How AI Decides Which Local Businesses to Recommend in Australia (2026 Study)

More and more Australians have stopped scrolling Google and simply ask an AI: “find me a good plumber near me.” ChatGPT, Google Gemini, Perplexity and Claude now answer that kind of question millions of times a day. For a local business that raises one nobody has solid Australian data on: when an AI recommends a business, how does it decide, and is it recommending you?

So we ran the numbers. Across June 2026 we put 84 real “find me a [trade]” questions to all four engines, covering 12 trades and every Australian capital plus Geelong, Ballarat and Newcastle, then logged exactly which businesses and which sources each engine cited. Around 330 answers in total. This is first-party Australian data, not a US study repackaged for a local audience.

The most useful thing we found was also the most unsettling, so we will start there.

If you have ever typed a “best [trade] near me” question into ChatGPT and wondered how it picked, this is the answer, measured rather than guessed. Below we cover how we ran the study, why the answers are so unstable, why each engine cites completely different sources, the role your own site plays, how cost questions flip the rules, why reviews decide who gets listed, and what I would actually do about all of it.

How we ran the study

This is first-party Australian data, which matters because most of what gets published about AI search is American. We ran roughly 330 successful AI answers in June 2026. The question set was 84 real “find me a [trade]” style questions: 12 trades (plumber, electrician, house cleaner, gardener, painter, roofer, landscaper, carpenter, air-con, pest control, removalist, locksmith), across every Australian capital plus Geelong, Ballarat and Newcastle, in three intents: recommendation, cost, and find.

We put each question to four engines: ChatGPT with web search, Perplexity, Gemini, and Claude. Then we did something most studies skip. We took eight of the questions and re-ran each one three times on every engine, 96 extra calls, just to see whether the answer held still. It did not. That is where the headline finding came from.

Why does AI give different answers to the same question?

Because AI recommendations are volatile, and how volatile depends entirely on which engine you ask. When we re-ran identical local-business questions three times, ChatGPT kept the same recommended businesses only 8% of the time. Claude kept them 89% of the time. The other two sat in between.

There is a pattern underneath the noise. Volatility is worst in dense markets. Ask for a Melbourne plumber or a Sydney electrician and each run pulled a different directory page, so the businesses churned. Ask for a Ballarat locksmith or a Geelong removalist and the answers held steady, because there are fewer sources to choose from and the engine keeps landing on the same handful.

So what does this mean for you? In a small town, getting cited once can carry you. In a capital city, it will not. You need broad, repeated presence across the sources an engine draws from, because no single mention survives the churn. One listing on one directory is a coin that might not come up heads next time.

Do the AI engines all use the same sources?

No. There is no shared “AI directory” that feeds all four engines. Each one leans on its own set of trusted sources, and the overlap is smaller than you would expect. Win a spot on ChatGPT’s preferred sources and you have done nothing for Gemini.

Here is the source signature we saw for each engine across the 12 trades.

Look at how little they share. Perplexity will quote a Facebook community group thread. Claude almost never touches forums and prefers the old-guard directories. ChatGPT lives and dies by aggregator sites like starworks.com.au. If you only ever optimised for one engine, you would be invisible on the other three. We have seen this play out with our own clients: a business strong in one engine often does not appear at all in the next one over.

Does my own website still matter for AI search?

Yes, more than anything else. Your own website is the most-cited source in every engine for recommendation queries. When an AI recommends a specific business by name, the link it reaches for first is that business’s own domain. The directories fill in around it.

This is the part people get backwards. They assume AI search is about gaming directories, so they pour effort into listings and ignore their own site. The data says the opposite. The directories matter, but they are the supporting cast. Your website is the lead, and if it is thin, slow, or vague about what you do and where you do it, you are handing the engines a weak primary source and asking directories to carry you. They will not carry you reliably, because, as we just saw, directory citations churn.

This is also why our SEO work for tradies still starts with the website, not the listings. The listings are real, but they sit on top of a foundation, and the foundation is the site.

Why do “how much does it cost” questions behave differently?

Because intent changes the entire source mix. Recommendation questions cite businesses and directories. Cost questions cite government regulators and independent cost-guide blogs, and they almost never cite a business’s own site.

When we asked “how much does an electrician cost in Melbourne,” the sources that came back were the Victorian Building Authority, Energy Safe Victoria, state .gov.au pages, and cost-guide sites like whatsthedamage.com.au and homeupkeep.com.au. Business services pages barely featured. The engine treats price as a question of fact and goes to what it reads as neutral authorities, not to someone trying to sell the job.

This is the content-marketing opening, and most local businesses are walking straight past it. You cannot win the cost answer with a services page that says “competitive rates, free quotes.” You win it by publishing genuinely useful, specific cost content: what the job actually costs, what drives the price up or down, what the ranges look like in your city. Answer the cost question better than the cost-guide blogs do, and you can become the source the engine quotes. Ignore it, and you have ceded every “how much” query before it starts.

Do reviews affect whether AI recommends me?

Yes. Reviews are the gate. Across the study, engines did not just recommend businesses, they explained the recommendation by quoting the review count and star rating, often word for word. A business would get listed with a line like “rated 4.9 stars across 200-plus reviews,” and that number was doing the justifying.

This makes reviews different from a vanity metric. The star rating and the volume are not just social proof for the human reading the answer. They are the evidence the engine cites to defend its own pick. A business with 12 reviews and one with 300 are not in the same conversation, and the gap shows up directly in whether you get named. If you are weighing whether a platform like Hipages is worth it, the review profile it builds is a real part of the answer, because that is the signal the engines lean on.

Review velocity matters too, not just the total. A steady stream of recent reviews keeps your rating fresh and your count climbing, which is exactly what an engine wants to see when it decides who to put forward this week.

What should an Australian local business actually do about this?

Four things, in priority order, based on everything above.

• Anchor on your own site. It is the most-cited source in every engine. Make it specific about your services, your service area, and your pricing approach. This is the foundation everything else sits on.
• Spread across each engine’s source layer. Because the four engines trust different sources, you need presence on the aggregators and listicles ChatGPT reads, the UGC and comparison sites Perplexity reads, the review platforms Gemini reads, and the established directories Claude reads. No single listing covers all four.
• Build review velocity. Reviews are the gate and the count gets quoted. Get a steady, recent flow of reviews on the platforms your engines actually cite, not just one.
• Publish answer-first cost content. Cost questions go to regulators and cost-guide blogs, not services pages. Write the genuinely useful cost guide for your trade and city and you can take a query type your competitors have abandoned.

None of this is a one-time fix, because the recommendations themselves are not stable. In a dense city especially, AI visibility is something you hold through breadth and repetition, not something you switch on once.

How solid is this data, honestly?

Directional, and I want to be straight about the limits. The per-cell sample is small: one query per trade, per city, per intent, so treat the source signatures as patterns rather than exact tallies. Gemini’s coverage was partial because the engine got rate-limited for part of the run, so its findings come from the queries that completed. We re-run the whole study quarterly, which is how we catch our own mistakes.

One of those corrections is worth flagging. An earlier, smaller version of this study (23 queries) found ProductReview was never cited. At this larger scale that did not hold. ProductReview does appear, mainly in Perplexity and Gemini. The bigger sample corrected the smaller one, which is exactly why we keep re-running it rather than trusting a single snapshot.

Want to know which AI engines already mention your business and which sources they trust? Get your AI visibility audit or talk to us about getting found in AI search.

Download the full study (PDF)