing. It is an interactive experience (a booking flow, a product configurator, an ordering interface, a lead-capture tool) that appears and functions directly in the chat when a user's request calls for it. For the right kind of business, this turns ChatGPT from a place that mentions you into a place that transacts with you on your own terms. It is early, it is gated, and it is worth understanding now because the direction of travel is clear.

The foundation is the Apps SDK, which OpenAI unveiled at DevDay on October 6, 2025, with Sam Altman framing it directly: "Today, we're going to open up ChatGPT for developers to build real apps inside of ChatGPT" - VentureBeat. The launch partners signaled the ambition: Booking.com, Canva, Coursera, Expedia, Figma, Spotify, and Zillow shipped first, with commerce-heavy names like Target, Instacart, DoorDash, and Uber following. When a user asks a relevant question, the app can surface inside the chat, render a real interface, and take action, which is why the press immediately called it ChatGPT becoming "the new app store."

The technical detail that matters for founders is that the Apps SDK is built on the Model Context Protocol (MCP), the same open standard for connecting AI models to external tools and data that the rest of the industry has converged on - OpenAI. This is genuinely good news, because it means the work you do to make your product accessible to ChatGPT is not locked to OpenAI. An MCP server that exposes your catalog, your booking system, or your configurator can be reused across every assistant that speaks MCP, which is increasingly all of them. If you are weighing this path, our guide to shipping an MCP server for your product is the practical companion, and if you would rather assemble the experience without deep engineering, our roundup of the top AI app builders covers the tooling.

Crucially, this is no longer a closed launch club. On December 18, 2025, OpenAI opened public app submissions and launched a browsable App Directory at chatgpt.com/apps, so any developer can submit an app for review and publication rather than waiting for an invitation - VentureBeat. That is the moment the door opened for ordinary founders. The submission itself is mostly paperwork: app name, logo, description, privacy-policy and company URLs, your MCP and tool details, and test prompts, all reviewed against OpenAI's developer guidelines - OpenAI Help Center. It is worth being precise about how this differs from the older GPT Store: a Custom GPT was a prompt-based customization with no backend, while an Apps SDK app is a genuine third-party service that takes authenticated, product-like actions on your own infrastructure. The former was a clever prompt, the latter is software you run.

What an app can actually do for selling is broader than checkout. Because it renders a real interface, an app can handle the interactions a static product card cannot: configure a complex product, book a service against live availability, build and price a custom order, capture a qualified lead into your CRM, or run the kind of guided conversation that a knowledgeable salesperson would. This is the same instinct behind putting a smart assistant on your own website, which we cover in build a support agent for your site, except the audience is ChatGPT's hundreds of millions of users rather than only the visitors who already found you. For businesses whose sale is too involved for a one-tap buy button, this is potentially the most valuable of the three layers.

Selling through an app is becoming concrete rather than theoretical, and OpenAI now documents the options explicitly. There are three monetization paths for physical goods, and the one you choose decides how much of the experience you keep control of.

• External checkout - send the buyer to your own checkout (recommended, generally available)
• Embedded checkout - an in-chat payment sheet via the requestCheckout method (beta, select marketplaces)
• Saved payment methods - checkout using the shopper's stored details, across processors like Stripe, PayPal, Adyen, and Worldpay

The commerce apps are already live and instructive. Instacart shipped its ChatGPT app on December 12, 2025 as the first to support end-to-end shopping with Instant Checkout inside the chat, DoorDash followed with recipe-to-grocery ordering, and Target ran an in-app shopping beta over the holidays - Digital Commerce 360. This is also where OpenAI's March 2026 pivot actually pointed: rather than killing in-chat purchase, it moved Instant Checkout into individual merchant apps, where a business controls the flow and can handle the messy realities (sales-tax remittance, live inventory) that a generic buy button could not. For a services or configurable-product business, an app plus embedded checkout is the closest thing to running your real storefront inside someone else's assistant.

