AI chatbot market share: usage is not citation

The short version

The Western chatbot market is ChatGPT, Gemini, Claude and Perplexity, plus Google's AI Overviews inside Search. I don’t think anyone’s yet coined a collective term for them but let’s call them “the big four”.

Usage data tells you which assistants people visit whilst referral data tells you which assistants send people to websites. Those aren't the same market.

ChatGPT has 53.9% of global chatbot web visits, but 77.92% of chatbot referrals. Perplexity ranks sixth for visits and third for referrals. Grok gets plenty of press and sends no measurable traffic.

If you want broad consumer reach, watch usage. If you want to understand where citations and visits can come from, watch referrals. If you’re a GEO/AEO consultant you’re going to want to watch both, but don't confuse them.

Which bots actually get used?

Similarweb's May 2026 figures measure visits to chatbot websites. They don't include every interaction inside an app, operating system or social platform.

A year earlier, ChatGPT held about 76% of chatbot visits. Its share has since fallen to roughly 54%.

That doesn't mean people are abandoning it. Far from it, ChatGPT reached 900 million weekly users in February 2026. Simply, the whole market is growing, and its competitors are growing faster.

Market share headlines often turn a changing denominator into a product obituary. ChatGPT is losing share while adding users. Both statements can be (and are) true at the same time.

Usage also hides distribution. A Gemini answer inside Google or Android isn't necessarily a visit to gemini.google.com. A Copilot interaction inside Microsoft 365 may never reach a public chatbot page. Grok's audience figures predominantly include people who encountered it inside Twitter/X.

There isn't one clean number called chatbot market share. There are visits, users, queries, survey answers and referrals. A table can easily become misleading without specifics.

Who actually sends referrals?

Statcounter measures traffic that chatbots send to websites. That makes it a better operational lens for generative engine optimisation, although it still isn't a complete citation count. A cited page can influence an answer without receiving a click, which is precisely why Google is getting so much grief from publishers over its AI overviews using their content and giving them nothing back.

Statcounter’s July 2026 worldwide figures look very different from the general usage table.

ChatGPT, Gemini, Perplexity and Claude produce 96.9% of measured chatbot referrals worldwide. The same four produce 90.9% in the US, 91.8% in the UK, 95.2% across Europe and 90.6% in Australia.

That regional consistency is why I use a big four monitoring set. As Dog on the Table's sole operator, I run this monitoring myself (with a bunch of delicately wired automations). It covers almost all observable referral behaviour without pretending every newly announced LLM deserves equal attention.

Perplexity is the useful contradiction. It has only 1.3% of chatbot web visits, yet supplies 5.88% of referrals. It ranks sixth in usage and third in outbound traffic. And honestly, it’s great. More people should use it. It hallucinates far less than its rivals.

That happens because Perplexity is built around research, retrieved sources and onward clicks. It can still guess or make things up, but it’s less likely to because it cites the sources it uses. Usage share measures audience size. Referral share also reflects product behaviour.

So clicks are emigrating unevenly. An LLM assistant can have millions of users and little interest in sending them anywhere. Ever notice how your prompts are often answered with suggested follow up questions or actions? These bots want to keep you on platform, continuing to engage. Sending traffic to other websites isn’t helpful for them in the same way as Google made its billions as the web’s biggest single source of traffic.

Which four assistants do I measure?

ChatGPT

Obviously. ChatGPT remains the default assistant and for many is as synonymous with AI as Google is with search. Its 900 million weekly users give it reach that no standalone Western competitor currently matches.

It also dominates observable referrals. For most brands, excluding ChatGPT from an AI visibility audit would make the result decorative rather than useful.

Gemini

Gemini is the fastest serious riser. Google reported 1 billion monthly Gemini users in August 2026, while Similarweb recorded 450% year-on-year visit growth.

The company claim and the web visit figure measure different things. Together, they show momentum backed by Google's existing distribution across Search, Android and Workspace. Gemini can be everywhere because Google is everywhere. Sadly its ubiquity isn’t matched by its quality at present: only one Gemini model barely cracks the top ten of the Arena.ai text model leaderboard at the time of writing.

Claude

Claude has the strongest quality-per-user story. Its chatbot visits grew 855% year on year, while its referral share rose from 0.30% in April 2025 to 2.91% in March 2026, almost 10 times higher.

Its audience remains smaller than Gemini's but its users appear unusually willing to research, build and follow sources. That behaviour makes Claude more important for visibility work than a simple user table suggests.

Perplexity

Perplexity is a different product wearing a familiar chat box. Its chief executive reported 780 million queries in May 2025, and it treats citations as part of the answer rather than optional furniture.

It also shows why files made solely for bots rarely solve the real problem. As I found when I tested the llms.txt file that nobody reads, visibility comes from usable evidence across sources. It doesn't come from announcing that a source exists.

AI Overviews

AI Overviews isn't a standalone chatbot, but leaving it out would miss Google's largest answer surface. Google said AI Overviews reached 2 billion monthly users in July 2025 and by July 2026, AI Overviews appeared on about 43% of US searches. Ofcom found them on roughly 30% of UK searches.

Those percentages don't belong in the chatbot tables because the denominator is searches, not chatbot visits. They belong in the monitoring plan because people still see the answers.

Though it’s powered by Gemini, users aren’t interacting with it in the same way and the queries it fields and responds to are far more like traditional keyword based search terms than the conversational format that Chatty G and co are nurturing.

Do Grok and Buzz change the market?

Grok Bot entered beta on 11th August 2026. It offers always-on cloud agents with separate logins, specialist roles and a coordinating chief-of-staff bot at about $120 a month. Buzz launched on 21st July 2026 as an open-source team workspace where each agent has a cryptographic identity.

