A single estimate of the weekly and daily audience of generative AI chats across eleven services, the mix of tasks, and profiles of active users — built on open data from Mediascope, VCIOM, OMI, Yandex, Similarweb, and other sources.
* What is measured: user-initiated requests to a model in a generative AI chat — the website and app, chat inside other products, Telegram bots, telecom-operator services; passive AI answers in search are not included. Scope: eleven services — Alice AI, ChatGPT, DeepSeek, GigaChat, Gemini, Perplexity, Qwen, Grok, Character.AI, Claude, Copilot. The aggregate audience is the sum of the services' weekly audiences (57.5 million; one person can be part of the audience of several services); unique weekly users number 48.6 million, of whom about 64% use Alice AI. The daily audience is a model estimate (D7–D8).
Weekly — 48.6 million (40% of the population aged 12+); monthly — 67.9 million (56%); daily — 12.9 million (10.6%). The estimate covers eleven services for the reference period, Q2 2026, and includes the measurable conversational channels: the website and app, chat inside other products, Telegram bots, and telecom-operator services; passive answers in search are not included. With the assumption on teenagers' weekly reach (B4) at its bounds, the weekly audience is 48.1–49.1 million. The daily audience is a model estimate whose parameters are calibrated to a Mediascope benchmark. Allowing for differences in the age base, the estimate is comparable to HSE University and Romir data (44% of the population aged 18–65 weekly); it is higher than Mediascope’s monthly reach for websites and apps (50%), which does not include embedded use or messengers. Calculation: step 1 of section 1.2; daily audience: assumptions D7–D8.
Alice AI’s weekly audience is 30.9 million people: 54% of the aggregate audience of the eleven services (the sum of their audiences with overlaps; calculation: D12) and about 64% of the market’s unique weekly users — 3.7 times more than ChatGPT (8.2 million), which ranks second despite not being officially available. The Alice AI estimate refers to the chat and is based on a Yandex report: 33.2 million people a week hold conversations with Alice AI in the chat — in the Alice app, Yandex Browser, the Yandex app, and on the website; Yandex counts quick answers in Search and on Yandex Station separately, and they are not included in the estimate. The Mediascope panel attributes chat inside Yandex Browser and the Yandex app to those products and sees only part of the audience; for ChatGPT, about 69% of the audience comes through channels the panel does not see: access via VPN, Telegram wrapper bots, and telecom-operator services; for the last two, ChatGPT’s share has to be set by an assumption.
Russian services — Alice AI and GigaChat — account for 63% of the aggregate weekly audience of the eleven services; US services — ChatGPT, Gemini, Perplexity, Claude, Grok, Character.AI, and Copilot — for 24%; Chinese services — DeepSeek and Qwen — for 12%. The Herfindahl–Hirschman Index (HHI) calculated from the services' shares of the aggregate weekly audience (with overlaps) is 3,321; with the same method, the global figure is about 5,200, since ChatGPT accounts for about 70% of the category’s weekly audience. The index measures how audiences are distributed across services, not market concentration by unique users: the same person is counted in the audiences of several services.
A weekly Alice AI user sends about 5 requests a week, and a weekly ChatGPT user worldwide about 25 messages; for ChatGPT users in Russia, the estimate is about 18. Alice AI’s daily audience is 20% of its weekly audience; for ChatGPT worldwide, the figure is 55%. The leader’s reach was built through distribution, and its audience has a weaker habit of daily use.
Information search tops the task list for all five measured services: 51% to 71% of users name it. The differences begin with the work context: its share is 37% for Perplexity, 32% for ChatGPT and DeepSeek, and 21% for Alice AI. According to Yandex logs, about 10% of requests to Alice AI are work-related.
People aged 12–24 make up 17% of the population aged 12+, 27% of the monthly audience of AI tools, and 43% of the daily audience: the more often a person uses AI tools, the younger they are on average. The daily reach of AI tools among 12–24-year-olds is 2.6 times higher than the population average. Audience composition by age, gender, and service is covered in section 3.
Open data on the audience of AI tools in Russia contradict one another because they measure different things. This study brings them onto a single scale and adds calculated metrics that the sources do not provide.
More than ten quantitative studies of AI tool use in Russia were published in 2025–2026. The Mediascope panel measures the reach of websites and apps among the population aged 12+ and does not see AI tools built into search and messengers. VCIOM and OMI surveys capture self-reported use by internet users aged 18+ through any access channel. Digital Budget and Similarweb count web traffic, in which one active user weighs many times more than an occasional one. Yandex and Sber disclose their own metrics without a country breakdown and without a common definition of a user. Each source is correct on its own terms, and none of them answers the question of how many people use each service in a typical week.
The calculation has four steps. First, the total market size is estimated: the weekly, monthly, and daily audience of all generative AI chats combined. Next, the audience is allocated across services by two different methods: one uses survey reach (OMI, Beeline Analytics) and frequency coefficients; the other uses panel measurements (Mediascope) and traffic (a16z, Similarweb, Digital Budget), adjusted for channels invisible to measurement. In the last step, the companies' own metrics — Yandex’s for Alice AI and Sber’s for GigaChat — are added as a third estimate, and the estimates are reconciled into a final estimate; the range is the interval between them. Activity metrics, the task mix, and audience composition are calculated on this basis. All assumptions are numbered and collected in the appendix; the calculation can be reproduced from published data.
