No dependable mechanism exists. WhatsApp has no South African business directory, so Meta AI answers local recommendation questions from general web patterns rather than verified listings, and it has invented business phone numbers. Allow Meta-WebIndexer in robots.txt, the crawler Meta documents as helping it cite and link your content. Expect no measurable referral traffic, because Meta AI passes no distinguishing referrer.
South African agencies have started selling Meta AI optimisation. Before you buy any of it, three facts are worth having.
WhatsApp’s business discovery feature, the one that lets a user browse local businesses by category, works in Brazil. WhatsApp’s own help centre says so. It has not reached South Africa.
Meta AI, asked in June 2025 for a UK rail operator’s contact number, produced a number in the correct format that belonged to a private individual. Meta confirmed to reporters that its assistant had invented it.
And Meta publishes the name of the one crawler that governs whether your content can be cited in Meta AI answers. South African guides skip it.
This article covers what Meta documents, what the South African numbers support, and where the evidence runs out.
Why this question matters more in South Africa
WhatsApp reaches between 93.9 and 96 percent of South African internet users, somewhere around 28 to 30 million people, on figures compiled from DataReportal’s Digital 2026 report. Against 51.7 million internet users at 79.6 percent national penetration, that makes WhatsApp the closest thing the country has to a universal channel.
Businesses have followed. Standard Bank’s first informal economy study, built on interviews across Gauteng, the Western Cape, KwaZulu-Natal, Limpopo and the North West between March and May 2025, found 74 percent of informal small businesses using WhatsApp as their primary tool for marketing and winning clients. Instagram reached 47 percent. Websites reached 30 percent.
Read that last pair again. Among the informal businesses Standard Bank surveyed, WhatsApp outranked having a website by more than two to one.
The reason sits in the price of data. No South African network zero-rates WhatsApp in 2026, but each one sells WhatsApp-only bundles at a steep discount to general browsing. Vodacom starts at R7 for 150MB over three days. Telkom starts at R6 for 150MB daily. MTN sells a weekly bundle around R20. Cell C offers messaging over USSD on *109# without a data charge.
For a customer choosing between R6 of WhatsApp and general data that costs several times more per megabyte, WhatsApp wins. That customer reaches your business through a chat window rather than your website, and Meta AI now sits inside that window.
What Meta AI does when a South African asks it to recommend a plumber
Less than the pitch decks suggest.
Meta AI reached WhatsApp in South Africa during April 2025, free, reachable by mentioning it in a chat. It searches the web for general questions. It does not query a verified South African business directory, because no such directory exists on WhatsApp here.
WhatsApp’s help centre is direct about this. The feature that lets users find a business by category or name “is currently only available to users in Brazil”, with expansion promised over time and no date attached.
A question like “who does geyser repairs in Randburg” therefore gets answered from general web patterns and training rather than from listings anybody has verified. Practitioners testing Meta AI against specific URLs report that it struggles to retrieve content from particular pages and often returns links as unavailable.
The Transpennine Express case shows the failure mode. Asked for the operator’s contact number, Meta AI generated a plausible one that turned out to belong to a real private individual in the UK. Meta told reporters the assistant had hallucinated it rather than leaked it from any contact record.
Draw the practical conclusion. Meta AI can name your business, and it can invent your phone number. Neither outcome is something you control through a content strategy.
Meta runs five crawlers, and one of them decides your citations
Meta documents five user-agents. Most South African guides mention two.
| User-agent | What Meta says it does | Honours robots.txt |
|---|---|---|
Meta-WebIndexer | Navigates the web to improve Meta AI search result quality | Yes |
Meta-ExternalAgent | Trains foundation models, or indexes content to improve products | Yes |
Meta-ExternalFetcher | Fetches individual links at a user’s request | May bypass |
Meta-ExternalAds | Crawls for advertising and business products | Yes |
facebookexternalhit | Builds link previews for content shared in Meta apps | May bypass during security checks |
Meta-WebIndexer is the one that matters, and Meta says why in its own documentation: allowing it “helps us cite and link to your content in Meta AI’s responses”.
That sentence is Meta telling site owners which crawler controls citation eligibility. Check your robots.txt for it before you spend money on anything else.
Two crawlers carry documented exceptions. Meta-ExternalFetcher may bypass robots.txt, because a user asked for that specific page. facebookexternalhit may bypass it during security and integrity checks for malicious content. Blocking Meta through robots.txt gives you partial control rather than a wall.
Meta states a preference for robots.txt over what it calls non-standard formats such as NoAI tags. No umbrella token exists here. Google offers Google-Extended as a single switch for its AI use; Meta requires you to name each crawler you want to stop. Changes take up to 24 hours to reach the crawlers.
Meta does not publish introduction dates for these crawlers, so treat any article claiming a launch month for one of them as guessing.
The two opt-outs that get confused
Business owners conflate these, and the confusion produces the wrong action.
