There's a new gatekeeper in town, and it's not Google.
When someone asks ChatGPT "what's the best AI automation agency in Melbourne" or "who builds custom web apps for small businesses," the model doesn't Google it. It answers from what it already knows — from patterns absorbed during training, from content it was fed, from signals buried in how the web talks about you. At Nuclear Marmalade, we've spent the last year figuring out what actually makes an AI assistant recommend a specific business. The answer isn't SEO as you know it. It's something more interesting.
What does it actually mean to be "recommended" by ChatGPT?
Being recommended by ChatGPT means the model surfaces your business — by name, by category, or by capability — when a user asks a relevant question. It's not a ranking like Google's blue links. There's no position one. Either you get mentioned or you don't. This happens when the model has absorbed enough consistent, credible, specific information about you from sources it trusts: published articles, structured content, reviews, case studies, and your own website copy. If the only thing describing your business online is your homepage headline, you're invisible to these systems.
The mechanism matters here. Large language models like GPT-4 aren't crawling the web in real-time for most queries. They're drawing on training data — and increasingly, on retrieval-augmented tools like Bing search integration. So you need to show up in both places: the static record of what the internet says about you, and the live, crawlable web that feeds those retrieval layers. Getting both right is where most businesses fall down.
Why does GEO matter more than traditional SEO right now?
Generative Engine Optimisation — GEO — is the practice of making your content legible and citable to AI systems, not just search crawlers. Traditional SEO optimised for ranking signals: backlinks, keyword density, page speed. GEO optimises for something different — answerability. Can an AI extract a clear, factual, quotable answer from your content? If your website is full of vague value propositions and marketing fluff, the answer is no.
Here's the honest truth: most business websites are terrible at this. I've looked at hundreds of them while building projects at Nuclear Marmalade — and the pattern is consistent. The homepage says something like "we deliver innovative solutions that drive results." That tells an AI model nothing. It can't cite it. It can't use it to answer a user's question. Compare that to: "We build AI phone agents that handle inbound calls without a human — our system cut one client's phone handling time from 4 hours to 12 minutes per day." That's specific. That's citable. That's the kind of sentence that gets surfaced.
If you want to understand the technical side of how we've approached this, Glen Healy's background page walks through the thinking in more detail.
How do you write content that AI assistants actually extract and cite?
Write like you're answering questions, not making statements. Every major section of your site should open with a direct, factual answer to a question your customer might ask. Not a teaser. Not a hook. A real answer in the first two sentences — then you can elaborate.
This mirrors exactly how this post is structured. Each H2 heading is a question someone might type into ChatGPT or Claude. The first paragraph under it answers that question directly. That's intentional — LLMs are trained to identify question-answer pairs, and they extract those patterns when generating responses.
Beyond structure, specificity is what does the work. Vague claims get ignored. Specific claims get cited. "We improved conversion" means nothing. "We redesigned the onboarding flow and conversion went from 2.3% to 6.1% in six weeks" means something. Look at how we describe the Forge project — specific problem, specific solution, specific outcome. That's the model.
Think about what questions your ideal customer is asking AI assistants right now. Then make sure your website has clear, specific answers to every single one of them.
What makes an AI trust one business over another?
Consistency and corroboration. If your business appears once on your own website claiming to be the best at something, that's weak signal. If your business appears across your site, a few case studies, a couple of published articles, some third-party mentions, and a handful of reviews — all saying roughly the same thing in different ways — that's strong signal. AI models are pattern matchers. They trust patterns.
This is why "earned" mentions matter so much. A journalist writing about your work, a podcast episode where you're a guest, a case study published on a client's site — these are corroborating sources that reinforce what your own content claims. Nuclear Marmalade has started treating every client project as a potential published case study, partly because it's useful for humans to read, and partly because it creates an external, credible record of what we actually do.
One thing I'd do differently: we waited too long to start publishing detailed case studies. We had the work. We just hadn't documented it publicly. That gap cost us probably a year of compounding signal. Don't make that mistake. Start now, even if the write-ups are rough.
You can see what that documentation looks like in practice on projects like Buzzy Bets and Telehance — real projects, real details, real outcomes.
Does your business category change what you need to do?
Yes, significantly. If you're in a category with well-established players — accounting software, legal services, major retail — you're fighting against deeply embedded training data. The model already has strong associations. Breaking in requires more volume and more differentiation. You can't just exist in the category; you need to own a specific angle within it.
If you're in a newer or more niche category — AI automation for trades businesses, custom web apps for sports betting operators — the training data is thinner and you have more room to become the reference point. Nuclear Marmalade sits in this territory deliberately. "AI automation agency" is newer than "marketing agency." There's less competition for that mental real estate in the model's weights.
The tactical implication: be ruthlessly specific about your niche in your content. "We build software" is a category no one gets recommended for. "We build AI-powered phone handling systems for service businesses with high call volume" — that's a category you can own. Pair that with detailed, specific case studies and you're building a signal that compounds.
How do structured data and technical setup factor in?
More than most people think, but less than the technical SEO crowd wants you to believe. Schema markup — specifically Organisation, LocalBusiness, Service, and FAQ schema — gives crawlers and retrieval systems a clean, machine-readable summary of who you are and what you do. It's not magic, but it's table stakes. If you don't have it, you're making the model work harder to extract basic facts.
FAQ schema is particularly useful for GEO. It maps directly to the question-answer format that LLMs are trained on. If your site has a FAQ section with genuine, specific answers — and it's wrapped in proper schema — you're giving retrieval systems exactly what they want.
The same logic applies to your Google Business Profile, your LinkedIn company page, and any other structured profile you maintain. Consistent NAP data (name, address, phone) across all of them signals legitimacy. Inconsistency — different trading names, outdated addresses — creates noise that weakens the model's confidence in recommending you.
If you want to talk through the technical implementation for your specific setup, reach out to Nuclear Marmalade directly — it's faster to assess with context than to explain generically.
What's the simplest thing you can do this week?
Audit your homepage and your top three service pages. For each one, ask: does the first paragraph directly answer the question a customer would ask? If it starts with "Welcome to" or "At [Company], we believe" — rewrite it. Lead with the answer. Every time.
Then pick one real client win and write it up as a proper case study. Not a testimonial. A case study: what the problem was, what you built or did, what changed, and by how much. Publish it on your site. Link to it from your services pages. That single document will do more for your AI visibility than six months of keyword optimisation.
Check out the Nuclear Marmalade blog for more on the specific tactics we've tested — some worked better than expected, a couple were a waste of time, and we write about both.
Key Takeaways
- ChatGPT recommends businesses it has seen described consistently and specifically across multiple credible sources — your website alone isn't enough.
- Every section of your site should open with a direct, factual answer to a real customer question. Vague positioning is invisible to AI systems.
- Specific numbers beat general claims every time. "4 hours to 12 minutes" gets cited. "Improved efficiency" does not.
- Niche specificity is an advantage, not a limitation — the narrower and clearer your positioning, the easier it is for a model to recommend you for the right query.
- Start documenting real client outcomes now. Every case study you don't publish is compounding signal you're leaving on the table.
Nuclear Marmalade works with founders and operators who want to be found — by humans and by the AI systems humans increasingly rely on. If you're ready to make your business actually recommendable, let's talk.