The limits are equally real and you should size them honestly. Discovery of apps is OpenAI-controlled through conversational relevance rather than a search box, the surface is gated by review and approval, and availability is uneven: apps rolled out to logged-in users outside the EU, EEA, Switzerland, and the UK, while in-chat checkout remains US-only - Storyboard18. This is not a channel you can flip on this afternoon the way you can submit a feed. It is a build, with a review process and a dependence on OpenAI's discovery decisions. For most founders, the correct sequencing is to nail discovery and, where it fits, checkout first, then evaluate an app once you have evidence that ChatGPT sends you meaningful, purchase-ready traffic.

Why this matters is that apps are how ChatGPT moves from recommending your product to hosting your actual sales motion, which is a categorically bigger opportunity for businesses with complex or service-based offerings. How to apply it is to treat an app as a considered investment, not a quick win: confirm the demand with your discovery data first, build on MCP so the work travels across assistants, and design the app around the one interaction that a flat product card cannot do for your business.

7. The competitive map: where else AI shoppers can find you

ChatGPT is the largest AI shopping surface but it is not the only one, and a founder who optimizes for it alone is leaving the majority of AI-driven shoppers on the table. The critical strategic fact, and the reason the assessment table at the top of this guide is ordered the way it is, is that these surfaces split into open ecosystems you can join for free and one walled garden you largely cannot. Where you invest should follow that split, not the raw size of each platform's audience. This section unpacks each row of the scorecard so you can decide where the second and third hours of your effort go.

Google is the highest-scoring surface for a simple structural reason: it sits on top of the world's biggest shopping index, the Shopping Graph with more than 50 billion product listings, 2 billion of them refreshed every hour, and it lets you into that index for free through Merchant Center listings - Google. Google's AI Mode reads those organic listings, and in November 2025 it added agentic "buy for me" checkout with Google Pay. In January 2026 it went further, announcing the Universal Commerce Protocol (UCP), an open standard co-developed with Shopify, Etsy, Target, and others and endorsed by more than 20 payment and retail companies - Google. For most merchants, enabling free listings in Merchant Center is the single highest-reach action available, and it costs nothing.

Perplexity scores third overall but first on the criterion founders care about most, openness. Its checkout, powered by the startup Firmly.ai, lets merchants join with a single API and no fees, keeping 100% of revenue and merchant-of-record status - Digital Commerce 360. It layered on PayPal Instant Buy in November 2025 and reported a fivefold increase in shopping-intent queries since launching its Buy with Pro feature - PayPal. Perplexity's audience is much smaller than ChatGPT's or Google's, which caps its score, but for a merchant deciding where to spend a limited integration budget, "smaller audience, zero fees, keep the customer" is often a better deal than "bigger audience, 4% fee." It is the merchant-friendliest surface in the market.

Microsoft Copilot entered the race in January 2026 with Copilot Checkout, auto-enrolling Shopify merchants and partnering with PayPal, Shopify, and Stripe, plus a Brand Agents product that lets merchants deploy AI shopping assistants on their own sites - GeekWire. The mechanics are sound and merchant-friendly, but Copilot holds only single-digit share of AI-chatbot traffic against ChatGPT's roughly 68%, so the audience is the constraint. It scores as a "set it up if you are already on Shopify, do not build for it specifically" surface: the enrollment is nearly free, so there is little reason to opt out, but it should not command dedicated effort until its shopping audience grows.

Amazon is the instructive opposite, and its low score is the whole lesson of the table. Amazon built its own Rufus assistant and a Buy for Me agent that can purchase from external brand sites, but it runs a closed system - Forbes. In November 2025 it updated its robots.txt to block OpenAI's crawlers, and it won (then later lost on appeal) an injunction against Perplexity's shopping agent, defending a business that earns it roughly 56 billion dollars a year in advertising - Modern Retail. For an independent brand, the takeaway is stark: inside Amazon you can only sell as an Amazon seller, on Amazon's terms and fees, and you cannot get your own site discovered through Amazon's AI at all. Bigness is not openness, and for a founder the open ecosystems are worth more than the closed giant.