They’re not the same but they’re generating a similar amount of hype at the moment and the multi-agent pattern matters because specialist agents can work in parallel while one agent coordinates the result. Agentic AI use is growing but it’s far from mainstream. This is the stuff of power users but isn’t something I need to monitor in the same way as LLM prompt responses.

Buzz and Grok Bot are currently making waves but open-source agent users have run versions of this fan-out structure for some time, which is a topic for another day. It began with OpenClaw, got another round of hype when Hermes Agent broke onto the scene and now the market is getting crowded with differing types of harnesses and ways of interacting with multi agent workflows. But I digress. It doesn't change which assistants answer most public questions or send most website traffic.

Grok's reported 117 million monthly users include people touching it inside Twitter-I-mean-X, while its standalone usage fell during 2026. I wouldn't use Grok for all the reasons that don’t need explaining, and the referral data means I don't need to argue the point here. Press attention isn't usage, and usage isn't citation.

What about the Chinese models?

The Chinese market is large and mostly separate. Doubao has 382 million monthly users, Qwen has 167 million and DeepSeek has 130 million.

By spring 2026, Chinese open-weight models supplied about half of OpenRouter's routed tokens (OpenRouter being the best-known platform through which AI power users can pick and choose from all the different models to route queries and tasks through). That shows their importance to developers and agent systems. It doesn't show Western consumer adoption.

I use these models more heavily than a normal person does for agentic work:

• MiniMax M3 is a reliable analysis workhorse. It excels at parallel research scouting, comparative judging and pre-mortems, but needs precise instructions. I give it grunt work and then have a smarter bot review it afterwards.
• Kimi K3 catches coding mistakes that other models miss. It also has a sharp eye for copy improvements. This is a very smart bot.
• GLM 5.2 routes my trusty in-house assistant DogBot each day. It's fast, cheap and dependable plumbing that is remarkably capable.
• DeepSeek v4 supports research briefs and citation checks. Its cheaper Flash tier verifies statistics for pennies.
• MiMo v2.5 handles mid-pipeline analysis and preserves local detail that other models drop. MiMo v2 Flash previously ran DogBot each day, before being retired, and I’m keen to see what v3 brings as I think Xiaomi’s MiMo models are quite underrated.

That isn't normal consumer behaviour. I find it hard to imagine any of these replacing ChatGPT on a typical person's iPhone today, even if they are all cheaper on a token by token basis. But as long as OpenAI give users a free version of ChatGPT, however limited it may be, the majority of users will find little reason to look elsewhere.

DeepSeek produces only 0.03% of worldwide chatbot referrals. Its downloads outside China reached 35 million so far in 2026, compared with 171 million across the whole of 2025.

Price could still change the equation. If ChatGPT starts costing £50 a month minimum and a strong Chinese app offered a paid tier for £10 or less, would Western users move? Lower model prices haven't shifted Western consumer share so far, and I don't know whether a larger gap would.

Is Copilot the likely fifth assistant?

Copilot is the strongest fifth candidate. It has Microsoft distribution, access to the Bing index and a product family that's being merged across consumer and enterprise apps.

The UK data also shows why it’s wise not to rely on surveys alone. Ofcom says 54% of UK adults use AI tools. YouGov gives Copilot an 11.6% UK preference score, just above Gemini at 11.3%.

Yet Copilot supplies 8.15% of UK chatbot referrals, while the big four supply 91.8%. Bundling can flatter awareness and survey use without producing equivalent answer behaviour.

A fifth assistant earns a permanent Dog on the Table monitoring slot when it meets one of these tests:

1. It holds at least 3% of chatbot referrals in the UK, or a client's main market, for 3 consecutive months.
2. It reaches at least 5% of chatbot web visits in that market.
3. It supplies at least 1% of AI referral sessions in client analytics for 2 consecutive months.
4. At least 10% of people say they've used it in a national survey for a market I serve.

Copilot currently passes the referral test in the UK, US and Australia. It fails the other tests. Grok and DeepSeek pass none in the West.

Will one assistant take the whole market?

My lean is no. Search consolidated because the products became interchangeable enough for distribution to decide the winner. LLM assistants are separating around answer quality, retrieval, workflows and distribution.

The trend also runs away from consolidation. The market leader fell from about 76% to 53.9% of chatbot visits in a year, even as its user base grew.

In May 2025, Meta said Meta AI had more than 1 billion monthly users across its apps, yet only 14% of US adults say they use it. It has no measurable referral presence, and Meta has little reason to send people away from its platforms. Its reach rivals Google's, yet its presence in observable answer behaviour is close to zero. Most people have instant free access to Meta AI but nobody wants to use it.

Regulation can redraw the tables faster than product quality. The UK Competition and Markets Authority ordered Google to let publishers leave AI Overviews without leaving Search on 3rd June 2026. The European Commission opened a formal investigation into Google's use of publisher content on 9th December 2025, while Italy blocked DeepSeek in January 2025.

This looks less like search in 2004 and more like one giant alongside 2 or 3 durable specialists. I’m not betting the house on it but I strongly suspect the current multi-model market is unsustainable.

Distribution is the force that could still create a 90% winner. Google holds the strongest card because Gemini and AI Overviews sit inside products people already use. But in my experience its model quality still trails Claude and ChatGPT.

I don't know whether that advantage will ultimately win. The watchlist tests will show movement before the press cycle does. Until then, I'll keep measuring ChatGPT, Gemini, Claude and Perplexity, and I'll keep publishing the numbers.