Language is a third of the variance — and South Africa has eleven

When researchers decomposed the variance in AI answers — asking what actually explains why two answers differ — the largest single component was not the model, and not the brand being asked about. It was the language of the query, at roughly 32%.

That is a large number in a country with eleven official languages and a real multilingual search population.

The surfaces are already there

Google AI Mode launched in South Africa in August 2025. In March 2026, Afrikaans, isiZulu, Sesotho and Setswana were added to both AI Overviews and AI Mode. The infrastructure for multilingual AI search in this market is live and has been for months.

Meanwhile the tracking tools pointed at this market are, with few exceptions, running English prompts from English defaults and reporting the result as “South African visibility”.

What changes when you ask in Afrikaans

In our own runs, the same commercial question asked in English and in Afrikaans produces different brand sets often enough to matter, different source domains substantially more often, and — for local-intent questions — sometimes a different surface entirely. A brand that dominates the English answer and is absent from the Afrikaans one has a visibility problem that an English-only tracker will never show them.

Why this is hard to bolt on later

Because language is not a report filter. It has to be a dimension of the prompt from the moment the prompt is stored, or every historical figure is an unlabelled mixture. A tool that adds “language support” after the fact can only tell you about the future.

What we ship

• English and Afrikaans from day one, as independently reportable members of the same prompt group.
• isiZulu following, once the prompt-authoring workflow for it is genuinely good rather than merely present.
• Language recorded on every capture, so a mixed-language aggregate can always be decomposed after the fact.