The First Time a Customer Found Us Through a Chatbot
A customer found us by asking ChatGPT, not Google. What I learned checking whether that was luck, and why you cannot see it happening to you.
He said it almost as an aside, twenty minutes into the call, the way you mention the weather.
He said it almost as an aside, twenty minutes into the call, the way you mention the weather.
"I asked ChatGPT who does this sort of thing in Brisbane. You came up."
Thirty years in this industry and I had never heard that sentence. Referrals, yes. A directory, occasionally, back when directories still meant something. Page one of Google, sometimes, on a good month. But not this. Something had formed a view about my business and handed it to a stranger, and nobody had told me it was happening.
So I did what you would do. I opened ChatGPT and asked it the same question.
It named us. It also named two competitors. One of them stopped trading in 2023.
That is the detail I keep coming back to. Not that we were in the answer. That a dead business was standing next to us in it, and neither of us had any say in the matter.
The part that should worry you
A search result is something you can inspect. You type the query, you see where you sit and who is above you, and if it is wrong you can usually work out why.
An AI answer is not like that. Two people asking the same question in the same week get different answers, and none of them leaves a page you can point at. The levers that do exist are ordinary ones, and what actually earns a place in an AI Overview is a shorter list than the opacity suggests.
Google has closed a sliver of the gap. Search Console now carries a generative AI report covering AI Overviews and AI Mode (Google Search Console Help). Worth turning on.
Read the limits, though. The metric it publishes is impressions, "how many times links to your site were shown to a user in a generative AI feature on Google Search" (Google Search Console Help), broken down by page, country, device and date. Google documents no clicks, no query dimension, and no control for separating AI Overviews from AI Mode. You learn that you were shown, not what was said or whether it was true.
Microsoft is further ahead than Google here: Bing Webmaster Tools has reported AI citations since February 2026 (Microsoft), so it will name the pages of yours that got used. For ChatGPT and Perplexity you get nothing beyond the referrals that land in your analytics, and only when someone clicks.
I went looking for the mechanism
I am a physicist by training and a network engineer by trade, and I do not like outcomes I cannot explain. So I spent a fortnight on it.
The models draw on three things at once, and only one of them is your website. Sometimes it is a live search at the moment the question is asked. Sometimes it is training data from a crawl taken years ago. And sometimes it is everybody else's pages about you, which for most small businesses means a directory listing nobody has touched in a decade.
You can influence the first. You cannot edit the second. The third is the one that quietly beats you, because a detailed directory entry is easier to quote than a thin page on your own site.
The uncomfortable arithmetic
There is an argument running about whether any of this costs you traffic, and both sides are telling the truth about different things.
Google says total organic click volume to websites "has been relatively stable year-over-year" (Google). Pew measured American users rather than global totals, and reported in 2025 that people who saw an AI summary clicked a search result in 8% of visits, against 15% where no summary appeared (Pew Research Center).
Both can hold. A lower click rate spread across more searching keeps the total roughly level while every individual page feels thinner. No equivalent Australian measurement exists that I would put in front of you.
Either way, some of the people who would once have landed on your site now get your answer without you. The question is no longer whether you rank. It is whether the answer they get is yours.
What I actually changed
Three things, none of which took a week.
I fixed the contradictions first. Our phone number appeared three different ways across our own site, an old listing and a profile I had forgotten existed. A model given three versions of a fact will pick one, and it cannot know which one you would have chosen. That was housekeeping I had carried for years, unpunished, until something started reading it.
I checked what we were telling the crawlers. Most people have never looked, and not every crawler treats robots.txt the same way. Across the Australian small-business sites in our September research, 13% had any AI-crawler rule at all*. A scan cannot tell you whether the rest decided to allow everything or never considered it. Only you know which one you are.
If you do go looking, the tokens are not interchangeable. OpenAI runs a separate agent for search from the one that collects training data, and sites opted out of the search one "will not be shown in ChatGPT search answers" (OpenAI). Block the wrong one and you remove yourself from the thing you were trying to win. The wider version of this argument is what people mean by generative engine optimisation.
Then I wrote down six questions a customer might actually ask. Not our business name. Our name is a test we always pass, and it tells us nothing. Six real buying questions, asked across four tools, written down with the dates and the sites that got cited. It took an afternoon. I do it every quarter now.
What I do not know
Whether any of it worked.
One customer is an anecdote. I have readings from one quarter, which is a baseline and not a trend. Anybody telling you they can reliably get your business into an AI answer is selling something they cannot prove.
What I am confident about is narrower and duller. If a model is going to describe your business to someone who has never heard of you, it will use whatever it can find. If what it can find is stale, thin, or contradicts itself, that is what it will say.
Nobody is going to tidy that up for you. Your web developer moved on. The directory does not know you exist. The model will not ask.
So here is the question I would sit with. If a stranger asked a chatbot about your line of work tomorrow, and it answered confidently, and you never found out what it said, how would you feel about the version of your business it described?
Sources
- Google Search Console Help, Generative AI performance report (Search) - retrieved 19 September 2026, https://support.google.com/webmasters/answer/16984139?hl=en
- Google, AI in Search is driving more queries and higher quality clicks - retrieved 19 September 2026, https://blog.google/products-and-platforms/products/search/ai-search-driving-more-queries-higher-quality-clicks/
- Pew Research Center, Google users are less likely to click on links when an AI summary appears in the results - retrieved 19 September 2026, https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/
- OpenAI, Bots and crawlers - retrieved 19 September 2026, https://developers.openai.com/api/docs/bots
- Red Bridge Cyber, Australian SMB Posture Baseline, September 2026, https://redbridgecyber.com.au/research/smb-posture-baseline
- Microsoft, Introducing AI Performance in Bing Webmaster Tools - retrieved 19 September 2026, https://blogs.bing.com/webmaster/February-2026/Introducing-AI-Performance-in-Bing-Webmaster-Tools-Public-Preview