AI citation clicks account for about 1% to 2% of calls for CallRail’s customers, a number that has roughly doubled since January.
Sean McCrohan, CallRail’s VP of Technology, shared this figure during SEJ Live‘s second session on August 26, revising a slide that put growth at 65% over six months in his own presentation.
He and Steve Wiideman, an SEJ contributor advising multi-location brands, discussed the gap between what AI assistants are doing for local businesses and what is reflected in reports. Catch the full session for free on demand.
2 Signals Are Trackable, And Both Are Small
McCrohan shared that CallRail has observed a common pattern: people often start their research with an assistant before moving on to a branded search. Interestingly, what you can measure looks just like a typical search, even though the assistant did the initial work.
He mentioned that tracking AI activity all the way to a call is very hard right now. Intermediary tools from ad networks like Google aren’t yet available from AI vendors. He also noted that the tools for ChatGPT ads are less developed compared to the well-established search ad systems that marketers rely on, and he mentioned that OpenAI has been reluctant to discuss this more.
There are two main things you can actually measure. One is when someone clicks on a citation within an AI response, which then directs them to your website, acting like a typical referral. The other is self-reported attribution, where a customer says that ChatGPT recommended the business. Call recordings also show instances without trackable clicks. Since January, both methods have seen a rise of over 100%, starting from a small beginning.
Filtering GA4 for activity from ChatGPT, Gemini, Claude, Perplexity, and Grok shows close to 1% for the multi-location and franchise brands Wiideman works with. He views this as quite normal for a discovery channel, rather than a direct sales platform.
While reading a live poll aloud, Wiideman mentioned that about a quarter of the audience said they could tell when an assistant sends a lead, and nearly half weren’t sure. McCrohan said he’d expected fewer yes responses.
AI Traffic Arrives Later Than Ordinary Web Traffic
CallRail has historically seen that about half of website traffic comes outside regular business hours that number rises to nearly two-thirds. It’s not uncommon for people to start browsing with an assistant in the early afternoon and continue browsing well past midnight
McCrohan noted that responses might highlight the top three businesses instead of just one, which means a missed call can send the lead somewhere else.
He mentioned, “You don’t have to use mine,” referring to CallRail’s voice agent. “Get something, because more and more of this business is a 24 hour business.”
Wiideman shared a story about clicking on a plumber’s ad, making a call, and hearing it ring, reminding us that the cost of that click is already paid, no matter what the outcome.
Watch the full session on demand.
Server-Side Number Swaps Work, In-Browser Swaps Don’t
McCrohan points out that giving a separate phone number for an AI crawler instead of Googlebot does work. However, this change needs to be made on the server side, which can be a bit technical.
He explains, “The crawlers that AI agents are using to retrieve this data do not execute scripts on your page when they go to look at your website. They just don’t.”
He said CallRail put a lot of effort into trying to get these crawlers to run page scripts, but it didn’t work out. The in-browser number swapping isn’t noticeable to the crawlers he mentioned. He believes that server-side handling will continue to be important, especially as AI attribution becomes more common and state privacy rules become stricter.
Wiideman raised cloaking as a reason to be careful with user-agent tests and stresses the importance of having consistent canonical local numbers everywhere Google can see because the name, address, and phone number need to match in Maps. He suggests doing tests on low-traffic pages to stay safe.
McCrohan mentioned that a site can still be valuable even if customers never visit, since that’s where the assistant got its answer.
Prompt Libraries Only Work As A Trend Line
Wiideman recommends starting with semantic triples, which are the specific claims a business wants to be recognized for. He suggests putting one in quotes, searching for it, and expecting to find few other sites stating it directly. These claims form a prompt library of about 100 to 125 entries that can be used for future reference.
He used the term ‘prompt drift’ to describe how the same question can sometimes lead to different answers, whether for different people or even for the same person asking twice. Steady Demand’s analysis shares some interesting numbers: about 40% of the time, repeating the same local search in Gemini results in overlapping citedo roughly 90% for Google’s local pack
McCrohan highlighted how important attention is, noting that tracking more prompts than the team can manage means they end up only following up on a random subset.
Wiideman also suggested that attendees shouldn’t focus too much on any single citatione covered this decline in August, and the reasons for it still aren’t fully explained. He recommended focusing on the type of site rather than just the specific one
Your Customers’ Words Are Already On Your Call Recordings
Call transcripts capture the actual words customers use to describe their problems, which might be different from the words on your website, McCrohan explained. When there’s a disconnect between the two, your business might miss chances to show up at all.
Wiideman also points out how important site chat logs are. Usually, visitors begin with a prompt in the chat widget, but many businesses overlook these conversations. If your team doesn’t have call recordings, he recommends using chat tools that can highlight conversations with specific keywords—this can be a great way to get started and better understand customer needs.
Ratings And Reviews Still Feed The Recommendation
Wiideman encourages businesses with multiple locations to train each location to ask customers for reviews on the other platforms they already use, not just Google Business Profile, like Reddit, TripAdvisor, and Yelp.
He notes that Yelp reviews show up on Bing and Apple Maps and that Yelp has a partnership with ChatGPT. He suggests aiming for an average review score of about 4.5 to 4.7 across various industries. He referenced Darren Shaw, who spoke later at the event, while talking about how the chances of getting recommendations drop off below that range.
Governance Beats Ranking Signals At Scale
No matter if a brand has 40 locations or 4,000, Wiideman emphasizes how important it is to have centralized control. The corporate team takes care of the schema, data feeds, and tracked phone numbers, which helps prevent issues like rogue listings and unauthorized manager cell numbers from showing up online. When managing so many locations, scrubbing calls and voicemails across all of them turns into a content roadmap.
He also suggests that for multi-location searches, keeping business data consistent is now even more crucial than traditional ranking factors.
Looking Ahead
Wiideman expects agents to verify a contact number and offer to make the call, all while the customer never needs to visit the site. Both speakers landed on the same note.
“It is an exciting new world, but it is not a 100% new world,” McCrohan said. “There is carryover. It will be OK.”
Watch the full session for free.
Join Us For Our Next Webinar!
AI Cites Your Brand. Now What? Turn AI Visibility Data Into Actions
Join Constance Tan of Ahrefs to learn which AI visibility metrics matter, how to prioritize your next steps, and which tactics win more citations in AI answers.
Reserve my Seat
CategorySEOAI SearchLocal Search
