App intelligence firm Appfigures published a report at the start of May 2026 that quietly invalidated the working assumption behind every "AI chatbot app" pitch deck of the last three years. Image model releases — not chatbot upgrades, not voice modes, not reasoning improvements — are now driving 6.5x more incremental app downloads than the model upgrades the entire industry has been organized around.
The headline example is Nano Banana. When Google released its first image model under the Gemini brand last August, the Gemini app picked up over 22 million new downloads in the following 28 days — a 4x lift on its prior baseline. ChatGPT did the same thing earlier: its GPT-4o image release in March 2025 drove a 12 million download surge in the same 28-day window. Every chatbot release since — every "smarter" model, every new context window, every voice update — moved the download needle a fraction as much.
If you operate a standalone AI chatbot app, that is the story you have been told you should care about. If you operate a chatbot embedded on a website, the actual story is different, and better for you.
What the Appfigures data actually says
The Appfigures report tracked app store performance across the major consumer AI apps from early 2025 through April 2026. The pattern is consistent enough to look like a category rule, not a product quirk.
| Model release | Download lift (28 days post-launch) | Multiple of baseline |
|---|---|---|
| ChatGPT GPT-4o image (March 2025) | +12M | +1.5x |
| Gemini Nano Banana (August 2025) | +22M | +4x |
| Gemini Nano Banana 2 (Feb 2026) | Significant, exact figure varies by region | New top-ranked image model |
| Average chatbot reasoning/voice upgrade | Single-digit millions | Marginal |
The aggregate: image launches drove 6.5x the incremental installs that chatbot-feature launches did over the same windows. That ratio is not random noise. It is the consumer market voting on what they think AI is for.
Revenue tells a different story. Nano Banana's 22 million downloads converted to roughly $181,000 in gross consumer spending in the same 28-day window. ChatGPT's GPT-4o image launch was a much better commercial event — closer to $70 million in spending lift over its 28-day window — but most of that revenue is people upgrading to ChatGPT Plus for everything ChatGPT does, not for image generation specifically. The honest read: image generation is a top-of-funnel acquisition channel, not a monetization product. It is what gets a person to install an AI app for the first time. It is not what makes them pay.
That distinction is the entire blog post. Hold onto it.
Why standalone chatbot apps stopped growing
For three years, the implicit market assumption was that AI chatbots would replace search, that the chat interface would be the new browser, and that every household would have one general-purpose AI app the way every household has Google. The race was to be that app.
A reasonable amount of that prediction has come true. ChatGPT crossed one billion users last year. Gemini, Claude, and Copilot all have meaningful consumer install bases. The category exists. What stopped happening is new growth — at least, growth driven by the chatbot capabilities themselves. The category appears to have reached the people who want a general-purpose AI app. New growth in those apps is coming from a different motivation: people who want to make a picture of their dog as an astronaut, or restyle a selfie, or generate a birthday card for a friend.
The implication for anyone building a standalone chatbot app is uncomfortable: chatbot capability improvements are no longer the growth lever. Image generation is. Video generation will be next. The marginal new install of ChatGPT or Gemini in 2026 is probably someone who saw a viral image trend on TikTok, not someone who heard the latest model is smarter at math.
This is a different problem than "AI is plateauing." AI is not plateauing — capability gains are if anything accelerating. What is plateauing is the marginal value of another generic chatbot app to a consumer who already has one.
Why this is good news for site chatbots
A site chatbot — the kind embedded in the corner of an ecommerce store, a SaaS onboarding flow, a support page — is a different product than a standalone AI app, in a way most people miss until they look at the data above.
The standalone chatbot app competes for the user's attention. It needs to be opened. It needs to displace a search query, a Slack message, a Google Doc, or a phone call. It competes against ChatGPT for the same finger-tap. That is a brutal competition, and the Appfigures data says it is one that capability improvements alone do not win.
The site chatbot competes for the user's commitment in context. The user has already navigated to your domain. They are not deciding which AI to ask — they are deciding whether to ask one at all, given that you have put one in front of them. That decision is dominated by the question "is this thing going to answer my specific question about this specific page in under five seconds," not "is this AI smarter than Gemini at coding."
The two products look superficially similar — both are conversational, both run on LLMs, both have a chat UI — but the unit of growth is entirely different.
| Standalone chatbot app | Site chatbot |
|---|---|
| Wins on virality, novelty, free tier | Wins on context, speed, completion |
| Acquisition: app store, social media | Acquisition: the user is already here |
| Competes against ChatGPT for installs | Competes against the exit button |
| Grown by image and video, not chat | Grown by site traffic and intent |
| Monetized via subscription | Monetized via conversion lift |
| Top-of-funnel for the AI ecosystem | Bottom-of-funnel for your business |
If you have been worried that the ChatGPT app is going to eat your site chatbot's lunch, the Appfigures data is the first quantitative reason to relax. ChatGPT is not investing in being a better generic chatbot — that is not what is driving its downloads. It is investing in being a better image studio, a better video studio, a better creative tool. Those features are excellent. They are also not in competition with the chatbot doing customer support on your pricing page.
