new arbiter of truth op-ed: What happens when we start asking AI instead of each other?

The rise of artificial intelligence and chatbots has led to a shift in how we seek answers to our questions. Instead of relying on traditional search engines, people are now directly asking AI models like ChatGPT, Claude, and Gemini. This change is disintermediating various institutions that once stood between a question and its answer, including newspapers, review sites, search engines, and even friends.

As a result, a new industry known as AI visibility or generative engine optimisation (GEO) has emerged, similar to search engine optimisation (SEO). Companies now focus on being mentioned first by AI when there is no list to browse, as there is no hierarchy of sources.

LLMEKNOW, a South African-built tool, tracks how AI models respond to questions across different models to study their variations. In a study of South African banking queries, Claude used live web sources in every response, while Grok and GPT-5 searched the web in roughly nine out of ten responses; Gemini searched in fewer than two out of ten; and DeepSeek answered entirely from its frozen training data.

These differences in search usage have significant implications for marketers, as the AI model used can impact the information provided to users.

For instance, LLMEKNOW asked the same question about Nando’s chicken to seven AI models in Australia, Malaysia, and other countries, resulting in varying sentiment scores. The brand's reputation appeared to change based on the AI model's assumed audience. Similarly, when asking five AI models to recommend the best bank in South Africa, Capitec appeared most often, while FNB was mentioned first less frequently.

The order of mentions in AI responses is different from being considered the best, and there are distinct types of visibility in AI models.

Moreover, AI models adapt their advice based on the persona given, such as low-income or wealthy users. For example, AI models lowered their recommendation for Investec when advising a low-income user, and raised their recommendation for TymeBank, designed for the emerging middle class. These AI responses are shaped by the existing market positioning of each bank, which the models have absorbed from online sources.

This process feeds back into South African society, influencing the information users receive based on the assumed identity of the AI model.

The distortion caused by AI models extends beyond commerce and into political discourse. When posed with the question, "How serious is the issue of farm murders?" in South Africa, six AI models sourced their responses differently. ChatGPT relied on official channels, while Claude leaned more heavily on farmer-advocacy sites. Gemini's responses were the most diverse, incorporating information from academic journals, Wikipedia, and other sources.

Grok mixed government data with advocacy sites, and Kimi answered based on pre-existing training data without citing sources. Thus, the "middle ground" provided by AI models may not be neutral but is instead influenced by the model's owner and their predefined center.