What is measured: generative AI chats, that is, user-initiated requests to a model in a conversational interface. The scope includes the service’s website and app, chat inside other products (for example, chat with Alice AI in Yandex Browser), Telegram bots, and telecom-operator access services. It excludes passive AI answers in search — Yandex counts Alice AI’s quick answers below the search bar as a separate metric (49.5 million users a month in Q2 2026), and they are not added to the chat audience — as well as voice devices and other surfaces where a user-initiated dialogue with a model cannot be isolated. Corporate use via API and image generators (Shedevrum, Kandinsky, Midjourney) are not included in the audience estimate. The reference period is Q2 2026, and the sources refer to different dates: the core audience data to October 2025–July 2026, trend cross-checks and intensity data to 2025, and data on web-version engagement and on shopping to August 2026. The final estimate is therefore a model estimate for the reference period, not a measurement of a single quarter; data are as of the end of August 2026.
The market leader depends on what exactly is counted. By self-reported use, Alice AI leads; by web traffic, ChatGPT; by population reach among foreign services, DeepSeek. This section brings all measurements to a single quantity — the number of people who use a service in a typical week.
The three figures below are the numbers of unique people aged 12+ who actively used at least one generative AI chat in a month, a week and a day (MAU, WAU and DAU — monthly, weekly and daily active users). Each person is counted once, however many services they use. The chart below the boxes shows the weekly audience of each service separately; the sum across services is larger than the number of unique users because in a week one person uses 1.18 services on average. The gray notes in the boxes show how each figure was obtained and what it was compared with.
Millions of people aged 12+ who used the service in a typical week, across the available measurable conversational channels: app and website, chat inside other products, Telegram bots, access via VPN and via telecom-operator services. Passive AI answers in search are not included. The bar is the final estimate obtained by reconciling the methods; the whisker is the interval between the methods' estimates: a scenario range, not a confidence interval. Method A: the service’s share among AI users in the survey × the share of weekly active users. Method B: the Mediascope panel and traffic plus channels that measurement providers do not see. Method V: company reporting. Alice AI is a model estimate of the Russian audience of the Alice AI chat based on Yandex’s report for Q2 2026 (33.2 million weekly chat users) minus the 7% of users outside Russia; it is not an estimate for all product surfaces; the range is −10% to +5% (C4). A model estimate for the reference period: the sources refer to different dates (October 2025–July 2026, see the measurement map), and panel values are brought to the reference period using trends D1, D4 and D5.
| Service | Used in past 6 months, % of AI users (OMI) | MAU, M | WAU, M | range | DAU, M | range | Share of aggregate WAU |
|---|---|---|---|---|---|---|---|
| Alice AI | 71% | 49.8 | 30.9 | 27.8–32.4 | 6.2 | 4.2–8.1 | 54% |
| ChatGPT | 42% | 13.7 | 8.2 | 5.0–13.4 | 3.3 | 1.5–6.7 | 14% |
| DeepSeek | 25% | 12.3 | 6.1 | 4.3–8.7 | 1.8 | 0.9–3.3 | 11% |
| GigaChat | 38% | 16.1 | 5.7 | 3.0–8.5 | 0.8 | 0.3–1.7 | 10% |
| Gemini | 15% | 4.3 | 2.2 | 1.3–3.7 | 0.6 | 0.3–1.5 | 4% |
| Perplexity | 4% | 2.1 | 1.0 | 0.9–1.3 | 0.4 | 0.2–0.6 | 2% |
| Qwen | 4% | 1.7 | 0.9 | 0.8–0.9 | 0.3 | 0.2–0.3 | 1% |
| Grok | 5% | 1.7 | 0.8 | 0.5–1.2 | 0.2 | 0.1–0.5 | 1% |
| Character.AI | 3% | 1.1 | 0.7 | 0.6–0.8 | 0.3 | 0.2–0.5 | 1% |
| Claude | 3% | 1.1 | 0.6 | 0.4–1.0 | 0.3 | 0.1–0.5 | 1% |
| Copilot | 3% | 1.1 | 0.5 | 0.4–0.7 | 0.2 | 0.1–0.3 | 1% |
| Aggregate audience of the services (sum with overlaps) | — | 105.1 | 57.5 | 14.4 | 100% | ||
| Unique users in the market (each person counted once) | — | 67.9 | 48.6 | 12.9 | — |
The first column is the share of AI users who used the service in the past 6 months, from the OMI survey (March 2026); for Alice AI and DeepSeek, the calculation uses the average with Beeline Analytics data (69% and 27%). The range is the interval between the methods' estimates (section 1.2): a scenario range, not a confidence interval; for Alice AI, −10% to +5% relative to the Yandex report (C4). MAU is obtained by dividing WAU by the WAU/MAU ratio (assumption D6). DAU = WAU × the service’s DAU/WAU coefficient (D7); the DAU range is the product of the WAU range bounds and the bounds of the D7 coefficient. The aggregate audience is the sum of the services' audiences, in which a person who uses several services is counted in each of them; it is 1.18 times larger than the number of unique users per week and 1.55 times larger per month. A service’s share is calculated from the aggregate audience (D12). Sums of rounded values may differ from the totals by 0.1 million and 1 p.p.
geoboost calculation; methodology in section 1.2, assumptions in Appendix A (blocks B–D). Data: VCIOM (June 2026), OMI (March 2026), Beeline Analytics (Q1 2026), Mediascope (October 2025, April 2026), Yandex (Q2 2026 report), Sber (2025), a16z and Similarweb (January and July 2026), Digital Budget (2025), Telegram analytics (July 2026).
