Kai ChenAndroid users can replace Google or Gemini with Perplexity as their default assistant, then ask it to act across apps. Yet that system role does not necessarily give Perplexity the same reach into apps, stored context, device surfaces, or live search inputs as Google’s own service.
Key takeaways
- AI-assistant competition now spans five interdependent layers: models, system invocation, app and tool access, portable user context, and live search infrastructure.
- Becoming Android’s default gives a rival such as Perplexity a route to users and actions, but it does not guarantee access comparable to Gemini’s integrations with Gmail, Maps, Docs, and device-level surfaces.
- Stored history, preferences, permissions, and routines create switching costs; importing chats into one service helps entry, but it is not equivalent to market-wide context portability.
- AI assistants still depend on upstream search infrastructure for current information, as shown by OpenAI’s use of Bing and reported reliance on Google-derived results.
- Two EU Digital Markets Act decisions make parity a practical engineering test by requiring Google to provide rival assistants and search engines comparable Android access and some Search data—not merely the ability to be installed.
A strong model can win a benchmark. To replace a workflow, an assistant must also open the right app, retrieve current information, preserve context, and complete the task.
Task completion split one model race into five contests
Users initially encountered generative AI as if product and model were the same thing: open a chatbot, type a prompt, compare the answer. Builders broke that frame when they began treating the model as one component in a larger system.
At Google I/O in 2023, Google Assistant had become an afterthought as Bard took the foreground. The old voice-assistant category was being absorbed by generative AI. Later that year, OpenAI announced an Assistants API that let developers combine its models with tools to perform tasks. One shift happened at the product layer and the other at the developer layer, but both moved the assistant beyond text generation.
A model cannot complete tasks alone. It needs external interfaces to retrieve live information, query services, and execute multi-step workflows. Once builders optimized for completed tasks rather than plausible answers, APIs, permissions, and application surfaces became scarce inputs.
Builders were no longer competing in one race. They were contesting five interdependent layers: who supplies the model, who controls invocation, which apps can be called, whose context follows the user, and which search infrastructure provides fresh information.
A default is a route to action, not a preference
Perplexity made the new competitive unit explicit. Its Android assistant launched free in 15 languages, could perform “multi-app actions” such as hailing a ride, and directed users to replace Google or Gemini as their default assistant. The important move was not putting another chatbot in the Play Store. It was seeking the system role from which a chatbot becomes an assistant.
A downloadable app and a selectable default may contain similar models, but they do not have the same route to use. The app waits for the user to find and open it. The default occupies an established assistant position on the device. If that position also carries the ability to call apps, default selection governs access to an action surface, not merely a settings preference.
Google’s own products demonstrate the same mechanism. Gemini on Android was judged more useful than the regular Assistant because of its Gmail, Maps, and Docs integrations, even while it remained frustrating at picking up context. Gemini’s advantage was not simply better prose. It could reach services that already contained the user’s mail, documents, places, and routines.
Pixel’s Magic Cue, call translation, and camera overlays extend the logic. Assistant value can appear inside calls, camera views, and contextual suggestions across apps. The interface is distributed through the device rather than bounded by a chat window.
Regulators testing AI distribution parity should ask whether a rival can occupy a comparable system role, not whether its icon is available to download.
Permissions set an agent’s practical ceiling
Assistants create value through tools: the model interprets and plans, while external systems supply current facts and executable actions. Whichever layer is weaker sets the practical ceiling.
Even a highly capable model cannot hail a ride without permission to reach a ride-hailing app. It can only describe the steps. A less celebrated model with reliable tool access may complete the task.
OpenAI, Perplexity, and Google converged on this architecture from different positions. OpenAI exposed tool-calling through an API. Perplexity organized its mobile product around cross-app actions. Google made Gemini more useful through service integrations and embedded AI in device-level surfaces.
| Competitive layer | What access changes | Observed signal |
|---|---|---|
| Operating system and default | Whether a rival can occupy the device’s assistant role | Perplexity asked Android users to switch their default from Google or Gemini |
| Apps and tools | Which tasks the assistant can execute rather than describe | Multi-app actions and Gemini integrations with Gmail, Maps, and Docs |
| Personal context | How much accumulated history must be abandoned during switching | Gemini added imports of chat history and context from competing AI apps |
| Search infrastructure | Whether the assistant can answer queries that require fresh web information | OpenAI drew on Bing and reportedly on Google-derived results for current events |
A rival can gain nominal access at one layer and remain disadvantaged at another. Regulators assessing parity—and users comparing assistants—have to follow the full route from intent to completed action.
Context makes yesterday’s use part of today’s moat
Assistants accumulate more than conversation logs. History can contain preferences, prior instructions, unfinished tasks, and the explanations that make later requests intelligible. As that accumulation grows, switching means more than installing a different app.
Google acknowledged the competitive value when it released Gemini tools for uploading chat history and context from other AI apps, explicitly making it easier to switch from those services. The feature reduces the cost of entering Gemini. More importantly, its existence identifies context portability as a product variable.
