Does ChatGPT decide differently for voice or image shoppers?

November 17, 2026
Written by:
Novi

Does the way consumers shop with AI influence which products get recommended?

To explore this, Novi conducted a series of controlled experiments examining how GPT selects, evaluates, and recommends products across different shopping experiences. The research focused on moisturisers for dry facial skin in the UK market, analyzing how product recommendations, evaluation criteria, sources, and search behavior varied between text, voice, and image-based interactions.

‍

How GPT Processes Voice and Images

While consumers can shop through text, voice, or images, GPT processes these interactions differently.

With voice, a separate model (GPT-Live) handles the conversation and passes more complex shopping tasks to the same reasoning model used for text interactions. That model researches products, evaluates options, and generates a full response. GPT-Live then delivers a shorter spoken summary, leaving out some of the supporting details and source links.

With images, the reasoning model evaluates the uploaded photos directly, including visible product information such as packaging, ingredients, and claims. It then conducts text-based web searches and generates recommendations using that visual information as additional context.

‍

What We Found

1. Voice didn't significantly change which products GPT recommended

Across both text and voice interactions, CeraVe Moisturising Cream remained GPT's most frequently recommended product, appearing first in 81% of text evaluations and 76% of voice evaluations. This difference was not statistically significant, suggesting that switching from typing to speaking did not meaningfully change GPT's top product selection under the conditions tested.

We also found no detectable differences in the volume or types of sources cited or the searches GPT performed. The key difference was what actually reached the shopper. In voice interactions, the same underlying reasoning model handled the product research, but the voice model delivered a shorter spoken summary of its findings, leaving out some of the reasoning and all source links. On average, spoken responses communicated 8.7 recommendation themes compared with 11.2 in written responses.

2. Product images influenced how GPT evaluated products

When GPT evaluated product images instead of written product information, we observed shifts in which moisturisers were recommended first. Aveeno led 58% of written-description evaluations compared with 38% of image-based evaluations, while Eucerin increased from 30% to 44% and CeraVe from 12% to 18%.

Although these differences in product selection were not statistically significant, the factors GPT emphasized in its recommendations changed considerably. Pack size and value appeared in 70% of written interactions but only 27% of image-based interactions. Image-based shopping also produced a statistically significant shift toward clinical and evidence-focused searches.

Across both image-based and written product comparisons, GPT was more likely to recommend products presented earlier. In fact, product presentation order had a greater influence on the top recommendation than whether GPT evaluated images or written product information.

3. Product presentation order influenced recommendations

Another finding was the influence of product presentation order. Across both image-based and written product comparisons, GPT was more likely to recommend products that appeared earlier in the information it received, regardless of the format.

This effect remained even after accounting for differences between images and written descriptions, suggesting that how product information is organized and presented can influence AI recommendations alongside the product attributes themselves.

‍

What This Means for Brands

The results highlight an important distinction for brands navigating AI-driven commerce: different shopping experiences may not necessarily change which products GPT recommends, but they can influence how those products are evaluated, which attributes receive attention, and how much information ultimately reaches the shopper.

Understanding these differences is increasingly important as consumers move between typing questions, speaking to AI assistants, and using images to explore products. Brands need to understand not only where their products appear in AI recommendations, but also the information and evidence influencing those decisions.

As AI shopping evolves, getting recommended will require more than visibility. Brands need to understand why AI chooses certain products over others and what they can do to influence those decisions.

‍

About the Research

This research was conducted by Donya Fozoonmayeh, Senior Machine Learning/AI Engineer at Novi, through a series of controlled experiments examining how GPT evaluates and recommends products across text, voice, and image-based shopping experiences. The study focused on the UK facial moisturiser category, analyzing differences in product recommendations, evaluation criteria, citations, and search behavior.

‍

‍

Get the Latest AI Commerce Insights

Thank you!

Your submission has been received!
Oops! Something went wrong while submitting the form.

See Novi in Action

Book a Demo