What you should actually take from this
Three concrete shifts in how to think about your site chatbot, given the May 2026 data:
1. Stop comparing your chatbot to ChatGPT's chatbot. The relevant comparison is your chatbot vs. the user's other option, which is leaving your site without an answer. The bar is much, much lower than "as good as a frontier model" — it is "answers this person's specific question about this specific product." A well-trained site chatbot on a 95-page knowledge base outperforms ChatGPT on questions about your business by definition, because ChatGPT does not know your business. We unpacked the mechanics of this in how to train a chatbot.
2. Add image as an input, not an output. The Appfigures data shows that consumers love generating images. That does not mean your site chatbot needs to generate images — it almost certainly should not. But it strongly suggests visitors are comfortable, and increasingly expectant, of uploading an image and asking a chatbot a question about it. An ecommerce chatbot that accepts a photo of a torn jacket and identifies the SKU, a real estate chatbot that takes a screenshot of a listing and answers questions about the neighborhood, a support chatbot that accepts a photo of an error message — those are the features that match how 2026 consumers have been trained to interact with AI. They reach for the camera, not the keyboard.
3. Lean into the channel you actually win. A site chatbot has structural advantages that no standalone app can replicate: it knows what page the user is on, what they clicked to get there, what's in their cart, what plan they're on, what their last support ticket said. Use those signals. A chatbot that starts a conversation with "I see you're on the Enterprise pricing page and you've been here for three minutes — anything I can help clarify?" is doing something ChatGPT cannot do, will not do, and is not trying to do. That is the moat.
The protocol layer is reinforcing this
The picture above gets sharper when you overlay what is happening at the agent layer. ChatGPT now sends roughly 3.6x more crawl traffic to commerce sites than Googlebot does, per Search Engine Journal — which we covered in AI shopping agents are already browsing your site. Google's Gemini Spark, launched May 19, is a 24/7 personal agent that goes out and interacts with third-party sites on a user's behalf — covered in our Spark post from yesterday.
What both of those stories share with the Appfigures data is a consistent direction: the future of the consumer AI experience is not one tab, one app, one chat window. It is dozens of agents and assistants reaching into sites, products, and tools where the actual work is. The site chatbot is one of those tools. The standalone chatbot app is one of those agents. They are not competitors — they are different positions in the same flow.
A concrete example: visual support
Pick the simplest case of where this matters. A customer of an ecommerce store has received a product with a defect. In 2024, the support flow was:
- Customer opens a ticket
- Support agent asks for a photo
- Customer emails a photo
- Support agent identifies the issue
- Resolution
In 2026, the version that beats that flow on every metric is:
- Customer opens the chatbot on the support page
- Customer drags the photo into the chat
- Chatbot identifies the defect, the SKU, the warranty status, and the policy
- Chatbot offers replacement, refund, or partial credit
The second flow takes 90 seconds instead of 36 hours. It does not require a human. It does not require the customer to type a description of what they are looking at. It uses the visual modality that the Appfigures data says consumers are now reaching for first. This is the kind of feature that turns a chatbot from a cost center (deflecting tickets) into a margin lever (closing returns faster than human agents could).
Agentkit's image upload, available on all paid plans, exists for exactly this case. The model behind it sees the image, reasons over your trained knowledge base, and decides what to do. The trained knowledge base is the part nobody else has — it is your product catalog, your warranty policy, your support history, your tone. None of that is in ChatGPT. None of that is in Gemini. That gap is your moat, and image upload is how you make it concrete to the visitor on the page.
The bigger picture
The Appfigures number — 6.5x — is not really a story about image generation. It is a story about where consumer AI is going.
The consumer AI market is splitting into two distinct categories. One category — the standalone AI app — is becoming a creative tool, with chat as the input and image/video/audio as the output. Growth in that category is driven by visual novelty and viral moments. Revenue follows users who treat it as a creative subscription, the way they used to subscribe to Photoshop.
The other category — the embedded AI assistant — is becoming an infrastructure layer that sits inside every product, app, and website. Growth in that category is driven by completion rate, conversion lift, and time-to-resolution. Revenue follows the business value the assistant produces for the operator who deployed it.
If you operate a website, you are not building the first kind. You are deploying the second kind. The Appfigures data is the clearest signal yet that those two markets have decoupled — and that the second one, the one you are in, is not subject to the same growth ceiling.
Build for the user who is already here, not for the user you want to acquire from ChatGPT. The chatbot embedded on your site is the chatbot they meet. The 95+ languages, the trained knowledge base, the lead capture, the form submissions, the image upload — these are not features you should benchmark against a frontier model. They are features you should benchmark against the next-best alternative, which is the user leaving without an answer.
That is the bar. It is much lower than the noise from the standalone app market would suggest. You are probably already clearing it.
No credit card required.
