Agentic search in pharmacy retail. How AI assistants answer where to buy medicine and what a pharmacy chain can do

According to Yandex, 46.5 million people a month see the quick answers that Alice AI gives in Yandex Search. On top of that come the Alice AI chat, GigaChat, ChatGPT and DeepSeek. A person types "where to buy ibuprofen in Tver" and gets a ready list of pharmacies instead of ten links. If a pharmacy chain is not on that list, the shopper never hears of it. If the list shows an old price or the address of a closed pharmacy, the assistant gets it wrong on the chain's behalf.

Searching through an assistant is called agentic search, and agencies sell work on it under the labels GEO, AEO and LLMO. Only GEO (generative engine optimization) has a scientific definition: researchers introduced it in 2023. The other two terms live in sales decks. For a pharmacy chain the useful question is a different one: what data the assistant reads and who is responsible for it.

Alice AI takes products from Yandex Products and search results

Yandex itself says that, for recommendations, Alice AI picks online-shop offers from search results and the Products section of Yandex Search. To let a shop's goods appear in the assistant's answers, Yandex advises sending them to Yandex Products as a feed and marking up product pages. A feed is a file with the assortment, prices and stock that a shop sends to Yandex every day. Markup is a set of service tags on a product page that tell the search robot where the name and the price are.

For a pharmacy this raises three difficulties. Yandex Products requires a site where a shopper can put an item in a basket and place an order, and a feedback form does not count as a basket. A site without a basket stays in ordinary search results but is not shown in every product block. Many pharmacy sites work through booking, and the help pages do not say whether Yandex will accept a booking as a basket.

Under Yandex rules, delivery in the feed must always be switched off for prescription drugs, dietary supplements, homeopathy and medical devices. The law does allow remote sale of supplements and medical devices, so here Yandex is stricter than the law. For pharmacies, the Products account has its own fulfilment option: collection from the pharmacy.

The Find Cheaper agent looks for a lower price for the shopper and is still in beta. Yandex has excluded medicines from its recommendations. If the shopper adds a medicine to the selection themselves, the agent will keep searching for a price.

A place in the answer itself cannot be bought. Asked about priority placement in the agent's selection, Yandex replies that it is "not provided". Under the Yandex Search partner integration policy, partners get no advantages, including in ranking. Paid advertising next to the answer does exist. In the Alice AI chat Yandex is testing ads under the answer, labelled "promo". According to the company, they do not affect the answer itself.

A pharmacy's card in Yandex Business can hold its address, opening hours and even a product list. Yandex does not say whether that data is used in Alice's answers. Since August 2025 the company has also run its own pharmacy aggregator, Yandex Apteki; at launch Yandex cited more than 200 cities. The help pages do not describe any link between its offers and Alice's answers either.

ChatGPT, GigaChat and DeepSeek search the open web

ChatGPT is not officially available in Russia: Russia is missing from the list of countries where OpenAI supports the service. People in Russia use it anyway, but OpenAI's documents give no way to tell how many. ChatGPT builds answers with links from web search. It searches through partners, of which the help page names Microsoft and Shopify, and through its own search robot, OAI-SearchBot. OpenAI says that a site closed to OAI-SearchBot will not appear in ChatGPT search answers. The company's other robot, GPTBot, collects text to train models. It can be blocked without any effect on search.

ChatGPT shows product results outside the United States too. A shop's own feed, however, can only be submitted for the US market, under the open ACP (Agentic Commerce Protocol) and only by an approved OpenAI partner. There is no way in for a Russian pharmacy chain. OpenAI describes stock availability in local shops as a plan for the future.

GigaChat can search the internet and cite sources. Sber does not disclose which search engine is behind it, and the GigaChat help pages say nothing about product answers. In search mode DeepSeek calls third-party search services and does not name them.

For a Russian pharmacy chain, a documented direct channel into assistant answers exists only at Yandex. It reaches the other three through what they find on the web: its own site and aggregator pages.

There will be no "buy" button for medicines in the chat

The step after the assistant's answer is called agentic commerce: the assistant places the order itself. For medicines, the big platforms keep this door shut.

In September 2025 OpenAI launched payment right inside ChatGPT, first for Etsy sellers in the US. In March 2026 the company allowed shops to send buyers to their own checkout. Its page for merchants now says that OpenAI is moving away from a standalone Instant Checkout. Prescription drugs are banned from ChatGPT feeds. In January 2026 Google opened the UCP (Universal Commerce Protocol). Checkout through it works for selected merchants in the US, and Google has excluded the healthcare/pharmaceuticals category from it.

On 27 February 2026 Yandex announced YCP, its own agentic commerce protocol. Products in Search, in Alice AI and in Yandex Ritm get a "Buy in 1 click" button. It opens the Yandex checkout, and the order itself goes to the shop. Once a shop has connected the Yandex checkout, the button appears on all its products by default and is removed with a separate flag in the feed. The help pages do not say whether Yandex lets medicines into its checkout.

In my view, such a button on a medicine amounts to remote sale. Government Resolution No. 697 defines remote sale as accepting, assembling, storing and delivering orders, as well as dispensing. Under it, the contract is concluded as soon as the pharmacy receives the order. Only a pharmacy organisation that has held a licence for at least a year and has a permit from Roszdravnadzor, the federal healthcare watchdog, may sell over-the-counter medicines remotely. Permits are issued for specific sales channels, so before turning the button on it is worth checking whether the permit covers this one. Without a permit, a basket for medicines is not allowed. What remains is a showcase: assortment, prices and the addresses of pharmacies where the product can be bought. A basket for supplements, medical devices and cosmetics is fine.

