How AI Overviews choose which Melbourne businesses to show

Someone in Brunswick types “best physio for lower back pain near me” and gets a paragraph of answer before they get a single blue link. Three or four sites are credited beside it. Those sites get the visit. Everyone else on page one gets scrolled past. The question worth asking is not whether this is fair — it is how those three or four get chosen.

AI Overviews don’t crawl the web live — they re-rank what’s already indexed

An AI Overview is not a separate robot going out to read your website when the question is asked. Google runs its own search behind the scenes, pulls a shortlist of pages it already has indexed, and writes a summary from what those pages say. The citations are drawn from that shortlist.

That has one blunt consequence. If your page is not indexed, or ranks nowhere for the phrasing behind the question, it was never in the running. No amount of AI-specific tinkering fixes a page Google has not stored. Before anything else, check that the page is indexed, that it is not blocked in robots.txt, and that it does not carry a stray noindex tag — we still find those on Melbourne sites months after a rebuild.

So the base layer is ordinary technical AI SEO in Melbourne work: crawlable, indexed, reasonably fast, reasonably authoritative. That gets you into the pool. It does not get you cited.

What gets cited is the passage, not the page

The model isn’t weighing your homepage against a competitor’s homepage. It is looking for the specific stretch of text that answers the specific question, and it prefers text where the answer is stated plainly and close to the surface.

This is why long, atmospheric service pages lose to modest little pages that say the thing. A page headed “Do I need a referral to see a physio in Victoria?” that answers in the first two sentences will beat a beautifully written 2,000-word clinic overview that mentions referrals somewhere in paragraph nine.

Practical version of this:

• Use the question as the heading, in the words a customer would use, not the words your industry uses
• Answer it in the first 40 to 60 words underneath, before you qualify or elaborate
• Keep one question to one heading — merged sections are harder to lift cleanly
• Put the concrete details in text, not in an image: prices, hours, suburbs served, turnaround times

Detail is what tips a citation. A page that says “we service the inner north — Brunswick, Coburg, Preston, Northcote — same week for urgent jobs” gives the model something to quote. “We proudly serve all of Melbourne” gives it nothing.

Agreement across sources is the tiebreaker

When a model has to decide which of six plausible sources to trust, it leans on whether the claim is corroborated elsewhere. Your site saying you are a Melbourne bookkeeper who works with hospitality clients is one data point. Your Google Business Profile, your industry directory listing and a local news mention all saying the same thing is a pattern.

You do not control most of those other sources, which is the uncomfortable part. What you control is whether they agree with each other.

Corroboration is also why inconsistency is expensive. A business name that appears three different ways, an old suite number on one directory, a phone number that changed two years ago — each mismatch weakens the pattern. Getting name, address and phone identical everywhere is unglamorous work with a measurable payoff.

Reviews feed the same mechanism. When an AI answer describes a business as “well reviewed for punctuality”, it is usually paraphrasing review text, not the star rating. Which means what your customers actually write matters more than the number above it.

Structure and freshness: the two cheap wins most sites skip

Two things sit under everything above, and both are within reach of a small business in an afternoon.

Schema markup does not force a citation, but it removes ambiguity about what your page is. LocalBusiness markup with your real address and opening hours, FAQPage markup on your question blocks, Service markup on service pages — these tell a parser what it would otherwise have to infer. On a small business site this is an afternoon of work.

Freshness matters more here than in classic search. Models are cautious about quoting a page with a 2022 date on advice that may have changed, particularly for anything involving rates, regulations or pricing. A page you revisit twice a year, with the date updated because you actually updated something, holds its position better than one left to age.

The same logic runs through answer engine optimisation generally, and it is most of what we do when a client asks about ChatGPT optimisation — the platforms differ, the preference for clear, specific, corroborated, current text does not.

What this looks like as a plan

For a Melbourne business starting from an ordinary site, the order that works:

1. Confirm indexing and remove anything blocking it. Nothing else matters until this is clean.
2. List the 15 to 20 questions customers actually ask you on the phone. Not keywords — questions.
3. Give each one a heading and a direct answer, grouped onto sensible pages rather than scattered across thin new ones.
4. Add LocalBusiness and FAQPage schema. Fix name, address and phone everywhere they appear online.
5. Ask for reviews that describe the specific thing you did, not just a rating.
6. Re-check in eight to twelve weeks. AI answers move slowly and unevenly; a fortnight tells you nothing.

If you sell to both English and Chinese-speaking customers in Melbourne, run the question list twice, in both languages. Translated pages rarely match how people actually search, which is a large part of why bilingual web design is worth doing properly rather than running the English site through a plugin.

Frequently asked questions