Designing AI into Opera for Android, 2023–2026
In May 2023, Opera put a free AI assistant called Aria into its Android browser as a chat window. Three years and 16 releases later, AI lives in the address bar, right next to Google. This case study is about how user research moved it there, and why we let people choose between search and AI instead of choosing for them.
At a glance
- My role: Design lead for Opera for Android. I led the Android side of this work from the first chat to the AI omnibox, working with the iOS, desktop, backend and marketing teams.
- Product: Opera for Android, a browser with 53M+ monthly users and a 4.7 rating on Google Play
- Timeline: May 2023 to March 2026, 16 releases (v75 to v96).
- Team: Android and iOS designers in Sweden, the desktop team in Poland, a shared AI backend team, and marketing
- Methods: cross-team concept workshops, user interviews, usability tests with prototypes, in-app answer feedback, usage analytics, coded prototypes
The challenge
When ChatGPT arrived, Opera set a clear direction: AI belongs inside the browser. That raised two design questions we kept answering for three years:
- What should AI in a browser do that a chatbot app doesn't?
- Where should it live in a product that people open for something else entirely?
1. Ship a chat, fast
2023 · v75 to v80
Speed to market mattered, and the simplest credible product was a chat assistant built into the browser. It was free, and aimed at people who didn't have ChatGPT or Gemini installed on their phones.
A cross-platform AI feature didn't fit how we were organized. Desktop is built in Poland, and Android and iOS in Sweden. We started with joint workshops where desktop, mobile and marketing brought all their concepts to one table before anyone drew screens. One backend team served every client. It combined several LLMs with live web results, so answers were current rather than frozen at a model's training date.
One technical decision shaped the design work for years: in v80 we moved Aria to a web UI inside the native app. The same components could then serve Android, iOS and, later, Opera Mini.
The first version was a chat behind the profile menu. You tapped your profile, then "Aria browser AI", then signed in with an Opera account.
2. The empty-chat problem
2024 to 2025 · v81 to v89
Many people were skeptical of AI, and those who did open Aria often didn't know what to ask. A blank chat gives no hint of what it can do. Professionals know what to ask an AI, but most of our users came to the browser to search, and they needed to see what was possible.
We answered in steps:
- v81: suggested prompts in every new chat, and "Ask Aria" on selected text, offering Tell me more and Summarize. Voice input too.
- v82: summarize any page from the menu, and a customizable toolbar so Aria could sit one tap away
- v83 to v85: the capabilities people expected from ChatGPT: image generation and image understanding. We also removed the login requirement and added sources to answers.
- v88: a new welcome screen with a carousel of real use cases that changes over time
- v89: follow-up questions
The use-case onboarding raised usage and engagement.
3. A product inside a product
Even with better onboarding, adoption stayed low. Aria was a separate app inside the browser, and people didn't relate to that model. Super apps work in some markets; for us, it didn't.
Over time we kept adding ways in: the profile menu, a Speed Dial icon, a homescreen search widget, a toolbar button, and a homescreen widget with an app shortcut. Each one made Aria easier to reach. None of them changed why people open a browser in the first place.
That led to the insight the rest of the project rests on:
You don't open a browser to use AI. You open it to find information or to reach a site you know, and both start in the omnibox, the combined address and search bar.
Looking something up is the same task whether you go to Google or to AI. Our assumption was that typing a question and getting an answer would beat a page of links. At the time, Google search had no AI mode, so the gap looked real. But it was still an assumption, so we tested it.
4. Testing AI in the omnibox
We ran user interviews and usability tests with prototypes of AI in the omnibox.
Two mental models, not one
The research split people into two clear groups:
- Daily AI users were already on ChatGPT, Claude or Gemini and wanted AI close at hand.
- Skeptics, a large group, preferred traditional search and wanted control over what they found.
The skeptics had a point. Models hallucinated, and answers could be out of date. On the backend, we mixed several LLMs with live search results to improve accuracy. On the design side, we had to treat skeptics as a group to design for, not a problem to convert. The wider data agrees with what we heard: the 2025 Edelman Trust Barometer (opens in a new tab) puts trust in AI at 49% across 28 countries.
The concept testing rejected
Our first idea was an automatic mode: whatever you type, the browser decides whether to send you to Google or to our AI. Testing said no. People wanted to choose.
| Concept A: the browser decides | Concept B: the person chooses | |
|---|---|---|
| How it works | Every query is routed to Google or to AI automatically | Google stays first; Ask AI sits beside it as a chip |
| Automation | Full | One nudge: questions starting with "what" or "how" suggest AI more prominently |
| In testing | Rejected: people wanted control | Shipped in v92 |
5. What shipped
October 2025 to March 2026 · v92 to v96
- v92: the new omnibox, AI answers in a full tab, PDF and document upload, and a rebuilt AI backend
- v93: attach files straight from the omnibox, and Ask AI about the page you're on or about your downloads. The current page comes along as context, with suggested actions like Summarize page and Translate.
- v95: Think deeper for harder questions, and the old chat bottom sheet retired in favor of AI in a tab
- v96: one voice input for both search and AI
We stated privacy plainly in the product: only the current tab is shared, data is encrypted, kept for up to 30 days, and never used for training (Opera blog, December 2025 (opens in a new tab)).
On desktop, the same thinking went further. Opera One has AI in the address bar, and Opera Neon is an agentic browser that carries out tasks on websites, asking permission first.
How people reacted
AI features found fans in Brazil (image generation above all), Turkey, India, the US and the UK. Germany was more reserved, and in Poland we saw some skeptics become fans. This matches Edelman's finding that trust in AI is higher in developing markets than in developed ones.
Impact
Moving AI from a separate chat to the omnibox and browser context menus had a positive effect on user adoption (daily active users) and engagement (number AI requests per user).
How we worked
Prototyping before handoff. We made a lot of motion design in After Effects and Figma prototypes to feel an interaction and its animation before it reached the codebase. That let us hand developers a decision rather than a static spec. Later, when Claude Code became available for us, interaction prototypes took hours instead of days, and larger test builds took days instead of weeks. I described this shift in Inside Opera's Evolving Design Loop (opens in a new tab).
Fundamentals matter more, not less. AI speeds up making things, but color, typography, layout, navigation and motion are still judgments a designer has to make.
Rituals that keep a distributed team close. We hold Design Talks every Friday, a design meetup twice a year, and dedicated Slack channels for designers. We also play drawing games (30 Circles, Skribbl, Exquisite Corpse, Gartic Phone) and run vibe-coding hackathons and concept workshops.
What I learned
- Put AI where the task already happens. People don't open a browser to talk to AI. As long as Aria was a separate chat, adoption stayed low, however many ways in we added.
- Offer AI; don't decide for people. Testing killed automatic routing. Keep people in control, and nudge only where intent is clear.
- Design for the skeptics, not only the fans. Sources, follow-up questions, plain privacy wording and keeping traditional search first are all design work.
- Show people what to ask. A blank chat stalls people. Real use cases get them started.