It searches, then writes up what it finds
Ask an assistant to recommend a builder in Tauranga and it does not consult a directory of builders. It turns your question into several searches, reads whatever comes back, and writes an answer out of that.
Google publishes how this works and has a name for each half. The retrieval part is grounding, "relying on our core Search ranking systems to retrieve relevant, up-to-date web pages from our Search index". The several searches are query fan-out, "a set of concurrent, related queries generated by the model". Its New Zealand launch post for AI Mode put it more plainly: "breaking your question into subtopics and issuing a multitude of queries on your behalf".
It is not assessing you. It is reading whatever ranked for a handful of searches you never saw. One of those searches was probably not "builder Tauranga" at all but something narrower, along the lines of "builders who do villa renovations Bay of Plenty". If nothing that came back for that one mentions you, there is nothing for the assistant to name you from.
Not every assistant reads Google's index. Claude's search runs on Brave, and on ChatGPT the people who follow this closely disagree: Bing and Google have both been named in the last year, twice by the same practitioner six months apart. OpenAI does not publish it. What you can act on is that being indexed in one search engine is not being indexed in all of them, and Bing is the one most businesses have never looked at.
You cannot reach into the answer itself. Once the pages are pulled, the model weighs passages against each other and discards most of them, somewhere you have no visibility. What you can affect is whether you are in the set it pulls from, and whether what it finds there is specific enough to be worth keeping.
So getting into AI results is mostly the same job as getting found on the web at all. Google says so directly: asked whether SEO still matters for AI search, its guidance answers "In short, yes!" Darren Shaw, who tests Google's local AI surface for a living, reckons ranking well in ordinary local results gets you 80% of the way into its AI recommendations.
The industry that grew up around the question
Google has published a list of the AI optimisation tactics that do not work. The question is new enough that plenty of people are selling them anyway.
- Special files for AI. You do not need an llms.txt file or any other machine readable version of your site, because "You don't need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search".
- Chopping your pages into fragments. "There's no requirement to break your content into tiny pieces for AI to better understand it."
- Writing in a special way. There is no target word count and no house style that AI prefers: "You don't need to write in a specific way just for generative AI search."
The one that costs owners real money is mentions. Services will sell you bulk mentions across the web on the basis that this is how you get into AI answers. Google's line on that is "Seeking inauthentic 'mentions' across the web isn't as helpful as it might seem."
The only requirement Google names for a page to be used as a supporting source is that it is "indexed and eligible to be shown in Google Search with a snippet". That is the same bar as showing up in ordinary search results.
None of the work is in making your website AI ready. It is in giving these systems something specific to find about your business.
The fake awards body that ranked in three days
Retrieving a page from the search index is not the same as checking whether it is true, and several sources agreeing with each other is not the same as any of them having checked. Google's language around grounding does not make that distinction obvious.
In August 2026 an SEO consultant called Charles Floate showed how big the gap is. He registered a domain, had AI build an awards body from scratch, and made up the judges. He added a Wikipedia page, seeded a Reddit thread, and asked a few people to write about it on their own sites. He built no links to it at all, on the reasoning that mentions of the name across enough trusted places would carry it on their own.
Within three days the site was ranking first for its term, and Google's AI Overview was describing his fabricated ceremony as "the most prestigious awards ceremony in the entirety of SEO". Google took it down not long after, and he published the whole exercise as a warning.
Set that against the mythbusting list above, which says chasing mentions "isn't as helpful as it might seem". Floate's site ran on mentions and nothing else. Both can be true, since Google's spam systems did catch it within days, but they are not describing the same system.
Whatever an assistant tells you about a competitor, or about your own business, it has repeated and not checked. Worth remembering before you panic about it or quote it at someone. It also explains why awards and best-of badges are being sold so hard: Joy Hawkins of Sterling Sky reports seeing businesses with 2 and 3 star averages surface for "best" queries on the strength of an award, because these systems cannot tell an earned one from a purchased one.
Apply that to us as well. Nobody can buy a position on a Top5.nz page and our badges are not for sale, but you have only our word for that until you check it, which is why the method is published in full. A best-of page that will not tell you how it ordered its list is the one to be wary of.
What to do about it
Before changing anything, run two searches. Ask Google for "plumber Tauranga", then ask it "who are the best plumbers in Tauranga", and compare the lists. Caleb Ulku ran that pair in Chicago and got no overlap at all: the map pack returned one set of businesses and the AI answer returned three different ones. That was a single test in one city, so run it on your own trade and see what you get.
The rest is ordinary local search work.
- Complete your Google Business Profile. Google states that "businesses with complete and accurate info are more likely to show up in local search results", and lists prominence as partly a matter of "how many websites link to your business and how many reviews you have". The specific gap on most profiles is the services list, which is filled in with no descriptions attached. That text is read as source data rather than decoration, so if you are the only firm in your city doing a particular job, write it down there.
- Keep your details consistent, then make the listings worth reading. Name, address and phone number matching across your own site, your profile and anywhere else that lists you is the floor, not the work. Google can usually reconcile a messy set of records, but conflicting names, numbers and addresses make it harder, and hardest where two similarly named firms trade in the same town. Past that, treat each listing as a page rather than an address record, with a real description and photos. A listing on a site that verifies a business before publishing it also carries further than one anybody can create, which is the job a Top5.nz profile page does for subscribers: verified details and services on a page we keep current rather than leaving to go stale.
- Get reviews with words in them. Review text is what matches you to a specific question, where a bare star rating matches nothing. Ask the customer to say what the job was and how it went. Do not hand them wording or ask for particular phrases, because a review written to your dictation is not evidence of anything. There is more on that in your reviews are your description.
- Put the awkward information on your site. Prices or price ranges, the areas you actually cover, your registration number where your trade has one, the jobs you do not take, and who you are not the right firm for. A customer skim-reading for thirty seconds never got that far, but a retrieval system can pull a detail from well down a page, and a question like "who can fix a burst pipe tonight in Papamoa" gets matched on those specifics rather than on your trade and city.
- Name yourself in the sentence. Most trade sites are written in the first person with the business name nowhere near the claim: "we have been doing bathrooms for 15 years". Anything assembling an answer has to work out who "we" refers to before it can use that. "Papamoa Plumbing has fitted bathrooms since 2011" is the same fact with the subject attached, and it still means something lifted out of the page on its own. That is not a special style for AI, which Google says above you do not need. It is writing that does not assume the reader already knows whose site they are on.
- Check you are indexed in Bing, not only Google. Bing Webmaster Tools is free and imports any domain already verified in Google Search Console, so it costs a couple of minutes. It is also blunter than Google about why a site is missing: where Search Console reports "crawled, not indexed" and leaves you to guess, Bing tends to say outright that the site does not have enough inbound links from good domains.
- Be on the pages that already rank for your trade. The association listing, the supplier's page, the local coverage, and the ranking pages for your category and city, ours included. Being accurate on a handful of pages that already get retrieved beats having profiles on twenty sites nobody returns. Bought links are a different thing and Google's spam policy has its own view about them, which our guide to judging a link gets into.
Then measure it, and do not use your own logged-in account to do it. Your results are shaped by your own search history, so what you are shown is not what a stranger is shown. The method Darren Shaw recommends is deliberately crude: pick the five questions you would most want to be named for, run each of them three or four times, and write down how often you appear. An appearance rate is the number worth having, because a position is not a stable thing on this surface. Once a quarter is often enough. Anyone offering to guarantee you a place in an AI answer is selling something they do not control.