The Machine Is the Shortlist
Discovery used to be a ranked page. You could be eleventh and fight your way to seventh. It is becoming a spoken sentence — three or four names, stated as fact — and there is no page two of a sentence.
When someone asks ChatGPT "who's the best conveyancing attorney in Cape Town?", they don't get options. They get an answer. Your firm is in it, or your firm does not exist for that buyer. No snippet below the fold, no sponsored slot, no long tail. The sentence ends.
I wanted to know who's in the sentence. So we measured it: 36 real buyer questions per industry, put to ChatGPT, Claude, Gemini and Perplexity — each with live web search on, the way real people use them — three times each, across property, law and accounting. 1,296 grounded answers, measured June and July 2026, with every prompt and the full method published. The result is The SA AI Visibility Index, which will now recur quarterly whether or not anyone reads it, because the interesting part turned out not to be the league tables. It was the shape of them.
Markets Have Shapes
I expected the three industries to look roughly the same: big brands on top, everyone else nowhere. That is not what came back.
Property is a monarchy. Pam Golding is named in 45% of all answers — and when it's named, it's recommended 91% of the time. The second brand is far behind; the twelfth is a rounding error.
Legal is a fractured parliament. The most-named firm in the country, Bowmans, appears in just 15% of answers. Six firms sit within three points of each other. Nobody governs.
Accounting is an upset. BDO — mid-tier, not Big 4 — leads at 28%. Every one of the Big 4 trails it. EY places ninth. Xero, a software product, is named more often than KPMG.
Same engines, same question format, same country. Three completely different competitive structures. Which means "how visible is my industry in AI" is not a real question. The real question is what shape your market's sentence takes — and where you stand in it.
The Machine Doesn't Know Your Reputation
Here is the finding I can't stop thinking about.
The most-named law firm in South Africa — the one the assistants put at the top of the profession — serves a hard 403 error to every AI crawler we tested. GPTBot, ClaudeBot, PerplexityBot: all turned away at the door, twice, reproducibly. PwC, a Big 4 audit firm, does the same. So do RE/MAX and Rawson in property. In most cases it will be a stock firewall rule nobody chose and nobody monitors.
And yet these brands rank. How do you top a list compiled by machines that cannot read a word you publish?
Other people's words. The engines answering legal questions cite attorneys.co.za 155 times, Lawzana 125 times, Best Lawyers 97 times — and the firms' own websites barely at all. The profession's AI story is being written entirely by its directory layer. One firm, Webber Wentzel, has the highest recommendation conviction we measured in any industry — 96% — with a website that serves no readable content to any crawler at all. Its reputation lives completely in other people's testimony.
That is the mechanism, and it's worth stating plainly: the machine doesn't know your reputation. It knows your paper trail. Decades of brand equity, the corner office, the billboard on the M1 — none of it transfers. What transfers is corroboration: consistent, question-shaped, machine-readable accounts of who you are and what you're good at, written by you and confirmed by others. A brand can be famous offline and absent from the sentence. Several are.
Early Still Beats Big
The optimistic finding hides in the noise floor.
Because we open-coded every answer — capturing every brand named, not just the ones we tracked — the challengers surfaced on their own. A suburban consumer firm, Burger Huyser Attorneys, was named 46 times across the legal answers. That is effectively level with the Big Five corporate firms, achieved without a fraction of their standing. In Gqeberha, a local independent estate agency surfaces alongside the national brands. In accounting, the SME end of the table is crawling with small practices the engines already trust.
This is what an unclaimed channel looks like. The engines reward corroborated, question-shaped presence — not size, not ad spend, not tenure. Those things correlate with corroboration, which is why incumbents mostly still lead. But the correlation is loose enough that a firm nobody outside Johannesburg has heard of can match the giants, and loose enough that giants with blocked front doors are coasting on borrowed testimony they don't know they're borrowing.
Winner-take-most dynamics plus a scoreboard nobody is watching. That combination never lasts. It resolves the moment the losers notice — and the arithmetic of a three-name sentence says most brands will be the losers.
What I'd Do About It
If I ran a brand in any of these industries, three moves, in order.
Open the door. Fetch your own site as GPTBot and ClaudeBot and see what comes back. If it's a 403, nothing else you do matters until it isn't. It is usually a one-line fix that nobody senior has ever been asked to approve.
Get corroborated. Find the mediator layer the engines actually cite in your category — the directories, the review sites, the rankers — and exist there, properly. Then publish the question-shaped material the engines want to quote: not brochureware, answers.
Measure all four engines, and date-stamp everything. ChatGPT names the fewest brands, Gemini the most, Perplexity leans on portals; checking one tells you almost nothing about the others. And the models drift — every number in this essay holds for June–July 2026 and will decay from there, which is exactly why the Index re-runs quarterly. Drift is the story now: who gained, who fell, who fixed their front door.
I spent years writing frameworks about acting before the window closes, and mostly not acting. This one I built. The window is the same as it ever is — open exactly until everyone else notices it's there.
— The data, the method, and every prompt: The SA AI Visibility Index, Edition 1. Built with Auto Alpha Advisory.