Google Study Maps National Differences in AI Image Prompting

By Amara Osei

Bangladesh recorded the world's highest prevalence of clothing and apparel modification prompts in AI image tools, at 35.2%, according to a Google research paper submitted to arXiv on November 5, 2026. The study, authored by Will Tucker-Ray, Marika Swanberg and 10 other researchers, analyzed Gemini conversations involving Nano Banana.

The dataset is large but the framing is cautious. The researchers drew on a random worldwide sample of 2.1 million Gemini conversations containing more than 5.2 million individual prompts, collected over 90 days in the first quarter of 2026.

Which countries lead which prompt categories?

India topped prompts involving people, at 65.5%, followed by Iraq at 65.3% and Somalia at 65.0%. Nigeria recorded 63.5%, Indonesia 63.0%, Malaysia 62.6%, Suriname 62.1%, Nepal 61.8%, the Philippines 61.7% and Congo 60.8%.

After Bangladesh's 35.2% on clothing and apparel modification, Morocco followed at 29.3%, Egypt at 28.4%, Sri Lanka at 27.9% and Iraq at 27.7%. India ranked seventh at 26.2%; Saudi Arabia followed at 26.0%.

Somalia led lighting modifications at 28.4%, with Malaysia at 28.3%, India at 28.1%, Nigeria at 27.7% and Sri Lanka at 27.5%.

Meme-style prompts were rare everywhere. Pakistan and the United States each recorded 0.7% prevalence — the highest of any country measured. The Philippines, Thailand and India each recorded 0.6%, Vietnam 0.5%, and Mexico, Indonesia, Egypt and Brazil 0.4% each.

The United States ranked tenth for precise positioning prompts, at 28.1%. It ranked third for prompts imitating art, people or characters, at 4.9%, and fifth for absurdity or humor, at 2.1%.

The takeaway: national patterns are measurable and consistent enough to rank, but the percentages describe intent, not behavior at scale.

What did the researchers actually measure?

The team said it examined "some differences in the geographic distributions of the topics" and "styles of image generation and editing (lighting, meme formats, precise positioning)." It also compared interaction styles, including "full re-prompting or not" — whether users rewrite a prompt from scratch or simply iterate.

In social and viral conversations, the paper reports: "The highest prevalence (at nearly 80%) is the inclusion of people." The researchers also found that "Full reprompts are about twice as prevalent as simple iterations (e.g., 'try again') in social and viral prompts." Meme style and absurdity or humor were the least prevalent features in that category.

In plain terms: users chasing viral content lean on people-centric images and rewrite prompts wholesale rather than tweak them.

How reliable are the country rankings?

The authors set limits on their own findings. Country comparisons included only countries with more than "100 conversations with those features." They stressed that "These are measures of intents, not outputs," and described the results as "directional and to an extent ordinal, but not necessarily precise."

That caveat matters for marketers reading the tables as market research. A 35.2% figure describes the share of prompts containing a feature, not the share of users, images produced, or commercial outcomes.

The paper raises an ethical question it does not fully resolve. The study itself is titled around "a privacy-preserving descriptive analysis," yet altering a person's appearance or identity without consent — a use case adjacent to the apparel-editing category Bangladesh leads — carries reputational and legal risk that no prevalence ranking captures. For brands planning localized generative AI strategies, the numbers sketch demand; the consent gap sketches exposure.