Someone in your city asked ChatGPT how AI chooses which business to recommend for a business like yours last week. It gave them three names. Yours probably wasn’t one of them.
That is the shift most owners haven’t caught up with. Buyers are asking an assistant instead of scrolling ten blue links, and the assistant doesn’t hand back a page of options. It hands back a shortlist, usually three names, sometimes one. The process is not random, and it is not the same thing as ranking first on Google.
I have spent the past year pulling this apart across my own business and my clients’. This post covers what actually feeds those answers, the six signals that decide who gets named, and the specific work that puts you on the shortlist.
Key Takeaways
- How AI chooses which business to recommend comes down to retrieval, not memory: it searches, reads a handful of pages, then names what it can verify.
- Assistants repeat corroborated facts. A claim that only appears on your own website rarely survives the summary; the same fact across three independent sources does.
- Inconsistent business names, addresses and phone numbers split your business into several half-entities and quietly disqualify you from being named.
- Question-shaped headings with the answer in the first two sentences are the easiest thing on a page for an AI to lift and quote back to a buyer.
- Review text matters more than the star average, because the words your customers use are what match the words a buyer types into the assistant.
- Third-party lists, directories and forum threads get quoted more often than your homepage, so being absent from them costs you the shortlist.
- Test monthly with ten real buying prompts across ChatGPT, Google AI Overviews, Claude and Perplexity, and log which sources each answer cites.
- AI referrals arrive pre-sold and impatient, so slow follow-up burns the visibility you just spent months earning.
How AI Chooses Which Business to Recommend: The Mechanics
Most owners assume the model has an opinion about their business stored somewhere. It doesn’t. Understanding what actually happens under the hood changes what you do about it.
It looks you up; it doesn’t remember you
When someone asks an assistant for a recommendation, the model doesn’t reach into training data and pull out a favourite. For anything local, current, or commercial, it runs a search first, pulls back a set of pages, reads them, and writes an answer grounded in what it just read.
That means your visibility is decided at the moment of the question, not at the moment of training. It also means the answer can change week to week, which is good news if you are currently invisible.
Different assistants search differently. ChatGPT leans on its own search index and Bing signals. Google’s AI Overviews sit directly on top of Google’s index. Perplexity runs its own crawl. Claude searches the live web. There is no single algorithm to game, which is exactly why the fundamentals matter more than the tricks.
It reads a handful of sources, then summarises
The model doesn’t read a hundred pages. It reads somewhere between three and a dozen, depending on the assistant and the question. It then writes a confident-sounding paragraph that stitches those sources together.
So there are two gates, not one. Gate one: does your content get retrieved at all? Gate two: once retrieved, does it contain a clear, liftable answer the model can restate without hedging?
Most businesses fail gate one because nothing they publish matches how people actually ask. Most of the rest fail gate two because their website talks about “solutions” and “excellence” instead of saying who they serve, where, and what happens next.
The shortlist is brutally short
Ten blue links gave you a decent chance at position six. An AI answer gives you three names, and the buyer usually contacts the first one. There is no page two.
That compression is the whole story. In traditional search, being the eleventh-best result still earned you a trickle. In AI search, being the fourth-best option earns you nothing at all. The gap between “we show up somewhere” and “we get named” is now the entire difference.
The Six Signals That Decide Who Gets Named
Across the tests I have run and the client accounts I have worked on, the same six things separate the businesses that get named from the ones that don’t.
1. Entity clarity
The model has to be certain that the business it is about to recommend is one real, identifiable thing. If your company name appears three different ways across the web, if an old address still sits on two directories, if your phone number changed in 2024 but half the internet didn’t get the memo, you have fractured yourself into several weak half-entities instead of one strong one.
Fix the boring stuff first. One legal name, one trading name, one address format, one phone number, everywhere. This is unglamorous work, and it moves the needle more than anything else on this list.
2. Corroboration across independent sources
This is the signal most people miss. A model treats a claim on your own website as a claim. It treats the same claim, repeated on a directory listing, an industry association page, a local news mention and a review site, as a fact.
If the only place on the internet that says you specialise in commercial refrigeration in South Auckland is your own homepage, the assistant has nothing to verify against. It will reach for a competitor whose specialisation is echoed in four places.
3. Structured data
Schema markup is how you hand a machine an unambiguous version of your business details instead of making it guess from your page copy. LocalBusiness, Organisation, Service and FAQPage markup all give assistants and the search engines feeding them clean facts: what you do, where you operate, when you are open, what you charge for, who vouches for you.
Google publishes the exact specification in its local business structured data documentation, and it takes a competent developer an afternoon. Almost no small business in New Zealand has done it properly, which is precisely why it’s worth doing.
4. What your reviews actually say
Star averages barely move an AI answer. The language inside reviews moves it a lot.
If a buyer asks for “an electrician who can come out same day,” the assistant is matching that phrase against text it can find. Five reviews that say “came out the same afternoon” will beat a 4.9 average with fifty reviews that all say “great service, highly recommend.”
