AI Strategy 9 min read

Bots Now Outnumber Humans Online: What SA Businesses Must Change

By ·

The biggest reader of your website is no longer a person. In mid-2026, Cloudflare, which carries a large share of the world’s web traffic, reported that bots now make more than half of the web requests on its network: 57.4% automated against 42.6% human (NBC News, 4 June 2026, retrieved 28 September 2026).

Most South African business websites are still built as if every visitor has eyes. That is now the minority case, and it changes what a website is for. I run a small web agency in Johannesburg, and in September 2026 I audited our own site against exactly this shift. Some of what I found was embarrassing. This article explains what has changed, why the usual advice no longer holds, and what to fix first.

Key Takeaways

  • Bots passed human traffic on Cloudflare’s network in the first half of 2026, up from roughly a fifth of traffic before generative AI.
  • AI assistants read enormously more than they send back: in July 2025, Anthropic’s crawlers fetched about 38,000 pages for every visitor they referred to a site (Cloudflare).
  • The AI audience is already here: Google’s AI Overviews have more than 2.5 billion monthly users, and 23.1% of South Africa’s working-age population used an AI product in the first quarter of 2026, ahead of every other African country.
  • A website now has to be readable by software, specific enough to be quoted, and measured by enquiries rather than visits.

What Cloudflare Reported

In early June 2026, Cloudflare’s CEO, Matthew Prince, said that automated traffic had overtaken human traffic on the company’s network, earlier than Cloudflare had forecast. By Prince’s own account, in late 2025 the company expected the crossover at the end of 2027, by March 2026 it had moved the estimate to the first half of 2027, and it happened in the first half of 2026 (The MAD Podcast, 25 June 2026; NBC News, 4 June 2026; Cloudflare, 1 July 2026, both retrieved 28 September 2026).

Three details matter for what follows:

  • Bots are not new. Before generative AI, bots were roughly 20% of traffic, mostly search-engine crawlers such as Googlebot, according to Prince (TechCrunch, 19 March 2026).
  • AI agents drive the growth. Prince’s example: a person comparing products might visit five websites, while an AI agent doing the same job on their behalf can visit thousands (TechCrunch, above).
  • He expects it to accelerate. On The MAD Podcast in June 2026, Prince predicted that within five years bot traffic could outnumber human traffic by a thousand to one. That is a forecast, not a measurement (The MAD Podcast, 25 June 2026).
Share of web requests on Cloudflare’s network, mid-2026A donut chart showing bots at 57.4 percent and humans at 42.6 percent of web requests.Who is requesting web pages? Cloudflare network, mid-202657.4%botsBots and AI agents: 57.4%Humans: 42.6%Before generative AI: about 20% botsSource: Cloudflare data as reported by NBC News, June 2026

Two caveats up front, because the number travels badly. Cloudflare measures requests on its own network, not the whole internet, and classifying a request as bot or human is never perfect. The direction is not in doubt; the decimal places are.

The Conventional View: A Website Is a Brochure for People

The standard advice to a South African business owner goes like this. Your website is where people go to check you out. Make it look professional, put your services and contact details on it, and do some SEO so it ranks on Google. Success is traffic: visits, page views and the click-through rate from search results.

This view has held for nearly thirty years and it was right for most of them. Search engines sent people to your pages, people read them, some phoned. Traffic was a fair proxy for interest because a human sat behind every visit. The whole commercial internet, from advertising to analytics, was built on that assumption.

It is also the view baked into most SA web design packages: a template, a services page, a contact form and a monthly SEO retainer that reports rankings and visits. I have sold versions of it myself. The logic was sound. The assumption underneath it has quietly stopped being true.

Why That View No Longer Holds

The problem is simple: a growing share of the readers you now need to persuade cannot see your design, will not fill in your form and do not click. They read, and then they answer the question themselves.

AI assistants read far more than they send back

Cloudflare tracks how many pages each AI platform’s crawlers fetch for every visitor they refer back to a website. In July 2025 the ratios were roughly 38,000 pages crawled per referral for Anthropic, 1,091 for OpenAI, 195 for Perplexity, 41 for Microsoft and about 5 for Google (Cloudflare, 29 August 2025, retrieved 28 September 2026). These figures move a lot from month to month, but the shape is stable: the more an assistant answers for the reader, the fewer readers it sends you.