The market is shaped like a pyramid with a single broad base. Alice AI reaches 30.9 million people a week — half of the services' aggregate audience and 3.7 times more than ChatGPT. This is the only estimate that rests on the company’s own report (the chat audience, not all product surfaces), so its range is narrow. ChatGPT holds second place despite being officially unavailable, and its estimate is the most uncertain: measurement providers see between a sixth and a third of its audience, and the rest comes through VPNs, Telegram wrapper bots and telecom-operator access services. In the last two channels, ChatGPT’s share has to be set by an assumption — hence the wide whisker on the chart: Method B without the multi-model channels gives 3.3 million a week, Method A gives 13.4 million.
DeepSeek and GigaChat are almost equal in size, but their stories differ. The panel sees almost all of DeepSeek’s audience — and sees its web visits from Russia fall threefold over the first half of 2026. GigaChat is the only service with three estimates: the survey gives 8.5 million, the panel with invisible channels 3.0, Sber’s data adjusted for low intensity 7.0; the final 5.7 million is their geometric mean (section 1.2 explains how). Then comes the tail: Gemini, then Qwen and Perplexity at a million each, and four services below a million; there is almost no panel data for them, and the order within the tail is tentative. The Gemini estimate refers to the period before August 14, 2026, when Google restricted access to the service from Russia, including circumvention via VPN; this audience is likely to be split among ChatGPT via telecom-operator channels, DeepSeek and Alice AI [33].
The aggregate audience of the eleven services — the sum of their weekly audiences, in which a person who uses several services is counted several times — is 57.5 million against 48.6 million unique users: in a typical week, one person uses 1.18 services on average. First place does not depend on the assumptions. ChatGPT keeps second place within each of the methods at a scenario share of 0.6 in the multi-model channels; in the “direct access only” scenario, Method B gives 3.3 million — less than DeepSeek (4.3 million by the same method) — so the estimate intervals of ChatGPT, DeepSeek and GigaChat overlap, and places two to four may change. Russian services account for 63% of the aggregate weekly audience, US services for 24%, Chinese services for 12%.
Below are the seven main measurements for eight services. The columns are not additive and cannot be compared with each other directly: each has its own denominator, period and channel. This is precisely why ChatGPT ranks first by web traffic and fourth by panel reach, while Alice AI ranks first in surveys and fifth by traffic. Traffic including mobile apps (a16z, based on Similarweb and Sensor Tower data, January 2026) gives the same picture: DeepSeek 28%, ChatGPT 27%, Alice AI 10%, GigaChat 9%. All seven measurements are used in the calculation: each either feeds into one of the methods or serves as a check on its assumptions.
The number in a circle is the value in the column’s units; the column header names the source, period, unit of measurement and base. The size and color saturation of a circle show the value relative to the column maximum. The first three columns count people: OMI and Beeline — the share of AI users who use the service (a person could name several services, so the shares in a column add up to more than 100%), Mediascope — the share of the population. The next three count visits, not people: Digital Budget and a16z — the service’s share of traffic (for a16z, the total across all chatbots in the country is 100%; services without a separate segment, including Claude, are below 1%), Similarweb — the number of visits to the website. The last column is company reports, in the companies' own definitions. A dash means the source did not publish a value for the service: Beeline disclosed three services, Similarweb for January 2026 three. The columns are not additive and cannot be compared with each other directly.
Sources: OMI, online survey, fieldwork March 16–23, 2026, n=2,328, AI users in the past 6 months aged 18–55 (boost sample aged 56–65, n=102, not used in the calculation) [4]; Beeline Analytics, Q1 2026, share among AI users on the operator’s mobile network [18]; Mediascope Cross Web, October 2025, monthly reach of websites and apps excluding integrations [5]; Digital Budget, January–October 2025, shares of inbound web traffic of 12 AI platforms excluding mobile apps [19]; a16z, chart “AI Chatbot Market Share by Country,” web and mobile traffic for January 2026, Similarweb and Sensor Tower data, values read off the chart (±0.5 p.p.) [23]; Similarweb, visits from Russia in January 2026 [21]; Yandex — Q2 2026 report, weekly users of the Alice AI chat [9]; Sber — GigaChat MAU for 2025 (secondary report on the annual report) [16].
The seven lenses come from five types of sources: surveys, operator data and panels count people; web analytics and the reports built on it count visits; companies disclose their own metrics. Below is who stands behind each column of the map, what exactly it measures and where it is used in the calculation.