Underneath all of these run the payment rails that make agentic commerce trustworthy, and they are worth knowing because they will outlast any single assistant. Visa Intelligent Commerce and Mastercard Agent Pay, both launched in 2025, add cryptographic verification so a merchant can tell a legitimate shopping agent from a malicious bot, with Visa reporting hundreds of completed agent-initiated transactions and more than 100 ecosystem partners by December 2025 - Visa. Nearly half of US shoppers already use AI for shopping tasks, per Visa's own research, which is why the networks are racing to build the trust layer.

Why this matters is that a multi-surface presence is nearly free to establish and dramatically widens the top of your funnel, while betting everything on one assistant exposes you to that assistant's next pivot. How to apply it is to establish free presence broadly and deep effort narrowly: turn on Google Merchant Center free listings, accept the auto-enrollments from Shopify into Copilot and others, join Perplexity's no-fee program, and reserve your real optimization energy for the two surfaces that send you measurable traffic. The same product feed and brand-signal work pays off across all of them at once, because their citation patterns overlap heavily.

8. The practical playbook: what to do, in what order

Strategy is worthless without sequence, and the most common way founders waste effort here is doing the advanced things before the basic ones. The order below is deliberately ranked by leverage per hour, front-loading the cheap, high-impact moves and deferring the expensive, speculative ones. Work it top to bottom, and do not skip ahead to an app or a checkout integration before the foundation underneath it exists, because the later steps quite literally depend on the earlier ones.

The foundation is presence and accuracy, and it is where your first day goes. Before any optimization, make sure the data ChatGPT reads about you exists and is correct, because everything else is built on it. The opening moves are unglamorous and high-return.

1. Confirm you are in the index - verify your Shopify or Etsy sync, or apply to OpenAI's merchant program with a complete feed
2. Unblock the crawlers - make sure OAI-SearchBot is allowed and not being throttled to a 429 error
3. Fix your product data - accurate titles, prices, availability, and images in your feed and in on-page schema
4. Publish complete policies - shipping, returns, and privacy pages, which are eligibility requirements, not nice-to-haves
5. Turn on free listings everywhere - Google Merchant Center free listings and any no-cost auto-enrollments

Those five steps cost almost nothing and put you in front of tens of millions of shopping conversations across multiple assistants. The reason they come first is that they are binary and cheap: each one either is done or is not, none requires ongoing creativity, and skipping any of them silently caps everything downstream. A brilliant GEO campaign pointed at a store that blocks OAI-SearchBot produces exactly zero results, which is why the boring foundation outranks the exciting tactics. Treat this as a one-time setup you verify quarterly.

With the foundation in place, the second phase is the recommendation work from section 4, and this is ongoing rather than one-time. Here you are building the brand signals and freshness that make ChatGPT choose you over an equally-eligible competitor. The highest-leverage moves are earning genuine third-party presence (reviews, YouTube mentions, credible Reddit discussion, comparison-article inclusion), keeping your key pages genuinely fresh given ChatGPT's extreme recency bias, and writing product and buying-guide content thick with the specifics and citations the models reward. This is real, continuous marketing work, and it is where the durable advantage compounds. It is also, candidly, the part founders struggle to sustain, because it never ends and rarely feels urgent, which is exactly why so much of it goes undone.

That sustainability problem is where the operating model matters more than any single tactic. Keeping a feed accurate to the quarter-hour, refreshing dozens of pages on a cadence, monitoring which assistants cite you, seeding and maintaining a real third-party presence, and doing it every week is genuinely more work than one founder has hours for. This is the same realization behind the shift to running a business as an always-on operation rather than a series of manual sprints, a theme we explore in the autonomous business guide and hire an AI workforce to run your company. Platforms like Founden take the most literal version of that idea: you describe the business and an AI workforce runs the recurring operation, which is one way (among several) to make the relentless upkeep of AI-channel presence actually happen instead of slipping. For the broader operational picture, our guide to automating your startup back office and the AI-native company tech stack map the full surface.

