ai-visibility · getting-started

How do you get your business recommended by AI?

What actually decides whether ChatGPT, Gemini or Perplexity name your business: how these systems pick sources, the two failures that disqualify you silently, and the seven things to fix in order.

Three things have to be true. An AI assistant has to be able to reach your website. It has to be able to work out what you do and who you are without guessing. And when it looks at your pages, it has to find something specific enough to be worth quoting.

Most businesses fail the first one and have no idea. Not because they blocked anything deliberately, but because a security setting somewhere between their site and the internet turns those systems away before any of the rest matters.

That is the honest shape of this. It is less a marketing problem than a plumbing problem with a marketing surface, and the plumbing is where nearly everyone loses.

What does AI visibility actually mean?

AI visibility is the outcome: being the business an assistant names when someone asks it for a recommendation. SEO, GEO and AEO are the three disciplines that deliver it, and they are not the same job.

  • SEO gets you the click. You appear in the list of links, someone chooses you.
  • AEO gets you the answer. Your page is the source a direct answer is built from.
  • GEO gets you cited in the AI’s answer. The assistant names you inside what it generates.

They stack rather than compete. Answer-engine work sits inside the generative-engine problem, and search fundamentals sit underneath both. You cannot skip to the top: if a business is not findable by conventional means, it is rarely in the pool an assistant draws from in the first place.

That is the practical reason we treat this as one service rather than three. More detail on the delivery side under getting found by AI.

How do AI assistants actually decide who to name?

Four mechanics, and understanding them makes almost every tactic obvious.

They retrieve, they do not remember. Most cited AI answers are grounded: the system runs a search, fetches a handful of pages, and writes an answer from them in real time. This is the single most useful thing to know, because it means the question “how do I get into the model” is the wrong question. You need to be in the search results it is drawing on, right now.

Extraction favours structure. These systems lift the parts of a page that come away cleanly. A self-contained sentence that answers a question is liftable. The same fact buried in the fourth paragraph after three paragraphs of positioning is not. Marketing preamble is actively expensive here.

Trust is weighted by entity, not just by page. The system has to work out who you are. When a name is ambiguous, it resolves to whichever entity carrying that name it is most confident about. If a larger business elsewhere shares your name, you can be describing yourself perfectly and still be overwritten by them.

Different assistants search different libraries. This one surprises people. Copilot leans on Bing. Perplexity runs its own index. Claude has used Brave’s. Google’s assistants use Google. Permitting a crawler is not the same as being in the index it reads from, and being visible in one assistant tells you very little about the others.

Research accepted to SIGIR 2026, the information retrieval conference running in Melbourne this month, puts numbers on how little these surfaces agree. Grossman and colleagues at NJIT compared what Google’s classic results, its AI Overviews and Gemini returned across 11,500 queries. Overlap was low throughout, and the least similar pair of all was AI Overviews against Gemini, despite both being Google products.

A useful way to picture it: the assistant does two jobs in sequence. First it finds candidate sources, then it decides what to quote. Search fundamentals get you into the candidate set. Structure and clarity get you quoted.

Does ranking first mean getting cited?

No, and this is the correction most people have not caught up with.

Ranking and citation are correlated but genuinely different games. Being retrievable gets you into the room. What gets you quoted is whether your answer is extractable and whether the system is confident about who you are. Plenty of cited sources are not the top-ranked result for the query they were cited on.

There is measurement behind this. A Washington University team (Xu, Iqbal and Montgomery) examined 7,583 Google AI Overviews in early 2026 and found that 29.8% of the domains cited did not appear anywhere on the first page of results for that same query. Two caveats worth stating: it is a preprint rather than peer-reviewed work, and it ran on US trending queries, which skew towards news rather than the commercial searches your buyers make. Treat the number as directional. The direction is the useful part.

That is good news if you are small. It means the citation bar is not “outrank a national competitor.” It is “be reachable, be clear, and have the specific answer they need.” Those are achievable in weeks. Outranking an incumbent is not.

It also means you should optimise for two different things at once: rank to enter the pool, structure and entity clarity to get chosen from it.

What are the two failures that disqualify you silently?

Check these before you spend a dollar on content. Both fail without any error message.

Your security layer is turning the crawlers away. A CDN or web application firewall can block AI crawlers with nothing in your robots.txt to indicate it. Your robots file can be perfectly permissive while the edge quietly refuses the request. Verifying this means checking that the actual retrieval bots get a clean response from your live site, not reading your own configuration and assuming.

Your content only exists after JavaScript runs. AI crawlers render JavaScript far less reliably than Google does. If your pricing, your FAQ answers, or your description of what you do are injected client-side, they can be perfectly visible to humans and to Google and completely absent to the model. Content that only appears after scripts execute is, for these purposes, uncitable.

We check both before anything else in an Audit, because no amount of good writing survives either one.

The seven things to fix, in order

Roughly in order of impact against effort. The first four are weeks of structural work with durable payoff. The last three are the ongoing engine.