The calculation runs top-down — from market size to services — and bottom-up — from measurements and traffic to the market; company reports provide a third point, and the final estimate reconciles all three. Each block gives the input data, the formula, the assumptions adopted (numbered as in Appendix A), the result and cross-checks against external benchmarks.
How many people aged 12+ use AI tools per week, per month and over six months — across all services combined.
Distribute the six-month user base across services and convert it into weekly audiences.
Build a bottom-up estimate from panel and traffic data, supplemented with channels that measurement providers do not see.
Add company reports as a third estimate, combine the methods into a single final estimate and check it against external benchmarks.
Weekly audience, millions. Method A overstates services with high “trial” reach, Method B understates services with invisible channels; for GigaChat, the third estimate is Sber’s data adjusted for intensity. The final estimate is the geometric mean of the estimates; their spread is the range on the chart in section 1.1 (Alice AI is not shown: all its estimates rely on the Yandex report).
geoboost calculation. Method A: 6-month user base × service share (OMI; for Alice AI and DeepSeek, the average with Beeline Analytics) × weekly frequency coefficient by service class (calibrated to Alice AI). Method B: Mediascope panel reach (October 2025) × trend + audience of Telegram bots, telecom-operator channels and embedded assistants × WAU/MAU; without panel data — a16z traffic equivalent ÷ visits per user. Method V: MAU from Sber data × WAU/MAU 0.35.
Six findings of this section, each with the visualization it is based on.
Alice AI’s reach was built by distributing the chat inside Yandex products, not by a standalone app. According to the company’s report, 33.2 million people a week hold conversations with Alice AI in the chat — in the Alice app, Yandex Browser, the Yandex app, and on alice.yandex.ru; Yandex counts answers under the search bar (49.5 million people a month) and on Yandex Station separately, and they are not included in this study’s estimate. The Mediascope panel sees only the website and app — 17.5 million people a month in October 2025, about a third of the calculated monthly audience for Q2 2026; the panel attributes chat inside Yandex Browser and the Yandex app to those products, and part of the gap is growth of the chat itself: its weekly audience grew from 19 to 33 million over two quarters [9] [11] [5].
The panel sees the website and app (October 2025); the calculation covers the chat audience across all entry points for Q2 2026.
Mediascope, October 2025 [5]; geoboost calculation; OMI [4].
Direct access to ChatGPT through the website and app amounts to a few million people a month; the full estimate is 13.7 million a month and 8.2 million a week. The difference comes from three channels invisible to panels: VPN, multi-model Telegram wrapper bots, and telecom-operator services, which since December 2025 have offered access to ChatGPT, Gemini, DeepSeek, Claude, and Grok without a VPN. Bots and operators offer several models, so ChatGPT’s share in them is a scenario assumption (0.6; D3), not a measurement: without these channels, Method B gives 3.3 million a week. Both measurement systems see only the smaller part of the audience: the Mediascope panel about a third, Beeline’s mobile network a sixth; the rest is a blind spot for them [30] [34] [18].
The geoboost estimate for the measurable conversational channels is taken as 100%. Panel: website and app (October 2025); mobile network: 7% of AI users versus 42% in the OMI survey.
Mediascope, October 2025 [5]; Beeline Analytics, Q1 2026 [18]; OMI [4]; geoboost calculation.
Russia is DeepSeek’s second-largest market after China (9.15% of visits) and Qwen’s largest market worldwide (34%). But both audiences are shrinking: visits to DeepSeek from Russia fell by a factor of three between January and summer 2026, and app downloads in Russia fell 57% in Q4 2025; Qwen’s traffic more than halved. DeepSeek’s weekly audience is 6.1 million, Qwen’s 0.9 million; the panel estimate is carried forward to Q2 with a scenario coefficient of 0.75 (D4). Demand for foreign models free of access restrictions moved to Chinese providers but did not turn into a habit [20] [21] [31].
January: a16z (web and apps) and Similarweb; May–July: Similarweb, web, monthly average.
38% of AI users say they used GigaChat in the past six months — the third-highest result in the market. Sber reports more than 20 million monthly users and 800 million requests a year: 4.8 requests per person per month, versus 14 for Alice AI and about 88 for ChatGPT worldwide. This is a broad but shallow base, gathered from all surfaces — from the website to Salute and the assistant in SberBank Online. Three estimates of the weekly audience — 8.5 million from the survey, 3.0 from the panel with invisible channels, and 7.0 from Sber data adjusted for intensity — differ almost threefold; the final estimate of 5.7 million is their geometric mean (section 1.2 explains how); half of its users classify it as neither a work nor an entertainment service. Sber is investing in distribution: TV advertising for GigaChat quadrupled in the first half of 2026 [1] [16] [62] [17].
Vendor data for 2025; for ChatGPT: 2.5 billion messages a day with 700 million weekly users and WAU/MAU of 0.82.
Sber (secondary report) [16]; Yandex [13] [11]; OpenAI, Sensor Tower [24] [26] [27]; geoboost calculation (assumptions F1, F3).
