The Local Signals AI Reads, and the Order to Fix Them

You rewrote the homepage for AI. You chased a dozen fresh reviews. Meanwhile the machine deciding whether to recommend you is reading four versions of your phone number, three of your address and two spellings of your business name. It hands the answer to the competitor whose facts agree with each other. For a local business, the signal that most often breaks AI recommendations is NAP disagreement: your name, address and phone listed differently across platforms you never log into. It belongs at the front of the repair list, ahead of reviews and far ahead of content.

Read the pages currently ranking for this question and you will get a list of local signals AI reads. None of them tell you which one is most commonly broken in a real business, or what to fix first when you have one afternoon and no budget. I can answer both, because I have been doing search work since 2008 and have fixed local visibility for over 300 businesses across the USA, Canada, UK, Singapore, Australia and New Zealand. The answer is almost never the thing owners are worried about.

Your address breaks more AI answers than your copy ever will

The audit goes the same way most weeks. The Google Business Profile is immaculate, because that is the dashboard the owner or the last agency actually opened. Then I check Apple Business Connect and find the old suite number from the move two years ago. Bing Places has a phone number that belonged to a receptionist who left. A data aggregator, one of the wholesale data suppliers that feeds address records to dozens of apps and directories, still lists the previous tenant's business name at that address.

Four sources. Three different stories about a single small business. A human customer shrugs and calls the number on your website. A language model does not shrug. It cross-references, finds a conflict it cannot resolve, and defaults to the option it can state with confidence, which is the competitor down the road whose four sources say the same thing four times.

You can be fully verified on Google and still be invisible inside an AI answer, because the machine is not asking Google for permission. It is reading a spread of sources, and your Google verification never travels to any of them.

The 29-point gap deserves a minute of your attention. For most local queries the summary is the storefront now, and the map pack you have been optimizing for a decade is the second thing on screen, when it appears at all. That changes what a wrong phone number costs. It used to cost you a handful of misdials. Now it can cost you the mention entirely, because the summary never names a business it cannot describe consistently.

Whitespark's split by search type tells a different story for a bare "plumbers in Phoenix" search, where the local pack showed 93 percent of the time. I work through the split in how to show up in Google AI Overviews.

Address changes are the sharpest version of inconsistent listing data. In the moves I have handled, the Google record updates inside a week and the rest of the ecosystem lags for months. That lag is behind the frustrating pattern a relocation so often produces, which I wrote about in why a new Google address does not move your rankings. The old data is still out there being read by something.

Owners fix the signal they can see and skip the one they cannot

Reviews feel like progress. You can watch the count go up, you can forward a nice one to your team, and asking for them feels like real work. Listing data feels like admin from 2014. The review campaign gets the attention and the listings rot. Eighteen months later nobody can explain why the phone is quieter than the rankings report suggests it should be.

Picture your own business for a moment. You run a physio clinic. You moved two doors down into a bigger space last spring, you are verified on Google, your star average is strong, your reviews are healthy, and your website is fine. A prospective patient with knee pain asks their phone for a physio near them. Apple Maps still holds your old unit number. The directions take them to a locksmith. Your Bing listing shows the pre-move phone. The AI summary, reading across those sources, mentions two clinics, and yours is not one of them. Your dashboard shows every metric healthy, because every metric that dashboard measures is healthy.

There is a second reason owners aim at the wrong target.

Rewriting your homepage is a bid for the roughly one-in-six slot where your own site gets cited. Worth doing, and I would still do it. It is also the smallest lever on the board, and every agency sells it first because it is the one they can bill for. Everything else the machine reads about you lives somewhere you do not own. The wider cross-channel version of this work, covering your site, your off-site footprint and your tracking together, is in my guide to improving brand visibility in AI answer engines. This post is the local half of it, done in order.

One answer can rest on a dozen sources, and not one of them is you

AthenaHQ's tracking across seven models put the average at about 12 cited domains behind a single AI response, with wide variation by engine: ChatGPT averaged 18.86, Grok 26.99, and Copilot just 5.77. Those figures cover all industries, not local queries alone. Uberall's quick-service restaurant benchmark, powered by the same AthenaHQ data, ran the same question on local queries and got a different spread by engine.