So when you ask for reviews, stop asking for reviews. Ask the customer what problem you solved and how fast. You are not collecting sentiment; you are collecting the vocabulary your next buyer will type.
5. Specificity of match
Models hedge when they are unsure and commit when they can justify the pick. Give them a reason to commit.
“We service commercial kitchens in Auckland and Waikato, twenty-four hours, with a two-hour callout guarantee” is recommendable. “We deliver quality outcomes for our clients” is not. The second one gives an assistant nothing to attach the buyer’s question to.
Say who you are for. Say who you are not for. Naming your floor (“we work with practices running three chairs or more”) makes you dramatically easier to recommend correctly, and it filters out the enquiries you didn’t want anyway.
6. Freshness
Content with a visible recent date, current pricing structure, current team, current service area, gets preferred over a page that reads like it was written three years ago. Assistants are cautious about recommending a business they can’t confirm still operates the way the page describes.
A quarterly pass over your core pages, updating dates, examples and service details, is enough. Nobody needs to publish daily.
Why Your Website Is Only Part of the Answer
Here is the uncomfortable part. When I run test prompts across assistants for a client, the sources cited are rarely the client’s own website. They are directory pages, “best of” lists, forum threads, review platforms, LinkedIn profiles and the occasional podcast transcript.
Your website is the thing that confirms the recommendation. Other people’s pages are what trigger it.
That reframes the work. Publishing more blog posts on your own domain has a ceiling. Getting your business named, described accurately and linked on surfaces the assistants already trust has almost no ceiling.
The practical version of that for a New Zealand founder-led business looks like this. A complete and current Google Business Profile with real photos and answered questions. Listings on the two or three directories that actually matter in your industry, not forty spammy ones. Membership pages on trade bodies and associations. A LinkedIn company page that says the same thing your website says. Being included in local “best X in Y” roundups, which usually just requires asking the person who writes them. A handful of genuine third-party mentions: a customer story on a supplier’s site, a quote in an industry newsletter, an interview on a regional podcast.
None of that is new marketing theory. What is new is the payoff. Those mentions used to be nice-to-have brand touches. Now they are the raw material an assistant reads before deciding who to name.
The same logic applies to community discussion. Reddit threads, Facebook group posts and industry forums get pulled into AI answers constantly, because they read as unfiltered human opinion. You cannot manufacture that, and you should not try. What you can do is be the business people have a reason to mention, and be findable by the correct name when they do.
How to Become the Business AI Recommends
This is the sequence I run. It works because it goes in order of what actually blocks you, not in order of what is most fun to build.
Step one: write the source-of-truth page
One page on your website that states, in plain language, exactly what you do, who you do it for, where you operate, how you charge, what happens when someone contacts you, and what makes you the wrong choice. That last part matters more than people expect. A page that admits who it is not for reads as trustworthy to both humans and models.
Write it as if you are briefing a new staff member on their first day. No positioning language. No adjectives you can’t prove.
Step two: rewrite your headings as questions
Go through your key pages and turn statement headings into the question a buyer would actually ask. “Our Process” becomes “What happens after I get in touch?” “Service Areas” becomes “Do you cover the North Shore?”
Then answer it in the first two sentences underneath, before any preamble. That two-sentence answer is the unit an assistant lifts. If your answer arrives in paragraph four after a windup, it never gets quoted.
Step three: fix your consistency, everywhere
Audit every place your business appears online and make the details identical. Name, address, phone, hours, service area, category. Kill or correct the dead listings. This is a two-hour job that most businesses have never done once.
Step four: add the markup
LocalBusiness or Organisation schema on the site, FAQPage schema on any page with a genuine question and answer block, Service schema on service pages, and sameAs links pointing at your verified profiles elsewhere. The schema.org LocalBusiness vocabulary lists every field available. Fill in the ones that are true and leave the rest.
Step five: build the corroboration
Pick five external surfaces this quarter. One profile you control, two directories or association listings, one roundup or list you ask to be included in, one genuine third-party mention. Then do it again next quarter. Twenty accurate mentions of your business across a year is a completely different position than three.
Pair that with a deliberate review habit. Ask for specifics, not stars. “What was the problem, and how quickly did we sort it?” produces text an assistant can match against a buyer’s question.
Step six: be ready to answer
This is where most of the effort gets wasted, and it’s the part nobody plans for.
An AI referral is not a cold lead. That person has already been told you’re the right choice by something they trust. They have not compared five quotes. They are ready now, and they are contacting one, maybe two businesses.
If that enquiry hits a voicemail, an inbox nobody checks until Tuesday, or a contact form that routes to a staff member on leave, you have paid for months of visibility work and handed the job to whoever answered second. Research on lead response consistently shows the first business to respond wins around 78% of the time, and AI-referred buyers are the least patient cohort you will ever get.
One of my clients, Dr Claire, was losing calls to a busy front desk despite having two receptionists on. Once every call was answered, qualified and booked automatically, her booked appointments went up 44%. Nothing about her demand changed. Only her ability to catch it did.