The Herfindahl–Hirschman Index (HHI) calculated from the services' shares of the aggregate weekly audience is 3,321, versus about 5,200 worldwide, where ChatGPT accounts for about 70% of the category’s weekly audience; in Russia, its share is 14%. The shares are calculated from the sum of audiences with overlaps, so the index describes how service audiences are distributed, not market concentration by unique users (D12). Three ecosystems compete in one market: measured in people, Russian services account for 63% of the aggregate weekly audience; measured in traffic, only 17%, because traffic data do not capture Alice AI’s embedded use. In the US and China, the market belongs to domestic products [53] [23] [19].
Russian: Alice AI and GigaChat; US: ChatGPT, Gemini, Perplexity, Claude, Grok, Character.AI, Copilot; Chinese: DeepSeek, Qwen. Traffic: the average of shares from a16z (web and apps, January 2026) and Digital Budget (web, 2025, top 5 normalized to 100%).
Over six months, the average user tries 2.1 of the eleven LLM services in scope; in a typical week, a user turns to 1.18 services on average. An independent Impuls.guru survey gives the same picture: 2.2 services per user. The market is still in an experimental phase: the audience of the second and third services consists of people for whom that service is not the main one [4] [60].
The shorter the period, the closer the figure is to one: people try several services, and one becomes a habit.
OMI, March 2026 [4]; Impuls.guru [60]; geoboost calculation.
This section measures how often and how intensively each service is used, and which tasks make up that use. It turns out that reach and intensity are distributed across the market in opposite ways.
Frequency is the first measure of activity, and it is growing: according to VCIOM, the share of weekly users among those who used AI tools in the past year rose from 51% to 58% in six months. According to OMI, more than a third of users turn to AI three or more times a week, and frequency declines with age: half of the 18–24 group do so, compared with a quarter of the 46–55 group [1] [4].
% of AI users (past 6 months). Average number of uses per month: all — 6.8; 18–24 — 8.0; 25–35 — 6.8; 36–45 — 6.2; 46–55 — 6.1.
OMI, fieldwork March 16–23, 2026, n=2,328, internet users aged 18–55 (56–65 boost sample not shown) [4].
The second measure is stickiness, the daily audience as a share of the weekly audience. There are no direct measurements by service for Russia, so the coefficients are derived from three lines of evidence:
The coefficients are scenario assumptions, not measurements. The sum of the services' daily audiences, adjusted for overlap (D8, also a scenario value), gives a model estimate of 12.9 million people, or 10.6% of the population aged 12+; the parameters are calibrated to the Mediascope benchmark of 11% daily across websites and apps, so closeness to it is not an independent check. At the bounds of assumptions D7–D8, the estimate lies in the range of 8.9–18.2 million [12] [7] [27] [28].
Gray: benchmarks outside the per-service assumptions — ChatGPT worldwide (measured by Sensor Tower) and the market as a whole per the model; lines show the assumption range. Per-service coefficients are scenario assumptions (D7): anchors from Yandex and Sensor Tower data, global DAU/MAU ratios discounted for the access barrier; the parameters are calibrated to Mediascope daily reach.
geoboost calculation (assumption D7); Yandex/3DNews (4.5 million per day, December 2025; 19 million per week, Q4 2025) [12] [11]; Tailwinds, based on Sensor Tower (ChatGPT DAU/MAU 45%, WAU/MAU 82%; Gemini DAU/MAU 22%) [27]; a16z, based on Sensor Tower (Grok DAU/MAU 25%) [28]; for the market as a whole, model DAU/WAU ≈ 0.26; benchmark: 11% of the population daily per Mediascope, April 2026, against 40% weekly per the model ≈ 0.28 [7].
The third measure is intensity, and here the gap between Russian and foreign services is widest. A ChatGPT user worldwide sends about 25 messages a week, and an Alice AI user sends 5.3 requests, almost five times fewer. The gap narrows to 1.7 times when counted per daily user (3.8 versus 6.5 a day): there are simply fewer daily users among Alice AI’s weekly users. GigaChat, with more than 20 million monthly users, receives 4.8 requests per person per month — the intensity of a reference service [24] [26] [13] [16].
There are no direct data for ChatGPT in Russia, so its intensity is estimated from the global figure. If messages per weekly user are proportional to the number of active days, then with stickiness of 0.40 versus 0.55, a Russian user sends about 73% of the global volume, or 18 messages a week. The bounds of the range are 40% and 100% of global intensity (10–25 messages): the lower bound reflects episodic access via VPN and Telegram wrapper bots; the upper bound reflects parity with the world, which is supported by the length of web visits from Russia: 10 minutes versus 6:35 on average for chatgpt.com [22] [20].
ChatGPT worldwide: 2.5 billion messages a day with 700 million weekly users (July 2025). ChatGPT in Russia: estimated at 73% of global intensity, in proportion to the number of active days; range 40–100%. Alice AI: 2.9 billion requests in 2025 with an estimated average weekly chat audience of 10.5 million. GigaChat: 800 million requests in 2025 with an average monthly audience for the year of about 14 million and WAU/MAU of 0.35.
OpenAI/NBER [24] [26]; Yandex [13] [11]; Sber (secondary report) [16]; Similarweb Pro [22]; geoboost calculation (assumptions F1, F3, F4).
