On July 1, 2026, Square shipped a ChatGPT app and a Claude plugin. A customer can now describe what they want, get a Square-powered restaurant recommended, read the live menu with prices and modifiers, and pay, all inside the conversation. Orders route into the seller's existing Square Point of Sale and Kitchen Display System just like an in-store or website order, and Square tags them as an AI-channel source in reporting. Eligible US Food & Beverage sellers with an active Square Online Ordering profile were opted in automatically, no new APIs, no setup, no contract. The only cost is Square's standard online processing fee of 2.9%–3.3% + $0.30, against the 15%–30% delivery apps charge.
Three weeks earlier, on June 11, DoorDash launched Ask DoorDash, a conversational search feature that lets users search the app in their own words rather than scrolling through restaurants or grocers.
The demand side has been moving longer than either launch. DoorDash's 2026 Restaurant Industry Trends Report found that 22% of consumers have used an AI tool like ChatGPT or Google Gemini to help choose a restaurant, and — citing Yext research — that restaurant listing sites account for more than 41% of the sources AI tools reference when recommending restaurants. Only 39% of operators have updated their listings to show up in those results. Uberall's May 2026 benchmark puts a harder number on the gap: 83% of restaurant locations are entirely invisible in AI-generated recommendations, even though 86% of those same restaurants maintain some presence on Google. From the Doorstep to the Dining Room: New DoorDash Survey Data Reveals the Full Picture of the Modern Restaurant Guest | DoorDash +2
This is no longer a thought experiment about where discovery might go. One player has working infrastructure in market, and it went live for millions of merchants without any of them lifting a finger. For everyone else in the POS and foodtech stack, the question is no longer whether the channel is real. It's whether your customers' data is in a shape that can be read by it.
Agentic ordering is a multi-step task executed by an AI assistant on a guest's behalf: interpret an intent, shortlist venues, read the live menu, assemble a valid cart with modifiers, and complete checkout, without the guest ever opening a restaurant app or website.
The distinction that matters is not "chat instead of clicks." It's that the entire funnel collapses into a single exchange. When a consumer asks ChatGPT or Claude to find a nearby specialty coffee shop and order a bag of house roast, the AI parses real-time data provided by Square. There is no browsing session. No homepage, no hero photo, no upsell banner, no "chef's picks" carousel doing quiet persuasion work.
As one analysis of the launch put it, the AI effectively becomes the front of house: it frames the menu, handles the modifiers, and sets the tone of the transaction.
That reframing has a blunt consequence. Everything the agent knows about a restaurant comes from structured data. Nothing else survives the trip.
Four categories of information an agent cannot infer, and is not allowed to guess:
Modifiers and option sets. "Large oat latte, extra shot, no foam" has to resolve to a real item plus a real modifier tree with real prices and real validity rules. If your modifiers live in free-text item descriptions — "add bacon +$2" written into a sentence — the agent can read the words but cannot build a cart that a POS will accept. It will either drop the customization or drop the restaurant.
Allergens and dietary attributes. This is the one with legal exposure. An assistant that infers "probably gluten-free" from an item name is a liability event waiting for a plaintiff. Models are trained to hedge on exactly this kind of claim, which means the practical outcome of missing allergen data is not a wrong answer; it's the agent quietly routing around your restaurant toward one that has the field filled in. Allergens have to be structured fields with explicit contains / may-contain values. Prose does not qualify.
Hours and availability. Not just "open 9–9." Holiday exceptions, per-channel hours, and the gap between when the kitchen stops and when the room closes. Platforms build against live feeds precisely because an agent that promises a closed restaurant makes the platform look broken.
Stop lists. Square's integration draws live from each merchant's catalog, surfacing item descriptions, pricing, modifiers, and real-time stock availability so that AI agents cannot present out-of-stock items to customers. An 86'd item that clears at the end of shift instead of in real time is a canceled order in the agent channel.
The asymmetry is what operators underestimate. In the old channel, incomplete menu data cost you a conversion; the guest was already on your page and worked around the gap. In the agent channel, incomplete data costs you the mention. You aren't ranked low. You're not in the sentence.
A static menu isn't a starting point; it's a disqualification. Square's integration requires an active Square Online Ordering profile built around a structured digital menu; a restaurant with only a PDF menu isn't eligible, because without live structured data, there is nothing for the AI to surface or transact on.
Classic SEO is a competition for placement among ten links, where the human does the final choosing. Answer Engine Optimization and Generative Engine Optimization are a competition to be one of the two to five names the assistant says out loud. There is no page two, and there is no scrolling past a weak result to a stronger one.
Two things drive that shortlist.
Trust thresholds are quantifiable. Uberall's benchmark found ChatGPT primarily recommends businesses averaging around 4.3 stars, Perplexity 4.1 and above, and Gemini 3.9 and above — and noted that a restaurant with a 4.0 rating will still rank in Google search results while falling below the threshold an assistant will name. A merely fine rating is now a hard filter rather than a soft ranking factor.
Data consistency is a ranking signal in its own right. Models like ChatGPT and Gemini pull from roughly 8 to 10 sources per restaurant-related query, cross-referencing data across directories before deciding whether to recommend a business. Hours that say one thing on your site, another on Google Business Profile, and a third on a delivery listing don't average out. They read as unreliability. Schema markup — Restaurant, Menu, LocalBusiness, FAQPage — is the cheapest way to make your own site legible enough to be one of those cross-referenced sources rather than the odd one out.
Worth reading with a little skepticism: the 41% listing-platform figure comes from Yext research cited in DoorDash's own report, and DoorDash is a listing platform. The direction is credible, the precision less so. But the strategic implication holds either way, aggregator presence now feeds the AI channel even for operators who never wanted a delivery order in the first place.