Being recommended is the hard part. Being ready to answer is the cheap part. Do both, or don’t bother with either. If your phone coverage is the weak link, after-hours call answering is usually the fastest thing to fix.
How to Test Whether AI Chooses Your Business Right Now
Stop guessing. Run the test. It takes forty minutes, and you will learn more than any report will tell you.
Write ten prompts the way a real buyer would type them. Not keywords. Full questions, with location and constraint included. Things like: “Who’s the best commercial plumber in Hamilton for a restaurant fit-out?” “I need an accountant in Auckland who works with trades businesses, who should I call?” “Best emergency locksmith Wellington open now?” “Which company should I use for X, and why?”
Run every one of them in ChatGPT, Google (watching the AI Overview, not the blue links), Claude and Perplexity. Four assistants, ten prompts, forty answers.
For each answer, log four things. Were you named. Who was named instead. Which sources the answer cited or linked. And, when you were named, was the description of your business actually correct?
That last column is the one that shocks people. Assistants regularly describe businesses using services they dropped years ago, suburbs they no longer cover, or a positioning they abandoned. Every one of those errors traces back to a stale page somewhere that you can go and fix.
The cited-sources column is your work list. If the same three directories keep appearing across forty answers, those three directories are where your listing needs to be complete and current. If a competitor’s name keeps surfacing through a roundup article, that roundup is worth an email to the author.
Re-run the same ten prompts monthly and keep the log. Answers vary run to run, so a single test is a snapshot and three months of tests is a trend. Once you can see the trend, you know whether the work is landing.
This is also exactly the kind of recurring, structured, slightly tedious task that should not live on a founder’s to-do list forever. Ten prompts, four assistants, one log, once a month, is a job for a system, which is the same argument I make about AI agents for small business generally. Do it manually twice so you understand the output, then hand it over.
What This Actually Adds Up To
How AI chooses which business to recommend comes down to something unfashionably simple. It picks the business it can verify, describe accurately and justify recommending, using sources it did not have to take your word for.
Which means the work is not a trick. It’s clarity about who you serve, consistency everywhere your name appears, other people saying the same thing about you that you say about yourself, and content that answers a real question in the first two sentences. Then the ability to actually pick up when the phone rings.
The businesses that will own this are not the ones with the biggest content budget. They’re the ones who set up a system to do this work every month without the owner remembering to. That’s the wider pattern behind everything I build: an AI brain that holds what the business knows, and an AI workforce that does the recurring work so the founder stops being the bottleneck. If you want the broader version of that argument, start with AI adoption for SMEs or the answer engine optimisation guide in this series.
If you want a straight read on where your business currently sits in AI answers and what to fix first, book a 30-minute Discovery Call. I’ll run a handful of live prompts for your market on the call and tell you honestly whether this is your bottleneck or whether something else is costing you more. No deck, no pitch, just the diagnosis.
Frequently Asked Questions
How does ChatGPT decide which businesses to recommend?
For local or commercial questions, ChatGPT runs a live search, retrieves a small set of pages, and writes an answer grounded in what it read. It favours businesses whose details are consistent across multiple independent sources, whose pages directly answer the question asked, and whose services are described specifically enough to justify the recommendation.
Why doesn’t AI recommend my business?
Usually one of three reasons. Your details are inconsistent across the web, so the assistant can’t confirm you’re one real business. Your claims appear nowhere except your own website, so there’s nothing to corroborate. Or your pages talk in general marketing language instead of stating plainly who you serve, where, and what happens next.
Can you pay to be recommended by AI?
Not directly, and be sceptical of anyone selling guaranteed placement. Assistants generate answers from retrieved sources, not from an ad auction. Paid advertising slots are starting to appear alongside AI answers on some platforms, but the recommendation itself comes from what the model can find and verify about you across the open web.
Does traditional SEO still matter for AI recommendations?
Yes, because most assistants search the web before answering, and several sit directly on top of existing search indexes. If you’re invisible in search, you’re invisible to retrieval. The difference is what you optimise for: clear question-shaped answers and verifiable facts rather than keyword density and link volume.
How do I find out what AI says about my business?
Write ten questions a real buyer would ask, including location and constraints, then run each one through ChatGPT, Google AI Overviews, Claude and Perplexity. Record whether you were named, who was named instead, which sources each answer cited, and whether your business was described accurately. Repeat monthly and watch the trend.
How long does it take to get recommended by AI?
Faster than traditional search rankings, because retrieval happens at question time rather than waiting on an index update. Consistency and schema fixes can show up within weeks. Building enough independent corroboration to be named reliably usually takes one to two quarters of steady work across directories, listings and third-party mentions.
About Octavius
Titus Mulquiney is the founder of Octavius AI, where he builds AI brains and AI workforces for founder-led businesses stuck running everything out of their own head. Twenty years in marketing, ex-Sony product manager, ex-GM Zeal NZ. Based in Auckland, working with operators across NZ, Australia, and the US. Connect on LinkedIn.