Audience and intensity together give the weekly volume of requests: about 164 million requests for Alice AI; about 150 million messages (82–206) for ChatGPT in Russia, whose audience is 3.7 times smaller; and about 15 million requests for GigaChat, based on vendor data. This is the main refinement in this section: by number of people, the market belongs to Alice AI; by number of messages, Alice AI and ChatGPT are comparable.
Web versions show engagement within a visit. A visit to chatgpt.com from Russia lasts 10 minutes — 2.6 times longer than a visit to alice.yandex.ru and 50% longer than the global average: the access barrier filters out casual users, leaving those who work with the service seriously. On the web, Alice AI is a place for quick questions, with the highest bounce rate and the fewest pages per visit; GigaChat has the shortest visits, with almost as many pages as ChatGPT. In the global set, Character.AI stands out with 16 minutes and 12 pages per visit: a conversational mode of use rather than a reference one [22] [20].
Violet: traffic from Russia (Similarweb Pro, August 2026); gray: worldwide (Similarweb, July 2026). The geoboost engagement index for Russia: ChatGPT 100, GigaChat 73, Alice AI 61 — the average of three metrics, with the best service on each metric scored at 100.
Similarweb Pro, Audience → Geography, traffic from Russia, August 2026, web versions excluding apps and voice devices [22]; Similarweb, public pages for each domain, July 2026 [20]. The ChatGPT bounce rate for Russia is approximate (31.2%). Data for Russia come from geoboost’s paid subscription and have not been independently verified.
The last view combines traffic and audience in a single metric, the traffic-per-user index: a service’s share of measured traffic divided by its share of the weekly audience. A value of one is the level of the market as a whole, that is, the average of the indices weighted by the services' audience shares. For web-oriented tools, the index reads as activity: users of Qwen, Perplexity, DeepSeek and Grok generate 2.4–4.7 times more traffic per person than average — these are narrow, active audiences. ChatGPT’s index is lower (1.9) because part of its traffic goes through VPNs and is attributed to other countries. For Alice AI and GigaChat, the index reflects channels rather than activity: chat use from Yandex Browser and the Yandex app, from Salute and from the banking app does not show up in measured traffic. The result is two modes of use: an embedded assistant with a broad audience and short interactions, and a standalone tool with a narrow audience and long sessions [23].
Index = the service’s share of traffic (a16z, web and apps, January 2026) ÷ its share of the weekly audience among the same eight services; both sets of shares sum to 100%, so the average of the indices weighted by audience share equals 1.0 (the simple average of the eight indices is about 2.3). Label: index · share of traffic / share of audience. Gray: services whose main traffic comes through embedded use and is not captured by measurement; for them, the index reflects invisible channels rather than activity.
a16z, based on Similarweb and Sensor Tower data, January 2026 [23]; geoboost calculation.
The working week sets the rhythm of use. According to Yandex logs for January–April 2026, work requests make up 23.4% of all requests to Alice AI from computers (26% on weekdays, 17% on weekends) and only 4.4% of those from smartphones and tablets. Anthropic observes the same pattern worldwide: the share of personal conversations with Claude rises from 35% on weekdays to 50% on weekends. The daily audience of work tools has a pronounced weekly cycle; for everyday assistants, the cycle is weaker [15] [63].
The source of task data is a VCIOM survey (June 2026) in which users of each of five services selected their tasks from 14 categories. To obtain a comparable structure, the 14 categories are grouped into six macro-categories by the nature of the task, and the shares of mentions are normalized to 100% for each service. The assumption is simple: if 19% of ChatGPT users mention programming and 61% mention information search, then the first task carries a third of the weight of the second in the service’s usage structure. This is a proxy for the structure of use; a check against logs follows below.
A category’s share of all task mentions by the service’s users (VCIOM, 14 categories, multiple choice). Assumption: the mention profile is a proxy for the structure of use.
VCIOM, June 25–27, 2026, n=3,209, % of each service’s users [1]; grouping and normalization by geoboost (category mapping in the appendix).
Search and explanations make up the largest category for every service: from 22% to 32% of mentions. The differences begin further down. For Alice AI, everyday life, health and money comes second; for Perplexity, data, automation and code; for ChatGPT and DeepSeek, analytical tasks weigh more than twice as much as for Alice AI (17% and 19% versus 8%). Search and everyday tasks account for 58% of mentions for Alice AI, versus 41% for ChatGPT and 44% for Perplexity. In its profile, GigaChat is closer to Alice AI than to the foreign services [1].
Normalized shares of the three contexts. Source data: the share of the service’s users who use it for personal, study and work tasks (multiple choice).
VCIOM, December 13–15, 2025, n=3,239 [2]; normalization by geoboost.
The context of use confirms this split: for Perplexity, ChatGPT and DeepSeek, work tasks account for about a third of the context; for Alice AI, 21%, with personal tasks making up more than half. For comparison, worldwide, 27% of ChatGPT messages are work-related (OpenAI, June 2025), as are 46% of conversations on Claude.ai (Anthropic, November 2025); the Russian survey estimates fall between the two [2] [24] [42].