The third phase is the optional layers, and it comes last because it should be evidence-driven. Only after discovery is working and you can see AI referrals in your analytics should you decide about in-chat checkout and apps. Enable Instant Checkout on the specific high-margin, impulse-friendly products where the 4% fee buys a sale you would otherwise lose, and leave everything else on click-through. Consider building an app only if your discovery data shows meaningful purchase-ready traffic and your sale is too complex for a product card, in which case build it on MCP so the work travels. The discipline here is refusing to build the expensive thing until the cheap thing has proven the demand, because an app or a checkout integration built on a discovery channel that sends you no traffic is effort spent on a foundation that is not there.

Why this matters is that the same finite effort produces wildly different returns depending on order, and the natural temptation (to start with the shiny checkout or app) is exactly backwards. How to apply it is to run the phases in sequence, verify each before advancing, and let your own analytics rather than the industry's hype decide when to graduate from free discovery to paid transaction to a full app.

9. Failure modes, limits, and honest risks

A guide that only lists opportunities is doing you a disservice, because the ways this can go wrong are specific and avoidable if you see them coming. The risks fall into three buckets: platform risk you do not control, execution risk you do, and economic risk that decides whether any of it is worth it. Walking each honestly is what separates a durable strategy from a fashionable one, and it is the difference between building on this channel and being burned by it.

The first and largest risk is platform dependency, and the last twelve months are the proof. A merchant who rebuilt their entire funnel around Instant Checkout in October 2025 was, by March 2026, holding an integration OpenAI had substantially repositioned. The platform changed the rules in six months, and it will change them again. Ads are already arriving: OpenAI began testing advertising in ChatGPT's free and lower-priced tiers in January 2026, reversing Sam Altman's earlier stance that ads would be a "last resort," and one study found ChatGPT now shows ads on about 26% of commercial prompts - Search Engine Land. The organic discovery that is free and merit-based today may become a pay-to-play surface tomorrow, exactly as search did. The defense is to treat every AI assistant as a rented channel, never your only one, and to use it to build the one asset you own outright: a direct relationship with the customer, captured on your own site.

The second risk is execution and data quality, which is entirely in your control and quietly fatal when neglected. Walmart's checkout underperformed partly because its feed of 200,000 products had frequently inaccurate inventory and shipping data, which the assistant then quoted to shoppers - CNBC. An AI that confidently recommends a product you cannot actually ship at the price it quoted does not just lose that sale, it erodes the trust that made the recommendation valuable. The failure modes here are mundane and preventable: stale prices, wrong stock levels, blocked crawlers, missing policy pages, and thin product data that gives the model nothing to work with. Because ChatGPT prizes freshness and accuracy so heavily, sloppy data is not a minor hygiene issue, it is a ranking and trust penalty.

The third risk is economic, and it is where honest math protects you. The 4% Instant Checkout fee, stacked on payment processing, can quietly turn a thin-margin sale into a loss, which is why blanket-enabling checkout is a mistake. The conversion data is genuinely mixed, with credible studies showing AI referrals converting both far better and somewhat worse than existing channels depending on the product and methodology, so you cannot assume the traffic will convert at your current rates. And the channel is still small in absolute terms, a rounding error in most merchants' total sales today, which means it deserves proportionate rather than all-consuming investment. Pricing your product to absorb new channel fees without eroding margin is its own discipline, one we cover in price your AI product to beat token costs. The correct economic posture is to position cheaply, measure with your own numbers, and scale spend only where the unit economics actually work.

Why this matters is that the founders who lose money here are not the ones who ignored AI commerce, they are the ones who over-committed to it, blanket-enabled fees, neglected their data, and treated a volatile rented channel as a foundation. How to apply it is defensively: own the customer relationship on your own site, keep your data ruthlessly accurate, run the margin math per product, and size your total investment to a channel that is real and growing but not yet dominant. Skepticism is not the enemy of opportunity here, it is how you survive to capture it.

