Most talk about AI image tools is about viral trends: action figures, Ghibli portraits, '80s yearbook photos. A new Google paper asks what people actually do with them. Twelve Google researchers analyzed 2.1 million Gemini conversations involving Nano Banana, with more than 5.2 million prompts, sampled worldwide over 90 days in the first quarter of 2026 (arXiv:2611.08749).
People are curious about it. In the past seven days, "google ai image generation study" was a breakout rising query next to "ai image" in US Google searches (Google Trends), and Search Engine Journal and Digital Information World both covered it this week.
This post sums up what the study found and turns its most useful findings into prompt habits, with three prompts we ran on Nano Banana so you can see the results.
How the study worked
The researchers used Urania, a Google framework for analyzing chatbot conversations under differential privacy. In their words, the analysis ran "without ever looking at a single image or individual prompt" (paper). Conversations were grouped into about 800 topic clusters across five use categories, and each was rated on 21 features such as lighting requests, precise positioning, exclusions and text instructions.
Two caveats matter. The analysis sees text only, not uploaded or generated images. And the authors describe the results as measures of intent, "directional" rather than precise.
What people make with Nano Banana
The five categories, with the number of topic clusters in each (paper):
| Category | Clusters |
|---|---|
| Social & Viral Sharing | 375 |
| Brand & Marketing Collateral | 191 |
| Product Visualization & E-commerce | 135 |
| Architectural & Spatial Design | 66 |
| Educational & Technical Information | 64 |
Cluster counts show how varied each category is, not how many conversations it holds, so don't read the table as market share. Still, more than half of the clusters are about work: marketing, products, rooms and diagrams.
Five findings worth knowing
- People are everywhere. In social and viral use, "the highest prevalence (at nearly 80%) is the inclusion of people." Even in educational prompts, including people was the most common feature, which the authors didn't expect.
- Memes are rare. "Perhaps surprisingly, meme style or format and absurdity / humor are the least prevalent," the paper says. Meme-format requests topped out at 0.7% in any country (Pakistan and the US).
- Positioning, lighting and exclusions dominate work prompts. Precise positioning was among the top features in every category. Lighting led in architecture, marketing and e-commerce. Negative constraints ("no people", "no text") were prominent in architecture, marketing, education and e-commerce.
- Marketing wants exact text. Brand & Marketing had the most requests for exact wording and for "full specification" of text, meaning both the words and the typography. Many of the brands were fictional or little-known: small businesses, side hustles and ventures that don't exist yet. Educational prompts included text about a third of the time. Most architectural, product and social prompts didn't mention text at all.
- People start over more than they tweak. In social prompts, "full reprompts are about twice as prevalent as simple iterations" like "try again".
Country differences made headlines too. Clothing edits peaked in Bangladesh at 35.2%, prompts with people peaked in India at 65.5%, and the US ranked 10th for precise positioning at 28.1% (Digital Information World).
Prompt 1: a room, built like an architect would ask
Architecture prompts in the study led with precise positioning, lighting and multi-subject scenes, and used few people. So this prompt describes the room shell first, then places every object, then sets the light, then lists what to leave out.
A warm Japandi-style living room, photographed as a real interior photo from the doorway at eye level, wide 16:9 frame. Architecture: white plaster walls, light oak floorboards, one large floor-to-ceiling window on the left wall, a low radiator under the window. Place a low cream linen three-seat sofa against the back wall, centered. Put a round pale oak coffee table directly in front of the sofa, a single tall olive tree in a clay pot in the right corner, and a wool rug in oatmeal tones under the coffee table. One paper pendant lamp hangs above the coffee table. Lighting: soft late-afternoon sunlight coming through the left window, casting long warm shadows across the floor, with the pendant lamp switched off. Exclude: no TV, no wall art, no clutter, no people, no pets, no text, no logos, no plants other than the olive tree.

Nano Banana put every object where we asked and kept the room empty of extras. Words like "centered", "directly in front of" and "in the right corner" do more work than style adjectives. If you're redesigning a real room, upload a photo of it and keep the same structure: what stays, where each new piece goes, the light, and the exclusions.
Prompt 2: a product shot with hands, not a face
The study found that product visualization prompts most often include people, usually a product on a model or held in hands, followed by positioning, lighting and negative constraints. This prompt uses all four for a fictional ceramics shop.
E-commerce lifestyle product photo for a small handmade ceramics shop. Two hands with short natural nails hold a matte sage-green stoneware mug from below, the mug centered in the frame with its handle pointing to the right. Steam rises from black coffee inside. The person wears an oatmeal knit sweater; only the hands and sleeves are visible, the face is not in frame. Background: a softly blurred kitchen counter with morning window light from the left, creating a gentle highlight along the mug's glaze and a soft shadow on the right. Shallow depth of field, sharp focus on the mug's speckled glaze texture. Exclude: no logos or text on the mug, no other products, no rings or jewelry, no visible face, no watermark.

Hands add scale and warmth without the problems a face brings. Naming the handle direction and the light side keeps a set of shots consistent, which matters when they sit side by side on a product page. For your own product, upload a clean photo of it as a reference and keep the rest of the prompt.
Prompt 3: marketing text, fully specified
Brand & Marketing was the category with the most "full specification" text requests. Here's what that looks like in practice: every line of copy in quotes, with its font style, color and position.
A printed promotional poster for a fictional neighborhood bakery, vertical layout on cream paper. Top third: the bakery name exactly "HALF MOON BAKERY" in a bold, rounded vintage serif, deep navy, centered. Below it, a crescent moon line icon in mustard yellow. Middle: a flat illustration of a golden croissant, a cinnamon bun and a loaf of sourdough on a navy plate, in a simple mid-century illustration style with mustard and terracotta accents. Below the illustration, a headline exactly "Grand Opening" in large mustard script lettering. Under that, one line exactly "Saturday, December 5 · Free coffee with every pastry" in a clean navy sans-serif. At the bottom, small text exactly "12 Maple Street". Generous margins, balanced layout, all text perfectly spelled, no other text, no real brand logos.

Every line of text came out spelled correctly on the first try. Writing "exactly" before each quoted line and ending with "no other text" stops the model from adding filler words. For seasonal versions of this layout, see the Black Friday flyer generator.
When to start over and when to edit
The finding that people reprompt about twice as often as they iterate is a reminder that the two moves do different jobs:
- Start over when the composition is wrong: wrong room layout, wrong product angle, text in the wrong place. Fix the prompt and run it again rather than stacking corrections.
- Edit when the composition is right and one detail is off: a misspelled word, a color, one object. Name the change and what must stay the same.
The three prompts above follow the same order: what it is, where things go, the light, the exact text, and what to leave out. It works on any model. Try them in the EasyImagen generator with Nano Banana, or switch to GPT Image 2 or Seedream 5 Pro to compare. For more ready-made layouts, browse the tools pages.