Schema markup won't get you cited by AI. Here's what does.

When a brand wants to be recommended by AI, the first reflex is almost always the same: add schema markup. More structured data, more JSON-LD, the full technical checklist. It is the move SEO trained everyone to make.

We checked whether it works. Across 600+ grounded AI answers in South African property, banking and crypto, dissected source by source, schema barely moved the needle. The single most-recommended estate agency in the country has zero structured data on its homepage. The brand carrying the most schema ranks mid-table.

AIV Index research — How AI Decides Which South African Businesses to Recommend — 600+ grounded answers across SA property, banking and crypto (Rule 6).

Schema isn't worthless. It is hygiene — fine to have, easy to do, and not the thing standing between you and a recommendation. The reason it disappoints is simple: structured data helps a machine parse a page it has already decided to read. It does nothing to make the assistant trust you, find you, or believe you belong in the category. Those are different problems, and they're the ones that decide the shortlist. Three other things move them.

The three that did move the needle, in order

When we ranked what separated the brands AI recommends from the ones it ignores, the same three signals came back across all three industries, in this order.

1. Corroboration. Are you present on the sources AI already trusts — the directories, comparison sites and guides it reads to build an answer? This is the floor in any crowded category. A brand that is on none of them is, to the assistant, a brand that doesn't exist.
2. Authority. Does your brand have a real organic search footprint? This correlated more strongly than any single number we measured. Genuine search presence is hard to fake and the engines treat it as a proxy for "is this brand real and substantial."
3. Category fit. AI will not recommend you in a category you don't credibly belong to, no matter how big you are. One of South Africa's largest financial brands scores a flat zero for "best business bank" — because it's an insurer, not a bank. All the authority in the world doesn't buy you a recommendation in the wrong category.

AI doesn't know who's respected. It knows who's corroborated.

Schema sits below all three. You can have perfect structured data and still be invisible if you're on none of the trusted sources, have no search footprint, and are asking to be recommended in a category you don't fit.

One nuance on corroboration, because it decides which way you point your effort. In a crowded category — property, banking — the trusted sources are third-party directories and comparison sites, and the job is to get listed and reviewed on them. In a thin one — newer or more specialised markets where no trusted middleman exists yet — the assistant has nothing to lean on, so a well-built brand site can become the source it cites. Same principle, opposite move. Schema doesn't change either answer.

AIV Index research — How AI Decides Which South African Businesses to Recommend (Rules 3, 4 and 6). Corroboration and authority outranked the technical checklist across all three industries; organic search footprint was the strongest single correlate measured.

Be the answer, not the address

When AI does cite a brand's own site, which page it picks gives the game away. The brands that win citations win them for pages that answer the buyer's question — "what commission do agents charge," "sole versus open mandate," "how to choose an agent." The brands that lose are cited, if at all, for thin contact and office pages. An address, not an answer.

The clearest example in the study: one big agency's most-substantial pages were 11,000-word directories of staff names. Enormous word count. Nothing a buyer asked. The assistant walked straight past them.

Word count is not the signal. Answering is. If your site doesn't answer the questions buyers ask, you are handing those citations to whoever does.

AIV Index research — How AI Decides Which South African Businesses to Recommend (Rule 5).

And it's four races, not one

There's a final reason the schema reflex misleads: it treats AI visibility as one number to optimise. It isn't. The four assistants behave like four different search engines. ChatGPT is cautious and names the fewest brands, often answering from memory and pointing at a brand's own site. Claude and Gemini cast wider and lean on third-party sources. Perplexity is the social one, citing far more Facebook, YouTube, TikTok and Reddit than any other engine. A brand can win one and vanish from another. No amount of structured data reconciles that — only knowing, engine by engine, where you actually stand does.

AIV Index research — How AI Decides Which South African Businesses to Recommend (Rule 1).

The playbook

Strip it back and the order of operations is short.

1. Get corroborated. Earn a presence on the sources AI reads in your category. This is the floor.
2. Build genuine authority. A real organic search footprint is the strongest signal there is. There's no shortcut around being a substantial, findable brand.
3. Write answer pages. Pages that answer real buyer questions get cited. Office and contact pages don't.
4. Fix category fit. Make sure you're asking to be recommended in a category you credibly belong to.

Schema is hygiene, not the lever. Do it once, then spend your effort on the four things above. None of this shows up in your Google rankings — it's a separate race, and it's already running.

Where you stand on each of these is specific to your brand, and you can measure it directly. The full method and the evidence behind every claim here is in the research. And for the household-name version of this story — why famous brands go missing from the AI shortlist — read the sibling piece: why South Africa's most famous estate agencies are invisible in AI.

Sources: AIV Index research — How AI Decides Which South African Businesses to Recommend (600+ grounded answers across SA property, banking and crypto, dissected at the source-URL level; read it here). AI answers shift week to week; this is a point-in-time read of patterns we measured, not a guarantee.