How to show up in AI search in South Africa

Zaiq — Thu, 25 Jun 2026 14:00:00 +0000

This is a cross-post. You can read the full version on Medium.

If you run a business in South Africa, here is a test that takes sixty seconds and quietly tells you how much money you are leaving on the table.

Open ChatGPT, Google's AI Overviews, Gemini, or Perplexity. Type the question a real customer asks right before they buy what you sell: "best [your service] in Johannesburg", "who builds [your thing] in South Africa", "affordable [your category] Cape Town". Read the answer.

Is your business in it?

For most South African companies the honest answer is no, and they have no idea, because they are still watching their Google rank while a growing share of buying decisions now starts inside an AI answer that never shows ten blue links at all. The buyer gets one paragraph, with a few businesses named inside it, and acts on it. If you are named, you might get the call. If you are not, the buyer never learns you exist and you never even see the miss in your analytics.

That is the whole problem this guide is about: how to show up in AI search in South Africa, on purpose, without a budget. The practice has a clumsy new name, GEO, Generative Engine Optimization, and it is not a replacement for SEO so much as the next layer on top of it.

Why AI search changed the rules

Classic SEO optimizes for a ranking algorithm that returns a list. You compete for a position. GEO optimizes for a model that returns a synthesis: it reads a pile of sources, decides what is true, and writes a paragraph. You are not competing for a slot, you are competing to be part of what the model believes.

That changes what actually matters:

Entities over keywords. The model needs to understand that your business is a thing, a specific category, in a specific city, at a specific URL, not just a string that appears near some keywords. That understanding comes from structured, consistent records across the web all stating the same facts.

Citations over backlinks. A backlink passes ranking signal. A citation, your name and URL appearing inside a trustworthy source the model has read, is what makes the model willing to name you in an answer. Consistency beats volume, hard.

Being correct beats being loud. Models punish contradiction. If three sources say you are in Cape Town and one says Johannesburg, you become a hazy entity the model hedges around. Clean, consistent facts make you a confident citation.

Why this matters more in South Africa than people think

Two things make this sharper here than in larger markets.

First, local intent. A lot of South African buying questions are explicitly local: "who builds WhatsApp chatbots in South Africa", "POPIA-compliant CRM for a small business", "how much does a website cost in Cape Town". When someone asks an AI a local question, the model leans on whatever public, structured, credible information it can find about local providers. If the web barely mentions you, the model has nothing to pull, and it reaches for a bigger or better-documented name instead.

Second, cost clarity. South African buyers are price-sensitive and tired of agencies that hide their rates. When an AI gives a straight answer in rand, that answer becomes the reference point for the whole conversation. Businesses that publish honest, sourced information get quoted. Businesses that stay vague get left out of the answer entirely.

How to show up in AI search: a practical first move you can do today

You do not need a vendor to start. Spend twenty minutes:

Run the test above across all four big AI engines. Write down, verbatim, what each one says about your category and whether you appear at all.
Note the sources they cite when they answer. Those sources are your real competitors for AI visibility, not necessarily whoever ranks number one on Google.
Check your own entity consistency. Does your business name, one-line description, city, and URL read identically on your website, your Google Business Profile, and every directory you are listed on? Every mismatch is a point of doubt for the model.

That audit alone usually surfaces the gap. The businesses that show up in AI answers are the ones with a clean, consistent, well-linked entity footprint. The ones that do not are usually invisible for a fixable reason, not a mysterious one.

What to do with the result

If the engines already name you correctly, protect it: keep your facts consistent and keep publishing honest answers, because the models re-check.

If they do not name you, or name you wrong, here are the fixes in order of leverage:

Fix your own pages first. State plainly what you do, where, and what it costs. Put real numbers where you can. Highest-return change, costs nothing but honesty.
Make your facts consistent everywhere they already appear, so the model resolves you into one confident entity instead of three fuzzy ones.
Earn a few independent citations over time from credible, relevant pages. Slow, but it is what converts a self-claim into a fact the model will repeat.
Decide what to automate. Some of this is worth handing to a tool or a small automation, and some is not.

Where this is going

The search box is not dying, it is dissolving into everything. Buyers will keep asking machines for recommendations, and the machines will keep getting more confident about which businesses to name. The companies that win the next few years will not necessarily have the biggest ad budget. They will be the ones the models understand well enough to recommend.