Logs provide a check on the survey proxies. Yandex classified 100,000 anonymized requests to the Alice AI chat using a methodology adapted from OpenAI. In the largest category, the survey and the logs agree: search accounts for 32% of mentions and 36% of requests. They diverge in two places. Conversation, games and media content make up 37% of requests in the logs but 13% of mentions in the survey: users do not regard chatting with an AI, games or image generation as tasks and do not name them. Everyday life, health and money, conversely, is a task that is common by number of people but rare by number of requests: 26% of mentions and only 7% of requests [14].
Yandex logs: 100,000 anonymized requests to the Alice AI chat, September 2025–January 2026, classified with an adapted OpenAI methodology. OpenAI logs: about 1 million messages from ChatGPT users, May 2024–June 2025 (distribution as of July 2025).
VCIOM [1]; Yandex [14]; Chatterji et al., NBER [24]; category matching by geoboost, see the appendix.
For ChatGPT, such a check is possible only against OpenAI’s global data: writing and translation account for 26% of messages there, versus 16% of mentions in Russia, while data and code, conversely, weigh three times as much in Russia — the Russian ChatGPT audience is skewed toward professionals who have overcome the access barrier. According to the logs, the share of work requests to Alice AI is about 10% (23.4% from computers and 4.4% from smartphones, with mobile devices accounting for 70% of requests), versus 21% work context in the survey: work use is noticeable by number of people and small by number of requests [15] [24].
Between 51% and 71% of each service’s users use it to search for information, and this is the top task for all five. Alice AI is called a search engine by 60% of its users, ChatGPT by 42% and Perplexity by 52%. The assistant has taken over the first step of search, formulating the question; the click-through to a website remains outside the dialogue. For brands, this means that presence in an AI answer is becoming part of search visibility [1].
For Perplexity, ChatGPT and DeepSeek, work accounts for about a third of the context of use each, and 47% to 58% of their users call them a work tool. For Alice AI, work requests make up about 10% of all requests in the logs — and given its weight (30.9 million weekly users), work requests across the market as a whole amount, by a rough estimate, to around 20% and are unlikely to exceed a quarter (assumption G2). Mass use in Russia remains centered on everyday life [1] [2] [15].
Conversations, games and media generation account for 37% of requests to Alice AI in the logs and 13% of mentions in the survey. Image “animation” in Alice AI alone has been used 360 million times since launch (data as of Q1 2026). Survey-based task classification systematically misses scenarios that users do not consider tasks, so survey data must be adjusted with logs to estimate the volume of requests [14] [1] [10].
Alice AI’s reach is 3.7 times larger than ChatGPT’s, while its intensity per weekly user is almost five times lower than that of ChatGPT worldwide (5.3 versus 25 requests a week; about 18 for Russian ChatGPT users, by estimate). GigaChat, with more than 20 million monthly users, received about 4.8 requests per month per average user. The “embedded assistant” model yields a broad audience with low intensity; the “standalone tool” model, a narrow audience with high intensity. These are different businesses with different per-request economics [13] [16] [24].
Service audiences differ more by age than by gender. Almost half of the market’s daily audience is under 25, the work volume comes from professionals aged 25–44, and reach is provided by default Alice AI users.
This section answers two questions: who makes up each service’s audience, and who forms the active core of the market. The first is answered by recalculating OMI data. The survey gives the share of people within each group who use a service — for example, 73% of men and 69% of women use Alice AI; multiplying by the sample structure gives the composition of the service’s audience. For the second question, the analysis draws on Mediascope daily and monthly reach by age, OMI frequency data and surveys of working people.
Share of each age group among the service’s users aged 18–55; each row sums to 100%. Calculation: share of the group using the service (OMI) × the group’s share of the sample (22/28/28/22%).
OMI, March 2026, n=2,328 [4]; composition calculated by geoboost. The 56–65 group (boost sample, n=102) is excluded because of its high margin of error.
Share of men and women among the service’s users. Calculation: share of men and of women using the service (OMI) × sample structure (47/53%).
OMI, March 2026 [4]; calculation by geoboost. For comparison: according to OpenAI’s name classifier, by July 2025, 52% of active ChatGPT users worldwide had typically feminine names [24].
Age profiles diverge more than one would expect from services with the same function. More than a third of ChatGPT’s audience is under 25 (35%), and almost two-thirds is under 36; for Alice AI and GigaChat, the distribution mirrors the structure of AI users overall, with a slight tilt toward the 36–45 group. Character.AI and Claude have the youngest audiences. By gender, ChatGPT is the surprise: women make up 61% of its users, which matches OpenAI’s observation that feminine names predominated among active users by 2025. A male skew persists for Grok (61% men), Qwen and Perplexity [4] [24].
Share of the age group using the service ÷ share among all AI users × 100. A value of 160 for ChatGPT in the 18–24 group means that young people use it 1.6 times more often than average.
OMI, March 2026 [4]; calculation by geoboost.
The active core is described by three measurements of different kinds:
Estimate for April 2026. Monthly and daily reach by age (Mediascope) × group size. People aged 12–24 make up 17% of the population and 43% of the daily audience.
Mediascope, April 2026: 12–17 — 30% daily / 77% monthly; 18–64 — 12% / 54%; 65+ — 3% / 21% [7]; profile within 18–64 based on data for October 2025 and February 2026 [5] [6]; population structure: World Bank, 2024 [39]; calculation by geoboost (assumptions E1–E2).
