AI Shopping Agents Are Mainstream Now, and Most South African Business Sites Are Invisible to Them

CI&T's new Retail Tech Report: Agentic Commerce Edition found that 86% of UK and Ireland consumers have used an AI shopping agent or are open to trying one, 68% already have. The rise of AI shopping agents in South Africa is coming whether local retailers are ready or not, and the question that matters isn't whether agents will shop for your customers. It's whether your site is structured well enough for one to understand what you sell.

Two Audiences Read Every Product Page Now

A website used to have one audience: a person scrolling on a phone. Now there's a second: an AI agent parsing the page, comparing it, and recommending it to a human who never visited the site. Most South African business websites are built only for the first audience. Product descriptions sit as loose paragraphs, pricing hides behind "contact us," and category structures exist for a human's eye, not a machine's query. An agent can't ask a follow-up question when it hits ambiguity; it either finds structured, queryable data or moves to a competitor's listing that has it. Comparing brands is the single most common use of an AI shopping agent, at 45% per CI&T, which means a business an agent can't read accurately gets left out of comparisons its own customers are running anyway.

The Fix Isn't More Content, It's Structured Data for AI Agents

Most advice about "AI SEO" says to publish more blog posts and hope an AI mentions you. That misses what's happening. An agent needs data structured and semantically searchable: attributes it can query, categories it can match to intent, pricing and availability it can extract without guessing. This is a database problem wearing a marketing costume, and it's why the businesses showing up in agent-driven comparisons will be the ones whose catalog was built for machine consumption from the start, not retrofitted with a chatbot bolted onto the homepage.

When we scoped the architecture for a South African local-services marketplace concept earlier this year (the idea: a user types "my geyser burst, I need someone today" and the system would match and dispatch a verified provider), the decision that would matter most was the one we'd need to make before writing a single line of matching logic. We planned it around PostgreSQL with the pgvector extension, which stores data as embeddings, numeric representations of meaning rather than keywords, so a plain-language request could match provider listings by intent. PostGIS for proximity, Smile ID for South African ID and CIPC verification so matching would happen against verified providers rather than a scraped directory, and Stitch for escrow, so funds would only release once a job was confirmed done. None of that was cosmetic; it was the answer to the question every SA business now faces: is your data structured for a machine to query, or only formatted for a human to read?

What to Check on Your Own Site

If you run a South African business with a catalog, price list or service menu, ask your developer one question: can our data be queried by attribute (category, price, location, availability) without a human reading a paragraph first? If the answer is no, that's the gap, and it's rarely fixed by a chatbot widget. Retailers with SKUs, price tiers or bookable slots should treat this as an infrastructure priority for 2026. A bespoke legal or consulting practice has more time, since agents are weakest where a human judgment call still matters most.

67% of consumers in CI&T's report already say AI makes them smarter shoppers, meaning the agent is becoming the first filter customers run before they ever reach your site. The businesses that win the next few years of search won't be the ones with the cleverest content. They'll be the ones whose data was built to be read by something other than a person, which is exactly the groundwork we scope into every web app development project before a single interface gets designed.