The agent buys.

What already works

On 29 June 2026 Yandex unveiled a platform for integrating AI agents into Alice AI. An agent is not a chatbot with buttons but a service that understands a task in natural language and carries it through to the end. Two agents are already live for some users: Yandex Taxi orders a car on the phrase “let's go to work”, and Yandex Lavka fills a basket from a description of a dish — “get everything for pasta carbonara” — and offers to place the order. Agents for Yandex Delivery and Yandex Market are promised next, and by the end of 2026 the first outside companies will get access to the platform.

The road here was short. Since January 2026 Alice has been picking products right inside the dialogue: comparing specifications, finding offers. In April Yandex reported that around 10% of requests to Alice AI involved searching for, choosing and buying products. By May, according to the company, 19 million people were talking to the assistant every week.

Google is moving the same way: in May 2026 “preferred sources” appeared in AI Mode, and in August a button for websites through which readers can pin them as a favourite source. According to Google, by launch users had marked more than 345,000 sources and click through to them twice as often. This is no longer about search but about whom the neural network trusts by default.

How an agent differs from search

Search shows options; the person chooses. An agent chooses itself and shows the result: “taxi booked, arriving in four minutes”. The user sees neither the results page, nor the supplier's homepage, nor its design. They see only the outcome and the price.

Three consequences follow. First — the site as a shop window does not exist for an agent: it reads data, it does not look at banners. Second — the choice is made on formal criteria: price, lead time, availability, rating, completeness of the description. Third — the only way into the selection is through an integration or through data the agent can read. A beautiful landing page with “leave a request” simply takes no part in this scenario.

We already saw the first half of this shift in Yandex's AI answers: there the neural network answers instead of websites. The agent goes further — it also acts instead of the user.

Who will be affected first

Everything where the purchase is routine and repeated: food and grocery delivery, taxis, pharmacies, booking services, household orders, business consumables. This is exactly where the user does not care whom they order from as long as the terms are the same, and exactly where the agent picks the supplier without human involvement.

Complex purchases with emotion and consultation will hold out longer: renovation, property, medicine, B2B projects. But even there the agent will take over the first stage — gathering options and the initial cut. A company missing from its list will never reach the consultation.

What to change on your site right now

Structured data. Schema.org markup for products, services, prices, availability, opening hours, address, reviews. This is the language in which the agent reads a site. A page without markup is, to it, a picture without a caption.

Prices and lead times on the page, not “on request”. The agent compares what it can read. “Price on request” and “check lead times with a manager” exclude you from the comparison automatically.

An up-to-date feed. A catalogue as a feed (YML, the Yandex Market product feed) with availability and prices that is kept updated rather than left over from last year. This is a ready-made source for aggregators and agents, and where any integration will start.

Listings in Yandex services. The first agents work inside the Yandex ecosystem: Market, Maps, Services, Eda. A company with a complete and current listing there is already visible to the agent; without one, it is not. This costs less than any development and takes a week.

Reviews and rating. For an agent the rating is one of the few available quality signals. Working with reviews stops being a “reputation” matter and becomes part of the results.

Fast, clean answers to questions. Pages that answer directly “how much does it cost”, “how long”, “what is included” — a source both for AI answers and for agents during the initial selection.

What not to do yet

Don't commission “your own agent” before the platform opens: Yandex promises outside companies access by the end of 2026, but there are no integration rules yet, and anything built now will have to be redone. Don't rewrite the whole site “for neural networks” — the agent reads data, not copy, and correct markup will deliver more than new copywriting. Don't expect the format to “not catch on”: 19 million weekly users of Alice AI and agents from the country's largest services — this is no longer an experiment.

In brief

• The Taxi and Lavka agents are already live in Alice AI; outside companies will get the platform by the end of 2026.
• The agent does not see design — it reads price, lead time, availability, rating and markup.
• Routine, repeat orders will be the first to move to agent purchasing.
• Prepare now: Schema.org, open prices, an up-to-date feed, listings in Yandex services, reviews.
• Building your own agent before the platform opens is premature.

Frequently asked questions

Can a small business connect to Alice's agents?

Not yet: the platform is open only to Yandex's own services. Outside companies have been promised access by the end of 2026.

What does the agent see on my site?

Structured data, prices, availability, lead times, reviews. Design, banners and slogans it ignores.

Will the agent replace ordinary search?

In routine purchases — partly it already does. In complex services it will take over the initial selection, while the decision stays with the person.

Where to start if the budget is limited?

With listings in Yandex Maps and Yandex Market and with markup for prices and services on the site. These are the cheapest steps with the fastest effect.