What a tracker can and cannot see about AI Overviews in South Africa
Best-in-class detection is about 68%. That third matters more than the two thirds.
Google's AI surfaces are now the main event — AI Overviews reach billions of users monthly, AI Mode has passed a billion and is available in around 200 countries, and South Africa has had AI Mode since August 2025. Any serious tracker has to cover them. Very few are honest about how well.
The detection problem
Third-party access to these surfaces runs through SERP vendors, and published benchmarks put best-in-class AI Overview detection at roughly 68%. About a third of genuinely-present AI Overviews come back as nothing.
That is survivable. What is not survivable is recording it as *absent*. A trendline built on “we did not detect one, therefore there wasn't one” accumulates false negatives at an unknown rate and cannot be corrected retrospectively, because the information needed to correct it was never stored.
Tri-state presence
So AI Overview presence is recorded as one of three values, not two: present, absent, not_detected. The third is excluded from rates rather than counted against the brand, and the report shows how many observations fell into it. A period where a fifth of Google observations were undetermined is a period where the AI Overview figures are weaker, and you should be told.
Intent changes everything
AI Overviews trigger on roughly 36% of informational queries and around 5–8% of transactional ones. Local-intent queries frequently resolve to a map pack instead. A prompt set weighted toward transactional and local intent will show low AI Overview coverage that reflects *the surface's behaviour*, not the brand's performance — which is why the market config carries intent rules and the report separates them.
And the legal weather
The vendor route got materially safer in July 2026 when Google's DMCA claims against a major SERP API vendor were dismissed. Terms-of-service and blocking risk remain, which is why adapters are isolated behind one interface and the vendor is a swappable decision rather than an architectural one.
There is still no credible published South African AI Overview trigger-rate data. That absence is the reason we are publishing our own benchmark rather than citing someone else's.
See it against your own client list
A working demo runs your prompts, in your market, on live engines — not a sandbox with seeded data. Bring one client brand and three competitors.