10. The future outlook: when your customer is an agent

Step back to the structural question one more time, because the direction of travel changes what you should build. For the entire history of commerce, the thing on the other side of the sale has been a human with attention, patience, and a limited ability to compare. That assumption is dissolving. The trajectory of 2025 and 2026, from Instant Checkout to ACP to Visa Intelligent Commerce to Mastercard Agent Pay, all points at the same destination: a world where the entity discovering, comparing, and increasingly buying your product is not a person browsing but an agent acting on a person's behalf. Bain projects that autonomous shopping agents will complete roughly a quarter of online transactions within five years - Bain. When your customer is an agent, the rules of selling change in ways that reward preparation now.

The first implication is that legibility to machines becomes a competitive advantage, on par with brand and price. An agent cannot be charmed by a beautiful hero image or nudged by a limited-time banner. It reads structured data, weighs verifiable signals, and compares on the attributes it can parse: real specs, real reviews, real availability, real return rates. The merchants who win the agent era are the ones whose products are maximally legible, richly described in machine-readable terms, backed by genuine third-party evidence, and always accurate. This is why the unglamorous feed-and-freshness work in this guide is not just a 2026 tactic, it is the foundation of how you will be evaluated when the buyer never sees your homepage at all.

The second implication is that the standards, not the platforms, are the durable bet. ChatGPT may pivot again, a new assistant may rise, ads may reshape organic discovery, but the open protocols underneath, ACP, MCP, UCP, and the network agent-payment rails, are converging into shared infrastructure that outlasts any single app. Stripe already reports its agent-payment protocol processing more than 165 million transactions with tens of thousands of active agents - Stripe. Building your commerce so it speaks these open standards, rather than hard-wiring to one assistant's proprietary integration, is how you stay portable across whatever surface wins. An MCP server, a clean ACP-compatible feed, and Merchant Center listings are assets that travel; a bespoke integration with a single platform is a liability waiting for that platform's next reversal.

The third implication is about operating cadence, and it is the one most founders underestimate. In a human-paced web, a store that updated monthly was fine. In an agent-paced web, where ChatGPT rewards content updated in the last 30 days and feeds can refresh every 15 minutes, staleness is a structural disadvantage that compounds daily. The businesses that thrive will run more like always-on operations than periodic campaigns, continuously keeping data fresh, reputation current, and presence consistent across every AI surface. This is the throughline connecting this guide to the broader shift we document across our library, from what software is left to build in 2026 to the rise of the solopreneur and how to start a company in 2026: the leverage now comes from systems that run themselves, not from working faster by hand.

This is the lens through which Yuma Heymans, founder of the AI-workforce platform O-mega and co-founder of the AI recruitment engine HeroHunt.ai, has been building: his whole thesis (@yumahey) is that the durable advantage of the agent era belongs to businesses run by tireless software rather than stretched-thin humans, which is exactly the muscle that keeping a product recommendable across every AI assistant demands. Selling inside ChatGPT is an early, concrete instance of a much larger shift, and the founders positioning for it now are learning the operating model that the next decade of commerce will require.

Why this matters is that the specific features in this guide will change, but the structural forces (agents as buyers, machine legibility as advantage, open standards as the durable layer, always-on operation as the winning cadence) will only intensify. How to apply it is to build for the direction, not the snapshot: make your product legible to machines, bet on open standards over proprietary integrations, and run your presence as a continuous operation. Do that, and whatever the assistants look like in 2028, your product will be the answer they give.

The bottom line for 2026 is simpler than the hype suggested and more durable than the backlash implied. "Selling inside ChatGPT" is really about being the product a trusted AI recommends to a purchase-ready human, and everything else, checkout, apps, fees, is a decision you make on top of that. Get into the index, earn the recommendation, let people buy wherever converts best, and treat the whole thing as a fast-growing rented channel that feeds the one asset you own: a direct relationship with a customer who found you because, when they asked an AI what to buy, the answer was you.

This guide reflects the AI commerce landscape as of August 2026. This is one of the fastest-moving areas in technology: features, fees, model versions, and platform strategies change monthly, and OpenAI has already repositioned its shopping experience once. Verify current details with the official sources linked throughout before making decisions.