That is the work: make your business legible to the systems that now answer your customers' questions. Get the facts clean, get them everywhere that matters, and get cited by sources the models trust.

Zaiq is an AI engineering studio in Johannesburg that builds AI-search visibility systems, automation, and custom software, idea to shipped in a day. If you want to see exactly where your business stands across ChatGPT, Gemini, Claude, and Perplexity, that is the first thing we measure: zaiq.co.za

How we built a gate-accurate Game Boy emulator in Rust (and got dmg-acid2 pixel-perfect)

Zaiq — Sat, 06 Jun 2026 20:18:07 +0000

We are Zaiq, an engineering studio. We built Revenant, a Game Boy and Game Boy Color emulator, from scratch in Rust, compiled to WebAssembly, with a playable arcade that runs in your browser. No emulation libraries: every chip is written by hand.

Play it right now: https://zaiqltd.github.io/revenant/
Code and the full accuracy scorecard: https://github.com/zaiqltd/revenant

Here is what we built, and what "accurate" actually means.

"It runs Pokemon" is the wrong bar

Most emulators clear one bar: a game boots and looks roughly right. That is the easy 90%. The interesting part is the last 10%, the timing edge cases real cartridges depend on, the ones that only show up when you emulate the machine on its own 4.194304 MHz clock, cycle by cycle.

So we measured ourselves against the bar the hardware sets for itself: the community's test ROMs. The headline result: dmg-acid2 renders pixel-perfect, 0 of 23040 pixels different from real hardware, on both the original DMG and the Game Boy Color. We currently pass 130 of 279 of the canonical accuracy gates, and the full scoreboard is in the repo so anyone can check.

The CPU: stepped one T-cycle at a time

The SM83 core is not just a set of opcodes, it is a set of opcodes with precise sub-instruction timing. Reads and writes land on specific cycles within an instruction, and peripherals see those accesses as they happen, not batched at the end.

So Revenant steps the CPU one T-cycle at a time and advances the rest of the machine in lockstep. That is more work than the common "run the whole instruction, then total the cycles" approach, but it is the only way to get the halt bug, interrupt timing, and the PPU interactions right.

The PPU: a real pixel FIFO, not a scanline blit

The shortcut for graphics is to render a whole scanline at once. It is simple, and it is wrong: it cannot reproduce mid-scanline register changes, which real games and every serious test ROM use.

Revenant implements the actual pixel FIFO: background and sprite pixels are fetched and shifted out one at a time, with the fetcher and the LCD controller modelled as the state machines they are on hardware. That is what makes dmg-acid2, the PPU's torture test, come out pixel-perfect.

The rest of the machine

A 4-channel APU (two pulse channels, wave, noise), so the games have real sound.
The cartridge mappers: MBC1, MBC2, MBC3, MBC5, including the MBC3 real-time clock.
Full Game Boy Color support: double-speed mode, HDMA, and the colour palettes.
Battery saves persist in the browser, per cartridge.

Shipping it to the browser

The core is plain Rust compiled to WebAssembly, so the same code that passes the test ROMs in a headless harness also runs the arcade in your browser at full speed. On top we put a live CPU and PPU debugger that ticks alongside the running game, and instruction-level rewind, so you can step the machine backwards when something looks wrong.

The arcade has zero copyrighted content

We did not want a demo that asks you to "go find a ROM." So we wrote a tiny SM83 assembler (an original 8x8 font, APU blip helpers) and hand-assembled 11 original homebrew games: Snake, Breakout, Blocks, Flap, Blaster, Pong, Crosser, Maze, Memory, Dodge, and the first thing the emulator ever drew, a movable smiley. Each has a title screen, a score, a game-over screen, and sound. Or drop in your own .gb / .gbc file.

Why we built it

Not for a client. We built it because the team that can emulate a 1989 console at the cycle level is the team you want on your hardest problem. That is the standard we hold at Zaiq, and we aim it at real business problems, with AI as the edge.

Play Revenant: https://zaiqltd.github.io/revenant/
If you have a hard problem, bring it: https://zaiq.co.za

DEV Community: Zaiq