A few weeks back a new client told me how they found us. A mate recommended us, they googled us, then they asked ChatGPT if we were any good. The call only happened because all three came back clean.
Ask ChatGPT for a plumber or a physio yourself and you’ll see why that last step matters. You don’t get ten blue links. You get three names, maybe five, a line on each, and a reason to pick one. Being findable used to mean ranking. Now it also means being picked, by a system you can’t log into and can only partly measure.
Here’s what’s actually happening under the bonnet, what Google will and won’t tell you about it, and the work that still moves the needle. If you want the companion piece on why you might be missing from those answers entirely, we covered that in why ChatGPT recommends your competitors and not you.
How does AI search actually pick which businesses to mention?
Two techniques do most of the work: retrieval augmented generation and query fan-out. Neither is as complicated as it sounds.
Retrieval augmented generation (RAG) means the model doesn’t answer from memory alone. It fetches live pages from the search index first, reads them, then writes an answer grounded in what it found. This is why your website content matters as much as your general reputation. If there’s nothing fetchable that answers the question well, you can’t be part of the answer.
Query fan-out is the other half. Instead of running the one question you typed, the system breaks it into a set of related sub-queries and runs them all at once. Ask for “a good physio in Newcastle for a running injury” and behind the scenes it’s likely searching for physios in Newcastle, sports physiotherapy, running injury treatment, opening hours and reviews, plus a few angles you’d never have thought to type. It then assembles one answer from the best of all of that.
So you’re not competing for the one query any more, you’re competing across a dozen questions you never saw. There’s a number worth sitting with here. When Whitespark tested 540 local queries across six industries this year, Google’s AI answers showed up on only 15 per cent of the direct “physio near me” style searches. On hybrid questions, the ones that carry a buying decision inside them, “should I see a physio for this knee or wait it out”, they showed up 97 per cent of the time. Your customers are meeting the AI answer while they’re still deciding whether to buy, not just who from.
The practical upshot is that narrow keyword optimisation matters less than it used to. Covering your topic properly, in the language a real customer would use, matters a lot more. That’s the core of the GEO side of the work, and it’s a slower build than chasing a single phrase.
Why do listings and reviews still do the heavy lifting?
Because the retrieval step has to find something it trusts.
An AI assistant naming a business is making a small reputational bet on your behalf. It favours businesses it can describe confidently: the ones where the name, address, phone number, hours, categories and service description line up everywhere it looks. Where those details are inconsistent, the system plays safe and names someone clearer. We go into that in more depth in entity SEO.
Reviews are the other half of it, and not only for the star rating. Review text is retrievable content written in customer language, describing what you actually do and how well you do it. Volume, recency and specificity all count. A steady trickle of recent, detailed reviews is worth more here than a pile of three-word ratings from 2021.
It’s unglamorous work. It’s also the highest-return thing most local businesses can do, and it costs nothing but attention.
Can you actually measure any of this?
Partly. This is the uncomfortable bit, and anyone selling you a clean dashboard is overselling.
Google’s reporting on AI features will show you impressions and clicks for links surfaced there. That’s useful, and it’s more than we had a year ago. But it’s aggregate. You don’t see which question triggered you, you don’t see who else got named alongside you, and you certainly don’t see the answers where you weren’t mentioned at all. It also only covers Google. Nothing in it tells you what ChatGPT, Perplexity or Copilot said about your industry this morning.
So most businesses fall back on two workarounds.
Fixed query monitoring. Pick twenty or thirty questions your customers actually ask, run them across the major assistants on a set schedule, and record whether you get named and who else does. It’s crude, but it’s honest. The catch is that answers vary between sessions, accounts and locations, so you’re sampling rather than measuring. Run them logged out, and run them consistently, or the comparison is worthless.
Referral traffic analysis. Filter your analytics for referrals from the AI platforms. Real signal, but it undercounts badly, because plenty of people read the answer, form a view, and arrive later through a branded search or a direct phone call. In our experience being named in an AI answer shows up as a phone call more often than a session.
Neither method is a reporting suite. Treat both as directional, watch the trend rather than the number, and don’t make big decisions off a single week.
What does Google actually recommend you do?
Google’s own guidance for being eligible in AI features is refreshingly unromantic. Three things:
- Put your key business information in text on the page, not locked inside an image, a PDF or a video with no transcript.
- Keep your structured data accurate and matched to what’s actually visible on the page.
- Keep crawling permissions open, so the systems that want to read you can.
