AI assistants know who you are. They just can’t reliably tell a customer how to call you. On September 29, Insites published the first edition of its Local AEO Benchmark, built on 8,016 local businesses audited across ChatGPT, Gemini, Copilot, Perplexity, Grok, Google AI Overviews, and Google AI Mode. AI found 99.1% of those businesses. But when a business publishes a phone number, the assistants read it back correctly only 53.2% of the time. And 94.3% of businesses had their name, website, or phone number wrong in at least one assistant’s answer.
That second number is the one that should bother you. A wrong phone number is not a ranking problem. It is a lost job. The customer asked, the assistant answered, the customer dialed, and the call rang somewhere else: a disconnected line, an old tracking number that expired, or a competitor who bought your former number. You never see that lead. It never shows up in your call log, your CRM, or your ad reports. This post walks through where the bad number comes from, how to find it in under an hour, and the tracking tradeoff you have to settle before you fix it.
Where AI Gets the Wrong Number
The Insites data holds the most useful clue. Only 13.2% of the sources AI cited about a business were the business’s own website. The other 86.8% came from sites the owner does not run: directories, review platforms, data aggregators, old press mentions, chamber listings, and social profiles. The assistant does not trust your homepage more than it trusts Yelp. It weighs what it finds across the web, and if four sources agree on a number you stopped using in 2023, that number wins.
For most local service businesses, the conflict comes from one of four places. First, call tracking numbers that got syndicated. An agency set up a tracking line for a directory campaign, the campaign ended, and the number stayed on 30 citation sites. Second, a move or a phone carrier switch that updated Google but never reached Bing Places, Apple Business Connect, or the data aggregators. Third, a personal cell the owner used in year one that still lives on Facebook and Nextdoor. Fourth, a franchise or parent-company number that outranks the local line.
Google is adding friction of its own. Search Engine Roundtable reported on September 25 that Local Services Ads were routing calls to the Google Business Profile number instead of the Google Ads call-tracking number, even when advertisers set the Ads number in GBP advanced settings. Google told advertisers that ads with location assets “may direct calls” to the location’s number. Some accounts have since seen it reverse, but the lesson holds: Google treats your GBP primary number as the source of truth, and every other system follows it.
The 45-Minute Phone Number Audit
You do not need a paid tool to find the problem. You need a spreadsheet, seven browser tabs, and a clear head. Block 45 minutes on Thursday morning and run this in order.
- Write down your one true number. The line that rings a person or your answering service during business hours. Put it in cell A1. Every decision below compares against it.
- Ask each assistant the same branded question. Open ChatGPT, Gemini, Copilot, Perplexity, and Google (both a normal search and AI Mode). Paste the prompt below, swapping in your details.
- Ask the unbranded question too. Try “Who is the best [service] in [town]?” and note whether you appear, and if you do, which number the assistant attaches to you.
- Log every answer. One row per assistant: number given, correct or wrong, and any sources it cited.
- Follow the citations. Every wrong number traces back to a page. Click the sources the assistant lists. That page is your fix list.
Here is the branded prompt. Copy it exactly so your answers compare cleanly across tools:
What is the phone number, street address, and website for [Business Name] in [City, State]? List the sources you used for each detail.
Asking for sources matters. Perplexity and Copilot will usually show them. ChatGPT and Gemini often will when asked directly. AI answers also shift from one run to the next, so ask each assistant twice, a few hours apart. If the same wrong number shows up both times, it has a real source behind it. If it only shows up once, log it and move on. Here is what a messy result can look like for a 10-year-old plumbing shop: Google correct, Copilot pulling a tracking number from an old Angi profile, Perplexity citing a BBB listing with the owner’s former cell, and ChatGPT merging the shop with a similarly named competitor two towns over.
Fix the Sources, Not the Assistant
You cannot edit what ChatGPT says. You can edit what it reads. Work the fix list in this order, because each layer feeds the ones below it.
Google Business Profile first. Confirm the primary number is your one true number. Google’s own ad systems now lean on it, and so does Google’s AI. Then check your Posts. Google updated two Posts help documents in late September to say it may remove posts with unverified contact information, including phone numbers, emails, and social handles. If old posts mention a number that is not verified on the profile, delete or edit them. Use the Call or Learn More buttons instead of typing numbers into post text.
Bing Places and Apple Business Connect second. Copilot draws on Bing data. Siri and Apple Maps pull from Apple Business Connect. Both are free, both take about 15 minutes, and many owners have never logged in to either.
Data aggregators third. Data Axle and Foursquare feed hundreds of smaller directories. A service like BrightLocal or Whitespark citation cleanup runs roughly $3 to $5 per listing if you would rather not submit corrections by hand. Your own site last, but do it right. Put the number in plain text in the footer on every page, and add it to your LocalBusiness schema as the telephone property. Insites warns that schema alone will not get you recommended, and it is right. It is cheap hygiene that keeps your site from being the one source that disagrees.
The Tracking Tradeoff You Have to Settle
Here is the honest tension. Call tracking tells you which marketing dollars produce calls. Consistency tells AI which number is yours. Tracking numbers that leak into citations are the single most common cause of the wrong-number problem, so you cannot have it both ways by accident. You have to choose on purpose.
The setup that works for most owner-operators: keep your one true number on GBP, every directory, every citation, and in your site’s schema and footer HTML. Run tracking numbers only where the number never gets scraped into the public record. That means dynamic number insertion on your website through CallRail (plans start near $50 a month) or WhatConverts, which swaps the number in the visitor’s browser while crawlers still see the real one. It also means Google Ads call assets and unique numbers on printed mailers, truck wraps, and yard signs. Do not put a tracking number as the primary on any listing you do not control.
You will lose some attribution. You will not know for sure whether a caller found you through Yelp or through Perplexity. That is a fair price. A customer who reaches the wrong number costs you the whole job, while a customer you can’t attribute still pays the invoice. If you run LSAs, check your LSA call reports this week against your phone bill. If the counts drifted apart in September, the GBP routing change is the likely cause, and you should raise it with your Google Ads rep before October budgets lock.
One Signal to Watch in October
TechCrunch reported on September 24 that Google is testing letting Gemini call businesses on a user’s behalf, starting with US Pixel owners. Think about what that means for the 46.8% of the time AI currently reads a published phone number back wrong. Today, a wrong number costs you one customer who might notice and search again. When an agent places the call, nobody is there to notice. The agent dials, gets no answer or the wrong business, and moves to the next name on its list. That next name is your competitor.
Insites will publish its next benchmark in early November with October data. Run your audit now, fix the sources, and run it again after that report lands. If your wrong-answer count drops from four assistants to one, the work paid off. If it doesn’t move, the bad number has a source you have not found yet, and the citation list in your spreadsheet tells you where to dig.
We publish one tactical local marketing play every Wednesday, built for owners who run the site and the budget themselves. If this audit turned up something ugly, subscribe below and get the next fix in your inbox.