Pages crawled per visitor referred, by AI platform, July 2025A horizontal bar chart on a logarithmic scale. Anthropic about 38,000 pages per referral, OpenAI about 1,091, Perplexity about 195, Microsoft about 41, Google about 5.Pages crawled for every visitor referred (July 2025)Anthropic38,066OpenAI1,091Perplexity195Microsoft41Google5.4Logarithmic scale. Source: Cloudflare blog, 29 August 2025Ratios change monthly; Anthropic’s fell from about 287,000 in January 2025

The AI audience is enormous, and it is here

This would not matter if AI answers were a niche. They are not. Google’s AI Overviews had more than 2.5 billion monthly users by May 2026 (CNBC, 19 May 2026), and Google’s AI Mode passed 1 billion monthly users while the Gemini app reached 950 million (Alphabet Q2 2026 earnings, 22 July 2026; both retrieved 28 September 2026). South Africa is not lagging: Microsoft’s Global AI Diffusion report for the first quarter of 2026 found that 23.1% of South Africa’s working-age population used an AI product in the first quarter of 2026, up from 21.1% in the second half of 2025 and ahead of every other African country in the study (ITWeb, 19 May 2026).

Monthly users of Google’s AI products, 2026A vertical bar chart: AI Overviews 2.5 billion plus monthly users, AI Mode 1 billion plus, Gemini app 950 million.Monthly users of Google’s AI products, 20262.5bn+AI OverviewsMay 20261bn+AI ModeJuly 2026950mGemini appJuly 2026Sources: CNBC, 19 May 2026; Alphabet Q2 2026 earnings, 22 July 2026

Bots don’t click, and the business model is following them

Prince’s other point is about money. Advertising paid for the open web because humans saw ads. As he put it on the podcast, “The problem is bots don’t click on ads.” Cloudflare has responded by making AI crawling a choice rather than a default: since 1 July 2025, new domains on its network are asked whether to allow AI crawlers, with blocking as the default, alongside a pay-per-crawl scheme (Cloudflare, 1 July 2025). It has also started serving pages to AI agents as plain Markdown instead of HTML in a lighter form that uses far fewer tokens (Cloudflare changelog, 12 February 2026).

For a publisher who sells page views, blocking makes sense. For a business that sells engineering services or bathroom renovations, it is the opposite: being read by the assistant is the whole point, because the assistant is now doing the shortlisting.

What the Data Actually Shows

Traffic was always a proxy. What you wanted was the enquiry, and the visit was the only way to get one. AI assistants break the link between the two: the reading happens off your site, and only the enquiry, if you earn it, comes back. So the question is no longer how many people visited, but whether your business is the one the assistant names.

That reframing changes what good content is. Prince described the large language models as a map of human knowledge that shows where the holes are, and said what AI companies want is reputable sources filling those holes with specific, new information rather than another version of what everyone already says. His own example was the local newspaper he bought in Park City, which he expects to earn more from AI licensing than from display advertising, because it is the only source that knows which hotel is best or what the snow is like this week (The MAD Podcast, 25 June 2026).

Translate that to Johannesburg. The AI assistant answering “who refurbishes industrial valves in Mpumalanga” is looking for the same thing: a source that actually says which valves, what sizes, which standards, which plants, and where the workshop is. Generic copy about quality and passion is the part of the internet that is already full. The holes are specific, local and first-hand, and a South African engineering firm or contractor is often the only party in the world that can fill them.

Our own site is a fair illustration of the gap. When I audited digibannister.co.za in September 2026 against how AI crawlers read pages, the main heading and introduction on seven key pages were hidden until JavaScript ran, because of a scroll-in animation. A person saw a smooth fade; a crawler that does not run scripts saw an empty hero section. The pages also had no security headers, the sitemap carried no update dates, and blog posts were credited to the company rather than a named author. None of it was visible in the browser, and all of it affected what a machine could read and trust. The fixes took a morning.

The Better Approach: Build for Readers Who Are Software

The alternative is to treat search engines and AI assistants as a first-class audience for the website itself, not as something to bolt on with a monthly retainer. Five principles, in order of how much they change things:

  1. Readable without a browser. Every page delivered as complete HTML, with its text, headings and prices in the first response. Google renders JavaScript; most AI crawlers are not documented as doing so, and the safe assumption is that they read only what arrives in the raw HTML.
  2. Specific enough to quote. Service pages that state what you do, for whom, where, to what standard and at roughly what cost, in self-contained passages an assistant can lift accurately.
  3. Consistent about who you are. The same business name, phone number, addresses and service areas on your site, your Google Business Profile and directories, described in structured data so software does not have to guess.
  4. Deliberate about crawlers. A robots.txt that allows the search crawlers behind AI answers, with a separate decision about crawlers used only for model training. Blocking everything makes you invisible; allowing everything is a choice you should make knowingly.
  5. Measured by enquiries. Visits will fall as answers replace clicks. Track quotes, WhatsApp messages and calls, and check periodically how the assistants answer your buyers’ questions.

