Google studied 2.1M Nano Banana prompts, memes were the smallest share

Google's research team read 2.1 million Gemini Nano Banana prompts and grouped them into five jobs. Memes and humour are the smallest category in the real data. Marketing had the longest, most detailed prompts of the five. The paper is on arXiv and is dated 5 November 2026.

How Google ran the study

The method is called Urania. Google built it to look at prompts under differential privacy. The researchers never read a single prompt or image. The authors sort 2.1 million conversations into roughly 800 clusters. They then group clusters into themes. Each theme is compared across 21 features. Those features include editing style, composition, and how a user talks to the model between turns. The paper is by Will Tucker-Ray and ten co-authors at Google.

The five jobs people use Nano Banana for

The authors name five categories in the data. Social and viral sharing is the first. Brand and marketing collateral is the second. Product visualisations and e-commerce is the third. Educational and technical information media is the fourth. Architectural and spatial designs is the fifth.

Memes and humour are the smallest category, by the paper's own words. That is the shape most press coverage of AI image models uses.

Marketing uses the longest prompts

The brand and marketing category has the highest share of long, detailed text requests, by the paper's own measure. The authors say users in this group are mostly small businesses and individuals. Established companies are less of the group. The social and viral category has people in the frame nearly 80% of the time. That is the highest share of the five.

The paper gives no dollar figure for any of this. It does not say which prompts produced the best images. It describes what people ask for, not whether the model does it well.

Where this is useful, and where it is not

A marketer who uses a general image model as a quick meme maker is working against the common real use. The common use is a long, specific prompt for a brand or product scene. A small team or one person writes it. Google's own data says the heavy users are small businesses writing detailed prose. That is a shift from the viral-demo framing most product launches use.

The paper is Google reporting on Google's own tool, so the picture is partial by design. It does not compare Nano Banana against another image model. It does not say how often the output is good enough to use. A marketer still has to run the prompt and judge the result.

Source

Tucker-Ray et al., arXiv:2611.08749. Secondary coverage by Roger Montti, Search Engine Journal.