  1. Let the AI crawlers in, and ship an llms.txt. Explicitly permit the retrieval bots and confirm they actually get through. The llms.txt file is a plain statement of who you are and how to describe you. Worth doing as cheap insurance and entity clarity. Do not let anyone sell it to you as a ranking lever; Google has said it does not use it.
  2. Fix your entity identity. Organization schema, consistent naming, links to your other profiles, a clear one-line description of what you do and where. If anything shares your name, add disambiguating language: your category and your geography, stated plainly.
  3. Make your people machine-readable. Person schema for whoever the expertise belongs to, real bylines on articles, credentials stated rather than implied. Person-led businesses are easier for these systems to be confident about, which is an advantage worth taking.
  4. Structure what you sell. Service schema on each service page, and validate what you already have. Broken or fabricated schema is worse than none.
  5. Add answer-first blocks to the pages that matter. A question as a heading, phrased the way a human would ask it. The first sentence answers it completely on its own. Then the detail. Then something specific: a number, a named outcome, a real example. The Washington University study above found AI Overviews appeared on 64.7% of question-form queries against 9.5% of everything else, so phrasing a heading as a real question is doing more work than it looks like.
  6. Publish the content people actually ask about. Definitions, costs, comparisons, how to choose, what goes wrong. Cover the neighbourhood of a question properly rather than producing a thin page per keyword.
  7. Earn mentions elsewhere. Most of what an assistant cites about you is not on your own site. Directories, industry listings, genuine coverage, places your business is discussed by someone else.

Steps one to four are mostly technical and you can be ahead of nearly everyone in your category by finishing them. Steps five to seven never really finish.

Two cheap wins most businesses skip: Bing Webmaster Tools, because several assistants lean on that index and it is free, and a complete Google Business Profile, which functions as an identity anchor well beyond local search.

What should I be sceptical of?

More than usual, because this field has a measurement problem and some people are exploiting it.

Anyone guaranteeing you will be recommended. Nobody controls what a model says. Google’s own guidance names ranking and performance guarantees as a red flag, and that is a reasonable test to apply to anyone pitching you.

Anything that looks like gaming the system. Content written only for crawlers, fabricated Q&A stuffing, bought mentions at volume. Search platforms have explicitly named manipulation of generative answers as spam, and it carries real penalty risk.

Tools that promise precise measurement. There are no stable search terms here the way there are in search. Most tools model visibility by sampling synthetic prompts. That is genuinely useful directionally, and it is not the deterministic tracking some vendors imply.

Volume for its own sake. Publishing more thin pages is the most common expensive mistake. Structure and specificity beat quantity, and mass-produced pages targeting question variations can trip content-abuse policies.

The honest position is that this is directional work. It compounds and it is worth doing, and it is not paid media where you can forecast a straight line.

Where does Australia actually sit?

Worth being precise, because the two halves of this have very different urgency.

The discovery and citation game is fully live here now. AI Overviews and conversational search are running in Australia. If buyers in your category are asking assistants who to use, that is happening today, and everything above applies immediately.

The agentic side is not here yet. The AI-completes-the-purchase scenario is largely US-first. That is a build-the-cheap-foundation-now play rather than an emergency, and anyone telling you otherwise is manufacturing urgency.

The FAQ covers the questions this usually raises. One local specific worth two minutes: put your legal business name and ABN on your site. It is a concrete signal that you are a real registered entity, it costs nothing, and most Australian businesses leave it off.

We should be straight about the evidence here. Almost all published research on this is US or global. There is very little Australia-specific data, so treat the numbers you see as directional rather than as a description of your market.

If you want to know whether any of this is currently costing you, the first check is a free Snapshot or a discovery call.

FAQ

Make sure its crawlers can reach your site, make your identity unambiguous with structured data and consistent naming, and give it answer-first content specific enough to quote. ChatGPT grounds answers in a search index, so conventional findability still matters.

Is GEO just SEO with a new name?

Mostly it is shared foundations, and the difference is real but narrower than the marketing suggests. The same technical base serves both. What generative optimisation adds is crawler access for AI bots specifically, and entity clarity so the system knows who you are. Doing good SEO gets you most of the way; it does not finish the job.

Can anyone guarantee AI will recommend my business?

No. Nobody controls model output, and a guarantee is a reason for suspicion rather than confidence. What can be improved is whether you are reachable, identifiable, and quotable, which raises the odds substantially without promising an outcome.

How long does it take to see results?

The technical foundations can be fixed in weeks. Entity strength and citations elsewhere build over months. If a first check finds blocked crawlers, that single fix can change things quickly, since you were invisible rather than uncompetitive.

Do I need to rank on page one to be cited?

No. Being retrievable matters, but citation is decided more by whether your answer is extractable and your identity is clear. This is why smaller businesses can be cited alongside much larger ones.

What is llms.txt and do I need it?

A plain-text file stating who you are and how to describe your business. Cheap to add and useful as an entity-clarity artifact. It is not a ranking mechanism, and Google has said it does not use it, so treat anyone selling it as a growth lever with caution.

How do I know if AI can even see my website?

Confirm the AI retrieval crawlers get a clean response from your live site, behind whatever security sits in front of it, and confirm your key content exists in the raw page source rather than only after JavaScript runs. Both failures are silent, and both are common.

About the author

Paul Korber

Founder, Korbai  ·  AI consulting, automation and training for Australian businesses

Paul Korber is the founder of Korbai, an AI consultancy in Sydney working with small and mid-sized Australian businesses. He spent twenty years in commercial technology, including channel sales across Asia Pacific at Microsoft, before starting Korbai to do the part he kept finding missing: getting AI into the work a business already does, rather than running it alongside. He does not build custom models, and he will tell you when AI is the wrong answer to your problem.

All articles

Get started

Want this working in your business?

Book a discovery call and tell us how your business runs. If we can’t see a payoff, we’ll say so on the call.

AI Readiness Score

Before you go, how ready is your business for AI?

Twelve questions, three minutes, scored on the spot. No email needed to see your result.

Score your business