What DoorDash said
DoorDash says orders placed through outside AI agents have added up to only hundreds so far. Ars Technica reported the figure on 30 October 2026, and noted it is still early days for a relatively new way of buying. Hundreds of orders, for a platform of DoorDash’s size, is close to nothing.
The claim being tested
The claim for the past year has been that shoppers will hand their buying over to an agent. You tell an assistant you want dinner, it finds the place, it orders, it pays. Companies have been building for that moment, and a lot of smaller operators have been told to rebuild their business around it.
I read every piece of news like this with one question: is there a need for it? Ordering food already works. Open the app, tap, pay. An agent has to be clearly better than that before people switch, and so far they have not switched.
What changed: the agents exist, and some platforms let them in. What did not change: the customer’s habit. Hundreds of orders says the habit has not moved yet.
Why hundreds of orders is the 50/50 I expected
I have said agentic commerce is too early to call. It is 50/50. Some platforms will open up, make themselves indexable and welcome agents. Others, feeling their hold on shoppers slipping, will block them, which is what Amazon has done with AI shopping agents. DoorDash’s number does not settle it either way. It tells me the customer has not moved yet, and nothing more.
So I would not build a business around agents. I also would not dismiss them. Both are bets on something nobody can call yet.
DoorDash can afford to sit in that uncertainty. The giants can wait it out, or shape how it turns out. They have the scale and the relationships to decide later. A small operator has neither, and that changes the question entirely.
For a small operator the shift is already decided
Take a cafe owner in Singapore with two outlets. They do not get to wait. Their menu already shows up in places they do not control: Google’s AI answers, a delivery app, a maps listing, and whatever assistant a customer asks when they type “laksa near me, open now, halal”. Each of those surfaces reads the cafe’s data and decides whether to name it.
Whether the order eventually arrives through an agent, or through a person who asked an AI and then walked in, the input is the same. The menu. The prices. The opening hours. What is available today. What makes this cafe different from the one next door. If that data is vague, stale or inconsistent, the cafe loses either way. If it is precise, the cafe wins either way.
That is the one move that pays regardless of how agentic commerce turns out: get data-centric. The website is one surface of many. Wherever your data shows up, it needs more precision, more detail and an accurate reading of what you are.
Precise data is also what gets you named today
This is the same reason interest in GEO, getting named in AI answers, keeps rising. An AI answer does not pick the biggest brand. It picks the one it understands most clearly and can quote most precisely. Engines quote specific facts and drop marketing language. “A cosy neighbourhood cafe with great coffee” gets ignored. “Open until 10pm, 40 seats, oat milk at no extra charge, five minutes from the MRT” gets quoted.
There is also the long tail to think about. Everyone asks their agent the same question slightly differently. To catch that intent, your information has to cover a much broader spread of questions, with varied and detailed attributes. We do this ourselves at AeroChat. It publishes many pages, each one aimed at a single integration with a specific piece of software, so each page catches one long-tail question an answer engine can match.
What to do
- List every surface your business data appears on. Website, Google Business Profile, delivery apps, maps, social profiles, directories.
- Make the facts exact and identical everywhere. Prices, hours, availability, location, what is included, what is not.
- Replace adjectives with attributes an engine can quote. Seats, distance, dietary options, lead times, delivery areas.
- Cover the long tail. The public holiday question, the “do you deliver to Punggol” question, the allergy question, the “can I book for twelve” question. Answer each one somewhere an engine can read.
- Do not rebuild your ordering around one agent or one platform. Build the data layer any agent or AI answer can read, and let the agents arrive, or not, on their own schedule.
Source: Ars Technica: Basically no one is using AI agents for their DoorDash orders