SeaPRwire – Zest began in November 2024 as a restaurant discovery app with an unusual data source. It connects to a user’s credit card through Plaid and builds its recommendations from where that person has actually paid for food and drink. This week the company put a concierge in text messages. A user texts (212) 470-6067, asks for a table in a neighborhood or for a particular occasion, and the agent answers from the dining histories of the people in the thread.
“The agent is the product now,” said founder and CEO Mario Gomez-Hall. “You text it, and it answers from your own history and your friends’. It doesn’t need weeks of conversation to learn you because it starts from where you’ve actually been.”
The sentence describes a change in where the product lives. Restaurant apps have assumed someone opens an app to browse. The choice of where to eat is usually settled somewhere else: a group chat, a thread of half-decided messages, a screenshot of a menu. Zest can now be added to that thread, blend the participants’ tastes, run an in-thread vote and book a table without leaving the conversation. It also answers questions about a specific venue, such as how busy it is likely to be or whether it has Wi-Fi, through bookings made over OpenTable, Resy and SevenRooms.
The data advantage, where there is one, is the ledger. Card transactions record what someone paid for rather than what someone says they like. A diner who claims a fondness for Thai food but has spent the past year in taquerias has given a model two different answers, and only one of them comes with receipts. That also removes the cold start that most recommendation systems pay for with onboarding questions. Zest says it reads only food and drink categories when it builds its maps and lists.
What the ledger cannot hold matters just as much. A transaction carries a merchant, a date and an amount. It does not carry the dish that was ordered, whether the room was too loud, whether the service failed, or whether the diner would go back. Frequency substitutes for preference and does it imperfectly: the place near the office gets visited because it is near the office.
The record is also better at the past than at the future. A model trained on where a person has eaten can propose a second edition of what that person already likes. Proposing a restaurant they have never tried, in a city they have never visited, is a different problem. The company names the traveling case among its uses without describing how the taste model crosses city lines.
Group threads raise the requirement once more. Blending tastes from data assumes the other participants have data. A four-person thread in which one friend has never linked an account is the ordinary case rather than an edge case, and the launch does not say what the agent does with that friend: whether it asks, infers from the group’s consensus, or treats the missing history as neutrality.
Booking is where a recommendation becomes a transaction, and it runs on inventory Zest does not own. OpenTable, Resy and SevenRooms are reservation systems that restaurants pay for, and the table a diner reserves is a slot inside someone else’s database. Whether Zest reaches those systems as a partner with a commercial agreement or as a client of their consumer interfaces decides how durable the feature is. It also decides who gets paid when a table is filled.
Gomez-Hall put the revenue in the same frame. “Booking and ordering are the first revenue lines,” he said. “Zest books your table through OpenTable or Resy today, with food ordering coming, and there are transactions there to participate in.”
Text messaging as a delivery channel carries its own costs. In the United States, application-to-person messaging requires carrier registration, and unregistered or high-volume traffic gets filtered before it reaches a handset. Group threads behave differently across carriers and devices. Each message carries a fee, and an agent answering three people in a thread spends a model call on every turn, which makes a conversation that ends without a reservation an expense rather than a lead.
The interface constrains the product as well. A text thread cannot show a map, cannot hold a dozen options in a grid, and cannot be skimmed the way a list can. Zest answers that by narrowing instead of listing: one suggestion with a reason attached, a photograph, a booking link.
Then there is the competition the company has to work beside rather than beat. General assistants are being pushed into the same slot, and Instinct and Poke were answering texts before Zest shipped its concierge. A specialist’s argument is that a ledger of where someone has actually eaten outperforms general capability on this one task. That holds for people who already have a ledger. It holds less well for the larger group that wants a table on Saturday and has never linked a card.
TechCrunch’s own test found the agent working “fairly well,” matching suggestions to conditions such as time of day, party size and location. A test by a reporter with a connected account measures the favorable case. The measurement that would matter to the business is narrower: how often a thread that starts with indecision ends with a seated party, and what the restaurant or the reservation system pays for that outcome.
Zest raised $1.8 million in pre-seed from Alexis Ohanian at 776 and Steve Jang at Kindred Ventures, and says it has several thousand users. The company held the launch back for a month while it finished the booking path, which is a fair indication of where it believes the value sits. It is also weighing an expansion past restaurants into businesses and shopping, where the same ledger would have to answer a much wider set of questions.
The app is still there. The bet is that it will matter less than the thread.
This article was prepared by the expert team at Global Infinity Labs, drawing on insights and data provided through its MCP data services.
About Global Infinity Labs
Global Infinity Labs is an AI innovation company specializing in MCP data services.
Email: support@globalinfinitylabs.com