Square helps more than 4.5 million sellers surface and transact across search, maps, social, and marketplaces, owns a payment rail in Order by Cash App, and has direct partnerships with OpenAI and Anthropic. No mid-market POS vendor or reseller can replicate that combination. The moat isn't the technology. It's the distribution and the ability to auto-enroll a merchant base overnight.
But "Square won, we're late" is the wrong read of what actually shipped.
What Square built is a connective layer. Square describes its role as handling the infrastructure — syncing business information, menu data, hours of availability, and ordering information in real time — so sellers gain visibility across new AI-powered channels as they emerge without managing each integration individually. That layer is buildable at any scale. What Square has that you don't is the number of merchants plugged into it, not a secret technique.
More importantly, the interfaces are being standardized in the open right now. Square says it is participating in the AAIF Agentic Commerce Working Group and the W3C Web Payments Working Group, and is partnering with Google to co-develop the Universal Commerce Protocol spec for local food ordering and delivery, an open standard covering the full commerce journey so agents and systems can work together. It has also named Alexa+ as the next surface, via a partnership with Amazon.
An open protocol changes the strategic question. You do not need a bilateral deal with OpenAI. You need a menu data model clean enough to serve the protocol when it lands; and enough merchants whose menus are actually in that shape.
Three practical positions worth taking now:
Treat data readiness as product, not hygiene. Normalized menus, real modifier trees, structured allergens, live availability. If a merchant's data is clean, you can expose it to any surface. If it isn't, no integration rescues them.
Treat integration breadth as insurance. Nobody knows whether the winning surface is ChatGPT, an assistant inside a delivery app, Gemini in Search, or a voice device. Every bet on one is a bet against the other four.
Treat the non-Square base as the addressable gap. Square auto-enrolled its merchants, including ones whose menus aren't ready for the exposure. Everyone else — independents, multi-unit brands, and operators on other POS systems — is now reading the same coverage and asking their vendor what to do. That question is landing in reseller inboxes this quarter.
The risk of doing nothing is specific: the menu data layer becomes a commodity someone else owns and monetizes, and your POS is reduced to an order sink at the end of somebody else's pipe.
Three problems arrive together, and none of them has a settled answer yet.
Attribution. An agent order arrives with no session, no referrer, no campaign parameter, no landing page. Square makes the order source visible in reporting so operators can see how the new channels perform. That is the floor, and it's worth treating as a hard requirement of any integration you build or resell: channel-level source tagging on every order, or you're flying blind on a channel you're being asked to invest in.
The guest record. If checkout completes inside the chat through a wallet, the restaurant may receive everything needed for fulfillment and nothing that constitutes a relationship. The agent, meanwhile, retains the preferences: usual order, dietary constraints, spice tolerance, the fact that the guest always asks for extra sauce. That memory is the re-order mechanism, and it does not live in your loyalty program.
The economics look better than aggregators, but the position is structurally similar. No 30% commission is a genuine and large win. The discovery layer still sits above the operator. Square's version is the benign case — your catalog, your POS, your processing rate, your reporting. The version to plan against is an assistant that holds the guest profile and treats restaurants as interchangeable suppliers competing on data quality and rating alone.
Which is the argument for owning the data layer rather than renting it. The agent assembles what it says about a restaurant from that restaurant's own structured data. That is the one piece of the experience still fully under the operator's control, and the one place a POS vendor can add durable value.
A note on how early this is. The 22% figure is consumers who have used AI to help choose a restaurant, not to complete a purchase. NielsenIQ research cited by Square puts AI usage in shopping tasks at over 42% — again for discovery, comparison, and selection rather than autonomous buying. Morgan Stanley's projection that agentic shoppers could reach $190 billion to $385 billion in US e-commerce spending by 2030, or 10% to 20% of the market, is a wide-range forecast. The right response to all of this is data readiness, not a wholesale rewrite of your channel strategy.
A working checklist for operators, and a spec for the vendors serving them:
Menu data
- One structured source of truth. Items, categories, prices, images, and descriptions living in a system, not in a PDF or a designer's layout file.
- Modifiers as real option sets. Named groups with min/max rules, per-option pricing, and valid combinations; never pricing buried in description text.
- Allergens and dietary attributes as fields. Explicit contains / may-contain values. Never leave the agent to infer.
- Merchant-written item descriptions. DoorDash's internal analysis of SMB merchants who added their own descriptions to at least half of their menu items found a measurable sales effect, and in the agent channel, descriptions are literally what gets quoted back to the guest.
Sync and distribution
- Real-time hours, including holiday exceptions, per-channel hours, and separate kitchen vs. venue closing times.
- A live stop list that propagates to every channel within seconds of an item going out in the POS, not at the end of the shift.
- Order source tracking on every channel, so the AI channel can be measured rather than assumed.
Web and listings presence
- Schema markup on the website: Restaurant, Menu, LocalBusiness, FAQPage.
- A consistency audit across Google Business Profile, directories, delivery listings, and your own site, since the models cross-reference all of them.
For fifteen years, digital presence for restaurants meant being findable: on the map, in the listings, in the ten blue links. The work was marketing work.
Being in the answer is a different job. An assistant naming three restaurants is not ranking pages; it's evaluating whether it can read your data, trust your data, and transact on your data without embarrassing itself. That is an infrastructure problem, and it is solved upstream in the menu system, not downstream in the campaign.
Square just demonstrated what happens when a platform solves it for its merchants by default. The rest of the market now has a choice: build the same layer, or watch menu data become someone else's product.
