Why Australian businesses are losing Google traffic to ChatGPT

Where did the traffic actually go?

The pattern is consistent enough now to be recognisable. Technical SEO is clean, the primary keywords are still on page one, and organic clicks are drifting down anyway, hardest on the commercial searches that used to produce enquiries. There is no penalty to find and no missing tag to fix, because nothing about the site got worse. What changed is what the buyer did before they arrived.

When someone in Sydney or Melbourne needs a corporate tax accountant, a specialist logistics partner or a commercial fit-out crew, the old sequence was to search a short phrase, open three or four sites and compare them. The new sequence is to ask one full question and read one answer. Something like: which commercial fit-out companies in Sydney specialise in heritage office refurbishments under $500k.

The assistant does not hand back ten links to assess. It names two or three businesses, says why it picked each, and offers to compare them. If you are not in that sentence you have not lost a ranking position, you have lost the shortlist, and the buyer never learns you exist.

What do the numbers show?

Four published figures explain most of the decline, and they are worth reading in order, because each one narrows the problem.

The last of them is why so many businesses were caught out. You can hold position one for the phrase you have optimised for years and still be absent from the answer sitting above it, which is where well over half of the clicks now go.

Half the adult population now uses a chatbot

Pew Research Center found in February 2026 that 49 per cent of US adults use AI chatbots, up from 33 per cent in 2024, and that 42 per cent use them to look for information. There is no equivalent Australian survey, so this is a proxy rather than a local number, but the direction is not in dispute.

AI Overviews now appear on about half of queries

BrightEdge measured AI Overviews on roughly 48 per cent of tracked queries in February 2026, up from about 30 per cent a year earlier. This is the figure that reaches people who have never opened an assistant, because it sits inside an ordinary Google search.

An AI Overview costs the top result most of its clicks

Ahrefs re-ran its click-through study across 300,000 keywords on December 2025 data and found a 58 per cent lower click-through rate for the top-ranking page when an AI Overview is present, up from 34.5 per cent eight months earlier.

Ranking in the top ten barely predicts being cited

Across sixteen months of tracking, BrightEdge found only 16.7 per cent of AI Overview citations came from pages ranking in the organic top ten. Five citations in six were pulled from somewhere else entirely.

What is answer engine optimisation?

Search engine optimisation is the work of convincing a crawler to rank your pages. Answer engine optimisation, also called generative engine optimisation, is the work of convincing a model to recommend your company. Same buyer, different gatekeeper, and the new gatekeeper reads different signals.

Traditional SEO leans on keyword placement, domain authority, backlinks and page volume. All of that still matters, because an assistant runs ordinary searches to gather its sources. What it no longer does is decide the answer.

What decides the answer is closer to a background check on the business.

Entity confidence

How certain the model is that your business exists, trades in Australia, and does the specific thing the question asked about. A brand it cannot resolve is a brand it will not name.

Cross-web agreement

Whether review platforms, directories, trade press and forums describe you the same way your own site does. One source saying it is a claim. Five independent sources saying it is a fact.

Sentiment, and the recency of it

What the visible feedback says about you, and how old it is. A page of reviews from two years ago reads as a business that may no longer be trading.

Extractable answers

Whether your pages state prices, timelines, comparisons and process in plain sentences a model can lift, rather than holding them behind a contact form.

Why does ranking first not guarantee a recommendation?

A model does not read page one of Google and summarise it. It runs several searches from one question, draws on a wider retrieval pool than the visible top ten, and assembles an answer from the sources it trusts most on that particular claim. The BrightEdge citation figure above is the measurement of exactly that gap.

The brands that get left out are usually the ones whose entire visibility lives on their own domain. Years of on-page blogging and bought links produce a site that ranks and an entity nothing independent verifies, because nobody else has said what the site says. Asked to recommend a supplier, a model will not stake an answer on a single self-interested source, so it names a competitor that three other sites vouch for.

This is also the explanation for the reverse case that confuses people, where a small business with a thin site gets named ahead of the market leader. It was not better optimised. It was better cited.

What can you do this week?

None of this needs an engineering budget to begin. Three pieces of work move first, roughly in this order.

Run a prompt audit

Open ChatGPT, Claude and Perplexity and ask the five questions a buyer would actually ask, in full sentences, including the ones that never mention your name. Record whether you appear, who does instead, and which sources each answer cites. The source list is the most useful part of the exercise.

Fix the third-party footprint

Get your name, address and phone number agreeing across Google Business Profile, ProductReview.com.au, G2, the industry directories that cover your trade and any professional body listing you. Then work on recent reviews. This is the cheapest way to raise entity confidence.

Answer the specific questions on your own site

Publish the pages that state what a job costs, how long it takes, how you compare to the alternative and what the local regulatory requirement is. Short paragraphs, headings phrased as questions, tables where a table is honest. Models quote structure.

How do you tell a real decline from noise?

A single check tells you very little. Answers vary between engines, between cities and between days, so one absence is not evidence and one mention is not a win. What you need is the same set of questions asked on the same schedule, kept over weeks, with the cited sources recorded each time.

Two things become visible once you have a series. The first is direction, which is the only reliable signal that the work has landed. The second is which sources actually feed your market, because a platform that keeps appearing in answers about your category is where the next quarter of effort belongs.

Scout LLM runs that series automatically across ChatGPT, Claude, Perplexity, Gemini and Google AI Overviews, city by city, and ranks the sources cited in each market. Doing it by hand works too, as long as it is the same day each week and every answer is kept. The method matters more than the tool. A free check on one domain takes about 40 seconds if you want the starting picture before committing to anything.