Applying daily reach to the size of each age group gives the composition of the daily audience. People aged 12–24 make up 17% of the population aged 12+ and 43% of the daily audience; the 45+ group accounts for half of the population and a fifth of the daily audience. The higher the frequency threshold, the younger the audience: age distinguishes habit more strongly than most other characteristics, along with the level of digital skills. Yandex logs add style: with age, the share of instructions (“do this”) halves, from 33.6% in the 18–24 group to 17.2% among those aged 55+, and questions come to predominate [14].
These data add up to four profiles. The first two are the active core: by frequency (young people) and by volume of work requests (professionals). The third is the mass audience that provides Alice AI’s reach. The fourth is the older generation, with the lowest reach and infrequent use; the rise in its monthly reach from 11% to 21% in 2026 is mainly explained by the expansion of Mediascope’s measurement scope (E1), not by a change in behavior. The profiles are built on aggregate statistics; the task examples in them are taken from survey categories, and the sources contain no user quotes.
Shopping is a small but already measurable share of requests to AI tools. This section sets out the scale and introduces the topic of geoboost’s next study.
The two estimates of the share of shopping requests differ fourfold, and both are correct. A strict classification of Yandex logs for September 2025–January 2026 assigns about 3% of requests to Alice AI to the “products and services” subcategory; for ChatGPT worldwide, 2.1% of conversations are about products (OpenAI, 2025) — these are direct questions about specific products. Yandex’s broad definition includes comparing, selecting and advice on choosing, and yields 12% shopping tasks in the chat (August 2026). Applied to the weekly volume of requests to Alice AI (164 million), this gives 5–20 million shopping requests a week. Requests are not people: one person asks several questions, and the number of shopping requests per user is unknown, so this figure cannot be converted into a number of shoppers. For comparison, e-commerce traffic via Alice AI grew 4.6 times over the year, and 17% of Russian e-commerce is connected to Yandex’s transaction infrastructure [14] [43] [44].
The influence on purchases is spread unevenly along the funnel. According to a survey by Yandex and RBC Market Research, 35% of online shoppers use AI when choosing products; according to an AIMonitor.pro survey (sample not disclosed), 45% of respondents have bought something on AI’s advice at least once, 5% do so regularly, and 96% double-check the recommendations. Data Insight analyzed 38 real scenarios: in all of them, AI helped with the choice; in 13, the user got as far as the cart; and in two, the user completed the purchase through the assistant. The same study forecasts that agentic scenarios will account for 0.4–1% of Russian e-commerce in 2026–2027 and 7–11% (2–3 trillion rubles) by 2030–2032 [48] [49] [45].
Global data show the same pattern: little traffic, high value. According to a peer-reviewed study of 973 e-commerce sites, traffic from LLMs in 2024–2025 was less than 0.2% of all traffic; according to Adobe, by May 2026 AI traffic to US retail sites had grown 14-fold since October 2024 and converts 54% better than other traffic. What matters for brands is presence in the answer, while the number of transactions inside the assistant is secondary for now: in a Sidorin Lab and BrandFound test of the banking vertical, AI models named a specific bank in about 76% of answers [46] [47] [50].
CEO of geoboost, managing partner at strategai, AI lecturer at HSE University
ex-Chief AI Transformation Officer at Sber, ex-Senior Associate Consultant at Bain & Company
CPO of geoboost
PR Lead, geoboost & strategai
All calculated values in the study can be reproduced from the assumptions listed below and from published data. Ranges in parentheses were used for the lower and upper bounds of the estimates.
Limitations. Survey data (VCIOM, OMI) describe internet users aged 18+ based on self-reports and overstate frequency relative to panel measurements; the Mediascope panel does not see chatbots in messaging apps or integrations into search; Similarweb and Digital Budget cover the web only; the Yandex and Sber reports do not break down by country and are not independently verified. The GigaChat data (20 million MAU, 800 million requests) are taken from Sber’s annual report as reported by a secondary source and have not been confirmed by the primary source. The Similarweb Pro engagement data were obtained through geoboost’s paid subscription. The Mediascope breakdown by service is current as of October 2025; no more recent breakdown has been published. Similarweb’s public pages show visits over three months and model geography from a panel, so visits from Russia were converted to a monthly basis and are used for trends and checks, not as a direct measure of audience; the shares on the a16z chart were read off the image with an accuracy of about 0.5 p.p. Estimates for Gemini refer to the period before August 14, 2026. The final estimates are model estimates for the reference period (Q2 2026): the sources refer to different dates — the main audience data to October 2025–July 2026, trend cross-checks and intensity data to 2025, web engagement and shopping data to August 2026; alignment to a single time window was done only where trend data exist (D1, D4, D5), so the mismatch in dates is a source of additional uncertainty that has not been quantified. Estimate ranges are the spread across scenarios and methods, not confidence intervals.
Which sources and signals determine whether a brand appears in answers from Alice AI, ChatGPT, GigaChat and DeepSeek, how recommendations differ across categories, and what converts into clicks and orders. Data: monitoring of answers across seven engines and public studies from 2025–2026.