Search Engine Journal flagged these datasets as directional and overlapping. They are not independent studies confirming each other. Treat the numbers as the shape of the behavior, not a specification you can engineer against.

The first takeaway is that a Gemini answer built on four sources is brutal arithmetic for you. One conflicting record out of four sources is a quarter of the evidence pointing the wrong way. The second is the star column, which is where owners misread the data hardest. Those are averages of businesses that got recommended, not a bar you must clear. A 3.9 business was still being recommended by Gemini. Nobody was disqualified at 4.0.

The more useful question is what those reviews actually said. That is the part with practical value, because Uberall's work identified which review attributes AI systems found most useful when deciding what to recommend.

Change what you ask for and you change what the machine can read. "Mind leaving us a review" produces "great service, highly recommend," which is useless to a system trying to decide who treats runners' knees in Didsbury. "Would you mind mentioning what we treated and roughly when you came in" produces a sentence a model can actually use. Same effort, completely different asset. The recency finding also matches a pattern I have argued before: a steady trickle of recent reviews beats a high lifetime average.

The plumbing is shifting under this too. In July 2026, OpenAI signed a deal with Yelp that pipes live Yelp reviews, ratings and photos straight into ChatGPT with branded links back. Danny Goodwin reported it at Search Engine Land on July 23. Your Yelp page stopped being a directory listing you ignore and became a live feed into the answer.

Fix them in this order, not the order that feels urgent

Matt G. Southern's reporting for Search Engine Journal arrives at a strict sequence: listings and NAP consistency first, reviews second, third-party reputation third. My field experience says the same thing, and the reason is mechanical. Reviews and reputation are evidence about a business the machine has to identify first. If it cannot decide which business you are, better evidence about you has no business to attach to.

Step two is the whole ballgame if you only get one afternoon this month. Open Apple Business Connect and Bing Places yourself. Both are free, both take about twenty minutes each, and both hold stale data on more small businesses than I can count. You do not need an agency for this and you should not pay one a retainer to do it. Done means the name, address, phone, hours and primary category on both platforms match your master file character for character.

Steps three and five are where delegation makes sense, because tracking down a duplicate listing at an aggregator is tedious and slow. If you want to attempt it yourself first, search your business name plus your old address in quotes, then search the old phone number on its own. Work through the first two pages of results and request a correction or removal on each platform that surfaces. If you hand it over, ask something specific. "Which platforms did you check besides Google, and can you send me a screenshot of the name, address and phone on each one as it stands today?" A partner who has done the work answers with screenshots inside a day. Vague reassurance about citation building means it has not been done.

On step five, apply Southern's filter without sentiment. A roundup with a named author who explains how they picked the businesses is worth pursuing. A "top 10 plumbers" page with no byline, no criteria and a payment link is a spam signal wearing a nice outfit, and you should not pay to be on it.

Measure whether the machine changed its mind, not whether your dashboard did

Rank tracking will not show you this. You can hold position three in the map pack all quarter while every AI summary in your category names three competitors and skips you. I unpacked that disconnect in how to check if AI is actually recommending your business. The measurement has to happen inside the answer engines themselves, and it is manual work for now.

Owners watch the star average obsessively, and the Uberall data suggests it is a weaker lever than the words inside the reviews. A citation count sold to you as a win is just as weak a signal. Being mentioned in an answer and being the business the answer tells someone to call are different outcomes. Understand that before you buy a report about it.

Give it about eight to twelve weeks after the listing fixes before you judge anything. Records propagate slowly through the aggregators, and models refresh what they know on their own schedule. Fix, log, wait, re-ask.

Frequently Asked Questions

The London ADHD clinic I booked solid for three straight months, to the point where they hired more specialists and outsourced the overflow, did not win on clever copy. It won on fundamentals done in the right order, the same principle this post has been arguing. That is a boring competitive advantage and almost nobody has it.

Which raises a question worth answering before your next invoice goes out. You have spent years treating your Google listing as the source of truth about your own business. It never was. It was only ever the one source you could see, and the four you cannot see are the ones now doing the talking. If you want a second pair of eyes on which of them disagree about you, book a short AI visibility call and we will go through them together: https://cal.com/johntalaguit/ai-visibility-call.