Google Analyzes 2.1 Million Nano Banana Conversations

By Nathan Brooks

Google analyzed 2.1 million Gemini Nano Banana conversations and found five broad use groups, with commercial work, educational illustrations and technical instructions extending well beyond viral image making. The study did not inspect uploaded or generated images; it classified text prompts and conversational patterns instead.

Google described the method as a privacy-preserving descriptive analysis intended to show how people use image generation without examining sensitive images of themselves and their families. Its approach uses Urania, an existing framework Google already applies to insights from LLM chat interactions.

Researchers said outside voluntary self-reporting, evidence about why people use LLM image generation remained limited. The paper says: “With this growth in usage and emergent use cases, we sought to provide some descriptive analysis of Nano Banana use, to help understand the ways in which users might find tools like Nano Banana helpful for social connections, educational materials, scientific advancement, commerce, and more.”

Five use groups

Google grouped the conversations into five categories:

• Social & Viral
• Brand & Marketing
• Product Visualization & E-Commerce
• Architectural & Spatial Design
• Educational & Technical Information

Across those groups, researchers compared artistic styles, image editing, image composition, interactions and sessions, and multi-select features.

• Image editing: Requests to remove objects or people, change backgrounds, add elements, resize or restore images. This covers alteration, cleanup and changes to generated images or uploaded references.
• Artistic style: Requests for media such as watercolor, pencil sketch, photography or oil painting; moods such as minimalist, film noir or vintage; and conceptual mashups combining conflicting styles or moods.
• Absurdity or Humor: Nonsense, satire, surreal comedy and humorous exaggeration.
• Imitate Art, Person, or Character: Emulation of a specific artist, public figure or fictional character. The classification also covered specific stylistic and cultural influences.
• Composition: Structural arrangement, including subject placement, perspective, lighting, exclusions such as “what not to do,” people and the number of people.
• Interactions and sessions: The conversation as a whole, particularly how users iterate on and edit generated images.
• Multi-select features: Additional dimensions covering text and typography, transformation type, work context and content-creation stage.

Prompt completeness fell into four labels: Full Specification for precise style and wording; Precise Content Only for text without style; Stylistic/Generic Only for a general vibe without specifics; and N/A when the prompt mentioned neither text nor fonts.

What the prompts reveal

Google found use beyond trivial viral social media, including commercial activity, educational illustrations, technical instructions and product visualization. The researchers said some tasks that once required professional experience are now within reach of more people.

The most prevalent feature in social-and-viral conversations was inclusion of people, at nearly 80%. Composition and artistic-style requests were also common. Users asked the model to start over about twice as often as they requested minor changes.

The researchers said: “Perhaps surprisingly, meme style or format and absurdity / humor are the least prevalent.” That places humorous, meme-oriented prompting behind more conventional commercial and compositional uses.

Brand-and-marketing activity often involved original, fictional or unrecognized brands rather than established companies, suggesting that some users may be developing branding for new or speculative businesses.

Marketing conversations emphasized precise composition: element placement, people, lighting and instructions about what should not appear. Marketing also had the highest prevalence of requests specifying exact text and detailed typography among the five categories.

That pattern suggests marketers were specifying copy and design, rather than simply requesting attractive pictures. The Brand & Marketing category had the highest prevalence of conversations involving small businesses, individuals or side hustles, which were more prevalent than big brands.

Marketing work went beyond first drafts. Users refined logos, changed text and colors, produced promotional materials, and developed consistent visual identities for online engagement and traditional promotional uses.

What does this change for working photographers?

The practical change is that generative prompting is already operating inside production-shaped commercial work, not just novelty content. For photographers and studios, the data supports testing it against repeatable client tasks—promotional variations, logo revisions, instructional visuals and product scenes—while pricing the professional judgment needed for exact copy, controlled composition and consistent branding.

The study measures prompt patterns, not delivered-image quality, client revenue or labor savings. It therefore cannot justify a gear purchase, staffing cut or revenue forecast on its own; it is a workflow signal about where clients may demand more tightly specified iteration.

The researchers acknowledged that it remains unclear whether AI is enabling entirely new activities or simply making existing ones less expensive. They expect additional uses to emerge as users keep experimenting, making generative prompting a workflow to track rather than a settled replacement for commissioned image work.