Channel sheet · CH-11 · gain 4 min · logged November 12, 2026

Generative CreativeDirect input

Google analyzed 2.1 million Nano Banana image conversations

Google studied 2.1 million Gemini Nano Banana prompts without reading uploaded images, finding five use cases, precise marketing demands and memes among the least common social requests.

By Amara Osei4 min read795 words

Signal notes

  1. Google analyzed 2.1 million Gemini Nano Banana conversations.
  2. Researchers split the usage data into five categories.
  3. People appeared in nearly 80% of Social & Viral conversations.
  4. Start-over requests were about twice as prevalent as minor-change requests.
  5. Brand & Marketing led in exact-text requests and use by small businesses, individuals and side hustles.
Google Discovers Unexpected Patterns In AI Image Generation
Input monitorGoogle Discovers Unexpected Patterns In AI Image Generation — ZT_OSCAR / pixabay

By Amara Osei

Google analyzed 2.1 million Gemini Nano Banana conversations, and the findings look less like a meme factory than a working tool for people, small businesses and marketing teams.

The research answers a practical question for product teams: how do you learn what users ask an image generator without opening the personal images they upload? People send sensitive pictures of themselves and their families to Nano Banana, so inspecting those files would conflict with the study’s privacy goal.

How did Google study 2.1 million conversations?

The researchers used Urania, Google’s existing framework for extracting insight from LLM chat interactions, to analyze patterns in text prompts and conversations. They did not need to examine the uploaded or generated images themselves.

The researchers framed the evidence gap: “Beyond proactive self-reports in traditional media and on social networks, there is limited data on how and why many people are using LLM image generation.”

They said descriptive analysis could show whether the tool helps with social connections, educational materials, scientific advancement, commerce and other uses, including emerging ones.

Google grouped the conversations into five buckets:

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

Across all five, researchers compared artistic styles, image editing, image composition, interaction and session behavior, and multi-select features, which add extra classifications to each conversation.

Takeaway: The method measures behavior through prompts, without reading the images.

What do users ask Nano Banana to change?

Image editing covered removing objects and people, altering backgrounds, adding elements, and resizing or restoring images. The team defined this as explicit requests to modify an already generated image or an uploaded reference.

Artistic-style prompts named media such as watercolors, pencil sketches, photography and oil paintings, or moods such as minimalist, film noir and vintage. Some users paired clashing styles or moods for conceptual mashups.

The taxonomy also covered creative direction and visual aesthetics, specific stylistic or cultural influences, and two named subcategories: absurdity or humor, covering nonsense, satire, surreal comedy and humorous exaggeration; and imitation of an artist, public figure or fictional character.

Image composition controlled the scene’s structural arrangement: subject placement, perspective and lighting. Interaction features examined the whole conversation and how users iterated on and edited generated images.

Multi-select features added text and typography requirements, types of image transformation, work context and content-creation stage. Four text patterns described how much direction users supplied:

  • Full Specification: The user specifies both style and wording.
  • Precise Content Only: The user provides the text but not the style.
  • Stylistic/Generic Only: The prompt gives the intended feel without specifics.
  • N/A: The prompt mentions neither text nor fonts.

Takeaway: Users specify what to change, how it should look, where it belongs and what the copy should say.

What surprised researchers about social use?

Nearly 80% of Social & Viral conversations included people, the category’s most prevalent feature. Requests involving image composition and artistic styles were also common.

The most surprising finding undercut the assumption that image generation is mainly meme fuel. The researchers found: “Perhaps surprisingly, meme style or format and absurdity / humor are the least prevalent.”

Users asked to start over about twice as often as they asked for minor changes.

Users also applied the tool to commercial activity, educational illustrations, technical instructions and product visualizations, rather than treating it only as a novelty.

Takeaway: Social use centered on depicting people and refining images, not jokes.

What is marketing users actually making?

Brand & Marketing conversations had the highest prevalence of small businesses, individuals or side hustles. Much of the activity involved original, fictional or unrecognized brands rather than established companies, suggesting some users may be developing brands for new or speculative businesses.

Marketing prompts emphasized control over composition, including placement of elements and people, lighting, and instructions about what should not appear.

Brand & Marketing also recorded the highest prevalence of requests specifying exact text and detailed typography instructions. That pattern points to promotional work with defined copy and design requirements, rather than a generic request for an attractive image.

Users refined logos, changed text and colors, created promotional materials, and developed consistent visual brand identities for online engagement and traditional promotional uses.

Takeaway: Commercial demand centered on precise, editable brand assets from small operators, not generic campaign images.

What does the paper settle—and leave open?

The team calls its method a privacy-preserving descriptive analysis. It shows that business, marketing and creative tasks that once required professional experience now sit within reach of more people.

Researchers still cannot say whether AI enables entirely new activities or simply makes existing ones less expensive. They expect additional uses to emerge as people keep experimenting.

Takeaway: The paper documents practical adoption, not a proven new market.

via searchenginejournal.com (Source)

Filed under

  • nano-banana
  • gemini
  • ai-image-generation
  • brand-marketing
  • prompt-analysis
Share this article:

More from Amara Osei

Amara Osei

Show full bio

News editor covering money, funding and market shifts around marketing professionals running AI in the stack at Mart Signal.

126 articles

Bus out

‹ Previous articleNext article ›