An importer and a market-wide portability mechanism are not the same design. An importer makes one destination easier to enter. General portability would make accumulated context easier to move among services. Both respond to the same structural fact: once an assistant remembers, memory becomes part of distribution.
Google’s delayed migration from Assistant to Gemini reinforces the point. The company pushed its plan to upgrade Assistant on most Android devices beyond its previous end-of-2025 target and into 2026. Owning Android did not make the installed base frictionless to move; Google still had to carry users, device configurations, integrations, and product expectations across.
When users switch, they may leave accumulated context, app permissions, routines, and a familiar path to action. Any practical parity test must account for the cost of carrying those assets between assistants.
Search remains upstream when the answer must be current
Chat changes how information appears, but it does not remove the need to acquire it. An assistant answering questions about news, sports, or other current events still needs a live upstream source.
OpenAI’s behavior makes that dependency visible. In 2024 it was reported to be developing a web-search product partly powered by Bing. In 2025, sources said ChatGPT was partially using Google search results scraped through SerpApi for current-event answers. OpenAI competed with search interfaces while depending on infrastructure supplied by Microsoft and, indirectly, Google.
OpenAI can replace the search interface without reproducing the index, freshness, and retrieval infrastructure beneath it. The assistant changes how results are synthesized while drawing on an existing information-production layer.
Google supplies Gemini while controlling Search infrastructure valuable to rival assistants and search engines. Microsoft occupies a different position: Bing can serve as an upstream input to an assistant built by OpenAI. Perplexity’s Android strategy adds the device and app layers to the same contest.
After using AI search exclusively for a month, the Wall Street Journal characterized it as more of a user-experience revamp than a replacement for blue links. The experiment weakens a simple displacement story while preserving the upstream dependency: assistants can revamp the interface without replacing the infrastructure.
Comparable access turns neutrality into an engineering test
The European Commission’s two DMA decisions move the argument from abstract platform neutrality to practical comparison. They order Google to provide rival AI assistants and search engines comparable access to Android and some Search data.
By requiring “comparable” access, the Commission moved beyond nominal availability. A rival’s presence in an app store establishes that it can be installed. The new test asks whether the conditions under which it reaches users and uses platform capabilities compare with those available to the platform owner’s service.
Regulators can compare several distinct frictions. Can a user select the rival for the relevant system role? Which device capabilities and app interfaces can it use? Can the user carry enough context to make switching practical? Can the service obtain the live information required for current answers? Each answer contributes to the economic outcome.
Google made the same route-to-user issue visible in a different market. In a US court filing, Google said it was ready to add third-party app stores within Google Play. The product category differs, but the bottleneck is familiar: Google controls the route through which another product reaches users.
The DMA decisions do not establish universal assistant portability or prove that users will switch. They concern Google, Android, rival AI assistants, rival search engines, and some Search data. Their structural importance is narrower: regulators can now test access against the platform owner’s access rather than deem it adequate whenever a rival can technically be installed.
An open download layer can hide unequal action layers
Google did not wholly foreclose rivals from Android. Perplexity reached users through the Play Store, offered its assistant free in 15 languages, and told users how to make it the default. Comparable-access obligations may improve competitive conditions, but they were not the sole reason a rival could reach the platform.
Google also shows why platform control does not guarantee a smooth incumbent victory. Its delayed Assistant-to-Gemini transition left product-transition risk intact. Integrations have to work, context has to carry, and users still encounter the product that ships rather than the strategic diagram.
Rival assistants can obtain live information from Bing or intermediaries connected to Google results, but that access does not establish that users will abandon conventional search. The evidence supports dependence and interface change, not displacement.
A rival that wins the default may still lose every dependency behind it.
The Commission is not equalizing products, models, or user demand. It is asking whether Google’s control over defaults, integrations, switching, and upstream information predetermines the contest before users test product quality.
Rivals still have to build a reliable assistant. If they receive comparable access and users refuse to switch, access was not binding. If Google withholds an essential system role or capability, however, benchmark scores compare models built with different production inputs.
Perplexity can take Android’s assistant slot, but the slot is only the front door. Replacing the default is not the same as replacing the stack.
Frequently asked questions
Can Perplexity replace Gemini as the default assistant on Android?
Yes. Android users can select Perplexity as their default assistant and use it for multi-app actions, but that role alone does not ensure access equal to Gemini’s apps, stored context, or device surfaces.
Why are app permissions and integrations important for AI assistants?
They determine whether an assistant can execute a task or only explain how to do it. Even a strong model cannot hail a ride, retrieve mail, or use another service without the necessary interfaces and permissions.
What do the EU DMA decisions require Google to provide?
The two European Commission decisions require comparable Android access for rival AI assistants and search engines, plus access to some Search data. They do not mandate universal context portability, equalize products, or guarantee that users will switch.
Why do AI assistants still need Google Search or Bing?
Questions about news, sports, and other current events require fresh web information. OpenAI has used Bing and reportedly Google-derived results, showing that a new conversational interface can still depend on existing search indexes and retrieval infrastructure.
Does winning the Android default make an assistant fully competitive with Gemini?
No. The default is only the front door; meaningful competition also depends on app access, device capabilities, transferable context, and live-information sources across the rest of the stack.