Two things Google says it does not do, since both are being sold to local businesses right now: it ignores llms.txt and similar files, and it does not need your content chopped into special AI-sized chunks. Plain, well-structured pages are the whole brief. One caveat from the same guidance that lands squarely on the build side of our work: sites that render their content through JavaScript are, in Google’s words, generally more complex to work with. If your key information only exists after a script runs, you’re making the reader work harder than it needs to. We’ve inherited sites like that, and the fix is usually a rebuild decision, not a content one.
Nothing exotic. It’s the same fundamentals that have made sites findable for two decades, pointed at a new kind of reader. We wrote up the local application of that in local SEO and structured data.
Why has accessibility become an AI readiness issue?
This is the bit most businesses haven’t clocked yet, and it’s the one I’d put money on mattering more over the next couple of years.
AI browser agents that carry out tasks, booking an appointment or filling in an enquiry form, often read a page through its accessibility tree. That’s the structured representation browsers build for screen readers. It’s the same map, used by a different reader.
Which means the accessibility work you might have parked as a compliance chore now determines whether an agent can complete a booking on your site. Unlabelled buttons, form fields without labels, fake buttons built out of divs, custom widgets with no ARIA: a screen reader user struggles with all of it, and so does the agent. If the agent can’t work out which control submits the form, it moves on to a competitor whose site it can read.
Building an accessible site and building an agent-ready site turn out to be the same job. That’s a rare thing in this business, one piece of work that pays off twice.
Where should you start?
In this order, because the early items are cheap and the later ones compound:
- Get your business information identical and in plain text everywhere: website, Google Business Profile, directories.
- Check your structured data is accurate and reflects what’s on the page.
- Confirm nothing in your robots file, firewall or security plugin is blocking the crawlers you want.
- Fix accessibility on your conversion paths first: enquiry forms, booking flows, contact details.
- Keep reviews recent and encourage detail, not just stars.
- Set a baseline with twenty or thirty fixed queries, then re-check quarterly.
None of that is a trick, and none of it is beyond a small business. The businesses doing this work now, while most are still deciding whether AI search is real, are the ones that’ll be named for the next few years.
We see the gap most weeks. When we ran the check for one long-standing Newcastle manufacturer, holding page 1 Google positions for their core product terms for years, Perplexity answered their customers’ buying question by naming eight of their competitors. They weren’t on the list. Their listings were inconsistent, their site described products rather than answering questions, and nobody had ever asked the engines about them. Every one of those is on the list above.
Figures in this article are drawn from Google’s published guidance on AI features, Whitespark’s 2026 local query analysis and Uberall’s quick-service benchmark, as reported by Search Engine Journal.
Frequently asked questions
How do AI assistants decide which local businesses to recommend?
They fetch live pages from a search index and write an answer grounded in what they find, a technique called retrieval augmented generation. Most also use query fan-out, breaking your single question into several related searches and assembling one answer from all of them. Businesses that get named are the ones with consistent listing information, content that answers real questions in plain text, accurate structured data and open crawling permissions. AI answers appear far more often on questions that contain a buying decision than on simple “near me” searches, so answering those questions on your site matters most.
Can I track whether my business appears in AI search results?
Only partly. Google’s reporting on AI features shows aggregate impressions and clicks for your links, but not which query triggered them, who else was named, or what happened outside Google. The practical alternatives are running a fixed set of customer questions across the major assistants on a schedule and recording the results, and filtering your analytics for AI platform referrals. Both are directional rather than exact.
Does website accessibility affect AI visibility?
Increasingly, yes. AI agents that complete tasks on websites, such as making a booking or submitting an enquiry, often read pages through the accessibility tree that browsers build for screen readers. Unlabelled buttons, missing form labels and custom controls without ARIA make a site hard for both screen reader users and AI agents. Fixing accessibility on your conversion paths improves both at once.
Is optimising for AI recommendations different from local SEO?
They overlap heavily. Consistent business information, genuine reviews, accurate structured data and crawlable content serve both. The differences are in emphasis: AI answers reward content that covers a topic thoroughly in natural language rather than content built around a single keyword, and they put more weight on being describable as a clear, consistent business entity.
Want to know where you actually stand?
HyperWeb has been building websites and digital strategies for Newcastle and Hunter businesses since 2000, and getting our clients named in AI answers is now part of that work. If you’d like us to check where your business currently sits in AI search and what it would take to shift it, get in touch with the team. For the wider picture, our guide to getting found in AI search is a good place to start.
