AI Visibility Report: Methodology

The South African AI Visibility Report measures what the leading AI assistants tell consumers when asked real buyer-intent questions. The method is the same across all 19 industries, so results are directly comparable, and the same every quarter, so results can be tracked over time.

What we asked, and how

For each industry we wrote 20 buyer-intent questions - the questions South Africans genuinely ask when choosing a provider - grouped into seven intents:

Ranking ("what is the best X in South Africa?")
Ease ("which is easiest to sign up / claim / switch?")
Comparison ("Brand A vs Brand B")
Segment ("best X for a student / family / business")
Criteria ("cheapest", "best value", "most reliable")
Trust ("most trusted / reputable")
Problem ("I need X, who should I choose?")

Each question was put to three AI assistants - ChatGPT (OpenAI), Claude (Anthropic), and Gemini (Google) - with live web search / grounding switched on, so the answers reflect current, sourced information rather than stale training data. Every question was asked three times per assistant to smooth out run-to-run variation.

That is 20 questions x 3 assistants x 3 runs = 180 answers per industry, and 3,415 answers across the report (a handful of empty responses are excluded). For every answer we captured the full response and every source the assistant cited. The Q3 2026 data was collected during October 2026.

Which models

Each assistant is run on the model a South African gets by default when they open it, re-checked every quarter. People do not stay on old models: when an assistant updates, its users move with it. Measuring the current default means the report reflects what buyers are actually being told, and part of any quarter-on-quarter movement is the assistants themselves changing.

How brands are extracted

A second AI pass reads each answer and extracts every brand named, where it ranks if the answer is an ordered list, and whether it is an active recommendation, a neutral mention, or a caveated one. Brand names and their variants (for example "VW" / "Volkswagen", or "FNB Private Wealth" / "FNB Private") are mapped to a single canonical brand so visibility is never split or missed. A brand named through its parent company or in a combined name (for example "Matrix (by MiX Telematics)") is counted, while a bare group name is not credited to a differently named subsidiary. The extraction model and rules are held constant between quarters, so a change in the numbers reflects a change in the answers, not in how we read them.

The AI Visibility Quality Score

Every brand gets a single score from 0 to 100. It combines how often a brand is mentioned with where it ranks:

In each answer, a brand scores 1 ÷ its rank when it appears in a ranked list (1st = 1.0, 2nd = 0.5, 3rd = 0.33, and so on, capped at 10th)
A brand mentioned in passing, with no ranked position, scores just behind the last ranked item
A brand absent from an answer scores 0
The score is the average across every answer, including the ones where the brand was absent, multiplied by 100

So a brand recommended first in every single answer scores 100. A brand mentioned often but always near the bottom scores low. This is deliberate: it rewards genuine, prominent visibility rather than mere frequency. Recommendation strength (recommended vs neutrally listed vs caveated) is tracked as a separate measure, not folded into the score.

Mentions, rank, recommendations and citations

Alongside the Quality Score, each report shows:

Mention rate - the share of answers that name the brand at all
Average rank - the brand's average position when it appears in a ranked list (1 = top)
Recommendation rate - how often, of all answers, the brand is the active recommendation
Results across the three assistants - because the models often disagree, "AI visibility" is really three different audiences
Sources cited - the publisher domains the assistants drew on, counted by the number of answers that cited each (so a page cited several times in one answer counts once)

How movement is measured

Each industry page compares this quarter with the last, and the Over time page brings every industry together.

Movement is judged on the Quality Score, because it is the net of mentions and rank. Mention rate and average rank are shown alongside it, but on their own they can mislead: a brand can be mentioned more often yet placed lower down, and when the AI's lists get shorter, everyone who makes the list looks higher.
The same 20 questions are compared answer by answer across both quarters. A change only counts as movement when it is at least 2 points and clears normal run-to-run variation at 95% confidence. Anything smaller is shown as steady.
A race status describes each industry: a clear leader scores at least 50 with a lead of 15 points or more; a category whose leader scores below 35 is wide open; everything in between is a competitive field.

Restated Q2 2026 figures

To compare the two quarters on the same basis, the Q2 2026 figures were restated with the improved brand matching described above. No Q2 answers were re-collected. No industry leader or race status changed; a small number of individual brand figures moved, and the restated Q2 edition shows the corrected numbers.

Important notes

These are AI model outputs at a point in time. They will shift as the models and the live web change. The report is a snapshot, re-run every quarter to track movement.
This is a measure of AI visibility, not market share, revenue, or product quality. A brand can be excellent and under-visible, or modest and over-visible.
Where one assistant occasionally returned an unusable response, that answer is excluded; reported figures are built on complete answers.