A meaningful share of people no longer open search results one by one. They ask an assistant: which companies in Auckland build bilingual websites, roughly what does a renovation cost in New Zealand.
If the answer does not include you, you do not exist on that path — there is no second page to turn to.
This is entirely testable yourself, without tools and without spending anything. Here is how.
Rule one: do not ask whether it knows you
Most people’s first test is “do you know about such-and-such company”. The answer is worthless, for two reasons.
You have handed it the name, so it only has to confirm rather than retrieve and choose. In reality your customer does not know your name — they know their problem and not who solves it.
And models asked directly about a specific name tend to produce affirmative, plausible-sounding answers even on thin information.
Ask the questions your customers ask, without mentioning your name at all.
Prepare a list of questions
Write ten to fifteen covering different stages. Using ourselves as the example:
Finding a supplier: “which companies in Auckland build bilingual websites”, “where can I find a Chinese-speaking web developer in New Zealand”
Price: “what does a business website cost in New Zealand”, “annual cost of a small business website”
Approach: “I’m opening a clinic in Auckland, what kind of website do I need”, “should a Chinese-owned business have a Chinese site or a bilingual one”
Concepts: “what is GEO”, “does a small New Zealand business need SEO”
Write the equivalent for your industry. Use your customers’ words, not your industry’s. Nobody searches for “responsive web design solutions”; they search for “my website doesn’t work on my phone”.
Test four models
Their sources and retrieval differ substantially and the answers can be completely different. Test at least:
ChatGPT — the largest audience.
Claude — more conservative, higher bar for asserting things.
Gemini — closest to Google search.
Perplexity — retrieval-led, and it lists its sources.
Perplexity is particularly worth testing because it shows citations explicitly. Seeing who it cites tells you whose content is being treated as credible on that topic, which is more direct than any ranking data.
Start a new conversation each time. Context from earlier in a session influences later answers and produces optimistic conclusions.
Record three things each time
1. Whether you appear. If so, note whether you were recommended or merely listed.
2. Whether what it said is accurate. Frequently overlooked and more important than appearing at all. An assistant stating something wrong about you is worse than omitting you — a wrong price, an overstated service range, being confused with another company.
We found this ourselves: a model described our scope as broader than it is, and a client arriving on that basis spends the first conversation being corrected.
3. If you do not appear, who does. The most useful of the three.
Study what the named companies did
Open their websites and compare on these points:
Are prices published? “Contact us for pricing” cannot be quoted, because restating it conveys nothing. Companies with figures have a structural advantage here.
Is the scope clear? “We build all kinds of websites” carries nothing. “We build company sites and booking systems, not e-commerce platforms” is easier to recommend, because an assistant can judge whether to mention it.
How much is verifiable? Registration number, founding date, address, real client work — things that correspond to outside records, which makes them low-risk to cite.
Is the information consistent? The same company name and description on the website, the business profile and the social accounts. While they disagree, an assistant cannot confirm these are one company.
After that comparison the gap is usually obvious, and most of it is copy rather than technology.
How often to test
Quarterly is enough. These systems update with a lag, so testing frequently shows nothing and invites over-reading single fluctuations.
Two moments deserve an extra test: after changing prices or scope, to confirm the current information is what is being repeated, and after a significant change to the site.
Record the date. Without your own history there is no way to see a trend, and these systems provide no dashboard.
What to do about what you find
If it states something wrong: check whether that information is clear and consistent on your own site first. Most errors come from vague or contradictory source material rather than invention.
If you are absent entirely: that is a content problem, not a technical one. Look at what the named companies published that you did not — usually specific numbers, defined boundaries, and verifiable facts.
If you appear with outdated information: these systems are not real-time, so allow time after updating. Also confirm you updated everywhere — llms.txt, structured data and page copy — since changing one leaves the others stale.
A realistic judgement
How much effort does this warrant now?
Honestly: for most small New Zealand businesses, AI currently produces far fewer enquiries than search and referral. It is a growing entry point rather than a primary one.
So the sensible approach is doing the parts that cannot be wasted however these systems change — publish prices, define scope, make company information consistent everywhere, put verifiable facts on the page. All of it also underpins SEO and conversion, so there is no wrong bet available.
Model-specific techniques can wait until things settle. At the current rate of change, any technique has a short shelf life.
Our SEO and GEO service includes this multi-model testing, running the process above with more structure and a record over time. There is nothing proprietary in the method; you can run it yourself.