Website crawling runs through robots.txt, per crawler, as above. This controls whether Meta reads your pages.
Personal data in AI training runs through Meta’s Privacy Center objection, which rests on GDPR Article 21 and covers what Meta does with account data belonging to users in the European Union and the United Kingdom. Meta began training on that data in May 2025 after Irish regulatory sign-off, on an objection basis rather than consent.
Outside the EU and UK, Meta offers a third-party information contact form without the same guarantee attached. South African businesses have no equivalent statutory objection route for account data. Objections in either region apply to future processing, and Meta does not remove what its models already absorbed.
Filing a Privacy Center objection does nothing about crawling. Editing robots.txt does nothing about account data. Anyone selling you one as the other has misread both.
The one place llms.txt might do something
We told you in our Perplexity article that llms.txt files do nothing, and the evidence there is strong. Meta is the exception worth naming.
The same twelve-week server-log study across 83 sites that recorded PerplexityBot fetching llms.txt zero times found Meta-ExternalAgent fetching it 193 times, against 172 fetches of robots.txt on those same sites. OpenAI’s crawler managed 7. Anthropic’s managed 9.
Hold that finding at arm’s length. The study’s own authors note their panel skews toward small business sites whose owners had gone looking for an AI visibility tool, which is a real selection bias. Two other analyses disagree: a single-site log study traced most llms.txt requests to non-AI bots, and a 137,000-domain analysis put GPTBot as the leading named fetcher rather than Meta. These studies come from vendors rather than peer reviewers.
Our position: adding the file costs you an hour and might register with one crawler out of five. Paying an agency a retainer for it remains indefensible.
You cannot measure any of this
Standard analytics cannot show you Meta AI traffic.
Meta AI often passes no referrer at all, and in-app contexts inside WhatsApp and Instagram are the worst offenders. When a referrer does arrive, it reads as facebook.com, instagram.com or l.facebook.com, which puts it in the same bucket as somebody clicking a normal social post. Meta Business Suite publishes no report isolating discovery driven by AI.
Two partial routes exist. You can tag links you control with UTM parameters and hope Meta AI preserves them, which is unconfirmed. Or you can run prompts yourself and record which businesses Meta AI names, which measures the answer rather than the traffic.
Any provider quoting you a percentage of traffic from Meta AI is inferring it. Ask them which field in which report produced the number.
What to do about it
Open your robots.txt and confirm Meta-WebIndexer can reach the pages describing what you sell. That is the one lever Meta documents.
Set up a WhatsApp Business account with your catalogue, hours and address filled in. Do this because 74 percent of the informal businesses Standard Bank surveyed already reach customers this way, and because a customer who finds you through Meta AI still needs somewhere to arrive. Do not do it expecting Meta AI to read the catalogue.
Publish your phone number and address as text on your own site, in a place a crawler reaches without JavaScript. Meta AI invents contact details. Giving it the correct ones in a readable form is the closest thing to a defence you have.
Then run the test yourself. Ask Meta AI inside WhatsApp ten questions a customer would ask about your category and write down what it says. Check whether it names you, names a competitor, or invents a number. That record beats any dashboard, because no dashboard for this exists.
What we cannot verify, including a figure we have used
One number circulating in South African marketing material puts Meta AI at 49.6 percent against ChatGPT’s 53.2 percent among South African users. We have quoted it. Having gone back to the source, we are correcting ourselves.
That figure comes from Hypertext’s State of Work Survey 2025, which collected 844 responses through a Google Form embedded on its own website between June and July 2025. Respondents self-selected from a technology publication’s readership. The question asked which chatbots they had ever used, multi-select, rather than which they use or prefer.
The number measures lifetime trial among tech-publication readers. It does not show that half of South Africans use Meta AI. Treat any deck presenting it as adoption data with suspicion, ours included.
We also could not verify: Meta AI’s active user share in South Africa from any neutral measurement body, availability in any South African language beyond English, and whether Meta AI functions on Ray-Ban Meta glasses sold in South Africa, where Meta’s official country list and a local retailer contradict each other.
The Competition Commission’s Media and Digital Platforms Market Inquiry published its final report in November 2025 covering Meta and generative AI, and found that 77 percent of South Africans name social media as their main online news source against 1.2 percent going direct to news websites. We have not read its Meta-specific remedies in enough detail to summarise them, and we will not paraphrase a regulator from a table of contents.
Where this leaves you
Meta AI sits inside the channel most South Africans use to reach a business, which makes it worth understanding. It also has no local business directory, no measurable referral path, and a documented habit of inventing contact details, which makes it a poor place to spend a retainer.
Allow Meta-WebIndexer. Publish your real contact details in readable text. Run your own prompt test each quarter. Then put your budget where the evidence is stronger.
If you want that test run across Meta AI and seven other engines with the raw results handed to you, our AI search visibility service states the scope, the price and the things we cannot promise.