This is what we now build as standard. The full technical checklist, from structured data to sitemap dates, is on our SEO and AI search-ready websites page.

Matthew Prince on The MAD Podcast with Matt Turck, 25 June 2026. The episode opens with the bot-traffic discussion.

What to Do This Week

Start with a ten-minute check, because you may already be blocking the readers you want.

  1. Check your robots.txt (10 minutes). Open yourdomain.co.za/robots.txt. If it disallows everything for GPTBot, ClaudeBot, PerplexityBot or OAI-SearchBot, or if a security plugin blocks “bad bots” wholesale, you are invisible to those assistants. Decide deliberately which to allow.
  2. Read your site with scripts off (15 minutes). In your browser’s developer tools, disable JavaScript and reload your key pages. Whatever disappears is what a non-rendering crawler never sees. On WordPress page builders and some React sites, that can be most of the page.
  3. Ask the assistants (20 minutes). Put your buyers’ real questions to ChatGPT, Gemini, Perplexity and Google, and record whether you appear. This is your baseline; repeat it monthly.
  4. Rewrite one service page answer-first (half a day). Take the service that earns you the most and make the page state what it is, for whom, where, to which standard and roughly what it costs. Engineering firms should treat the page as a capability statement; trades should lead with service areas, credentials and real jobs.
  5. Fix the invisible basics (a day, or a developer). Structured data for your business and services, consistent contact details across your site and Google Business Profile, a sitemap with real update dates, and analytics that separate visits from chatgpt.com, perplexity.ai and gemini.google.com.

Expect technical fixes to be picked up within weeks as crawlers revisit. Citations and rankings depend on competition and on how much trustworthy information exists about your business, so judge results over months, not days. If you would rather have someone else run the check, we offer a free AI visibility check that does steps 1 to 3 and sends you a one-page scorecard.

Caveats

The biggest limitation is measurement. Cloudflare sees a large sample, but it is Cloudflare’s sample, and bot-versus-human classification is imperfect. The crawl-to-referral ratios in particular swing wildly from month to month, and 2026 figures circulating on third-party blogs are unconfirmed; that is why this article cites Cloudflare’s own 2025 numbers.

Second, for most South African SMEs today, Google and word of mouth still bring in far more enquiries than any AI assistant. This is about where the next few years are going, not about abandoning what works now. Do not rebuild a site that is bringing in work because of one statistic.

Third, some businesses are right to block. If your revenue is advertising or your content is the product, being summarised for free is a real loss, and Cloudflare’s default block exists for you. The argument here is for businesses that want to be found and hired.

Finally, the assistants themselves change constantly. What survives those changes is the boring foundation: pages that can be read, facts that can be checked, and a business that is consistently described everywhere it appears.

Frequently Asked Questions

Should I block AI crawlers on my website?

Only if you earn money from page views, such as a publisher selling advertising. For a business that wants enquiries, blocking AI crawlers removes you from the answers your buyers now read. Allow the search crawlers that feed AI answers, and decide separately whether to allow the crawlers used only for model training.

Will AI assistants kill my website traffic?

Some of it, yes. AI assistants read pages and answer the question themselves, so they send far fewer visitors per page read than Google does. In our experience the visitors who do arrive tend to be further along and closer to enquiring. Measure enquiries and quotes, not visits.

How do I know whether ChatGPT or Gemini mentions my business?

Ask them. Type the questions your customers ask, such as which contractors in Durban North do bathroom renovations, into the assistants your customers use, and note whether your business appears and what is said about it. Then check your analytics for visits referred from chatgpt.com, perplexity.ai and gemini.google.com.

Does any of this apply to a small trade business?

Yes, and often more directly than to large companies. Local questions such as who fixes geysers in Bryanston are exactly what people ask AI assistants, and the answers draw on your website, your Google Business Profile and your reviews. A one-page site with a phone number gives an assistant nothing to recommend.

Conclusion

Your website’s most frequent reader is now software, and the useful question has changed from “how many people visited” to “is my business the one that gets named”. The businesses that will do well are not the ones with the flashiest sites but the ones that are easy for a machine to read, specific enough to quote, and consistent enough to trust. That is a shift in what we all should be building, not just a new item on the SEO checklist.

If you want to know where you stand, request a free AI visibility check and we will tell you what the assistants currently say about you. If you are pricing a new site or an upgrade, the website cost guide for 2026 sets out what South African businesses actually pay.

References and Retrieval Notes

All sources retrieved 28 September 2026.

Interested in what we’re building?

Get in touch to learn more about our projects or discuss how we can help your business.