Nobody may currently sell prescription drugs remotely. The experiment in Moscow, the Moscow Region and the Belgorod Region ended on 1 March 2026, when Article 55.1 of the Law on the Circulation of Medicines lapsed. I found no extension. In September 2026 the Ministry of Health proposed replacing the remote-sale permit with a notification from 1 September 2027. Prescription drugs are excluded in the draft as well. While it remains a draft, the permit regime applies. I analysed it in Notification instead of a permit.

Booking with payment and collection at the pharmacy is a separate question. In a 2024 letter, No. 25-4/433, as reported by the legal service Garant, the Ministry of Health reiterated that remote sale means accepting, assembling, storing and delivering orders for over-the-counter drugs. The letter says nothing about booking without delivery, and a ministry letter is not a legal act anyway. I described how online booking works in Pharmacy online booking. For prescription drugs the risk is higher. I would only bring booking into assistant answers after talking to a lawyer.

What a pharmacy chain can do

Feed and site from one export. The Yandex Products feed is built every day from pharmacy stock and retail prices. The price in the feed must match the price on the site. The Products quality service compares them and may restrict display when they differ. The help pages do not say what deviation is tolerated. If the feed and the site draw on different exports, a mismatch will happen sooner or later. That is why there should be one export. The product master data is brought to a single reference list in advance: a pack entered under three names goes into the feed as three different products.

Page markup. Yandex advises marking up a product page with the Product schema and an Offer. A pharmacy page can be marked up with the Pharmacy schema, which Yandex Business supports. The Products help pages warn, though, that markup has no direct effect on ranking. Markup helps a site get into Products. It will not lift the site higher in search.

Robot access. robots.txt must not block YandexBot, bingbot or OAI-SearchBot. Check that YandexAdditional is not blocked either: blocking it removes the site from Alice AI answers. robots.txt alone is not enough, because the hosting provider can cut robots off by itself. On my own site, tiukon.ru, the hosting provider dropped GPTBot connections on its own server, even though robots.txt let it in. That did not hurt ChatGPT search, since OAI-SearchBot got through, but a host can just as quietly cut off a search robot. I only saw it in the access log. If your site runs on shared hosting, open the access log and look for robot names and response codes. New pages are easy to announce to Yandex and Bing via IndexNow, so that the robots learn about them sooner.

Aggregators and maps. An assistant that searches the web also finds pharmacy aggregator pages. A pharmacy that is missing there, or listed with an old address, loses this route to the shopper as well. Aggregators are building assistants of their own. Apteka.ru built one on the Yandex AI Studio platform. According to Yandex Cloud, in a test on 20% of the audience it raised order conversion by 77%. It does not suggest prescription drugs. Addresses, opening hours and phone numbers in Yandex Business and 2GIS need to be kept up to date in any case: shoppers see them on the map.

Some of what is sold today as "AI optimisation" is not backed by the documents.

• Yandex provides no priority placement for uploading a feed and gives partners no ranking advantage.
• A rule that "display is removed when prices differ by more than 5%" does not appear in the Yandex documents I read. They name no tolerance, so display can be restricted even for a small difference.
• The llms.txt file was proposed in 2024 by the developer Jeremy Howard and has not become a standard. Google says its search ignores such files. I found no statement from OpenAI or Yandex that they read llms.txt.
• The techniques from the 2024 GEO paper, that is, citing sources, adding quotes and adding statistics to page text, gave up to 40% more visibility. But 80% of the queries in that study were informational, and "where to buy" queries were not tested.

How to measure the result

Yandex itself warns that Alice AI does not use ready templates and builds each answer anew. So a single screenshot proves nothing, and measurement rests on a few principles.

The unit of measurement is one assistant answer to one query in a new session, with history switched off and the city in the query text. The query set is fixed before work starts and kept unchanged to the end. Each query is repeated in different sessions to see the spread. Some products are kept as a control group and left untouched. The effect equals the gain on the products that were worked on, minus the gain on the control products. If both groups rose equally, the assistant itself has changed.

The result is reported as a share with an interval. Suppose the share of answers that name the pharmacy chain rose from 20% to 30%. With 50 answers in each round, the 95% interval for this difference is ±17 percentage points; with 100 answers it is ±12. In both cases a 10-point rise cannot be told apart from noise. With 150 answers the interval narrows to ±9.7 points, and the rise only just clears it. The count must be of distinct queries, because repeats of the same query are correlated. The authors of "Quantifying Uncertainty in AI Visibility" (2026) arrived at similar volumes. They estimated the share of answers citing a site with an interval 15 points wide. For that they needed between 60 and 150 distinct queries per platform.

Some data can be taken from Yandex itself. Since 7 April 2026 Yandex Webmaster has offered the "Site visibility in Alice AI" report. It shows the share of queries for which Alice AI mentioned the site, is updated weekly and keeps three months of data. It has three limitations. It covers only quick answers in Search, not the chat. It has no clicks or impressions. Statistics are collected only for queries where the site already ranks high. If your report is populated, it is the best before-and-after benchmark: Yandex calculates it, not a contractor. The Yandex Products "Visibility in Search" report gives impressions and clicks for products, but the Alice chat is not in it. Yandex Metrica has no separate "Alice AI" traffic source, so visits from assistants are found by the addresses they came from: alice.yandex.ru, chatgpt.com, giga.chat, chat.deepseek.com. This is a lower bound: many people read the answer and go to the pharmacy without opening the link.

I start work with a pharmacy chain by looking at what assistants answer today: whom they name in its cities instead of it, and with what prices and addresses. This audit is part of the FlexPricer product line, which I take part in developing. I designed the methodology myself and run the audit personally. How it works is described on the page Pharmacies in AI assistant answers.