Ask Gemini the same local question twice and it recommends the same business 7.9% of the time. Google’s own local pack, tested on the same 500 metro-and-category combinations at the same cadence, returned the same top listing 90.2% of the time. That gap comes from Steady Demand’s AI Citation Ledger, published August 19, which logged 14,472 citations across 1,487 local queries in 50 US metros. The company calls the effect Grounding Drift. Repeat a query word for word and only about 40% of the cited sources come back.
Most operators are still checking AI visibility the way they checked rankings in 2019. Type the query, look at the answer, feel good or feel sick, close the tab. That habit is now worthless. A single check tells you almost nothing, and it tells you something different every hour. The fix is not a better tool. The fix is sampling.
One check is a coin flip. Ten checks are a number.
Rank tracking worked because the underlying result was stable. You were at position 4 on Tuesday and position 4 on Thursday, so a drop to position 9 meant something happened. AI answers do not behave that way. The same prompt, from the same city, minutes apart, pulls different grounding sources and lands on different businesses. Treating one result as your position is like judging your monthly revenue off one Tuesday afternoon.
So stop asking “am I in the answer” and start asking “how often am I in the answer.” That is a rate, and a rate needs repeated trials. Ten runs of the same prompt give you a usable number. Thirty runs give you a number you can defend to yourself three weeks from now when you want to know if the work paid off. The math is not hard. The discipline of actually running it on a schedule is the hard part, and it is also the part almost none of your competitors will do.
Sampling also protects you from panic. An owner who checks once and sees a competitor named will spend a weekend rewriting a homepage that was fine. An owner with a log knows that competitor shows up in 4 of 10 runs, they show up in 6 of 10, and the real problem is a service page that never gets cited at all.
Build the 20-prompt log this week
Here is the whole build. It takes about an hour to set up and 20 minutes a week to run.
Write 20 prompts a real customer would type. Not keywords. Sentences. Five of them should be discovery prompts (“best emergency plumber in Tempe AZ open now”), five should be comparison prompts (“who is better for tankless water heater install, X or Y, in Tempe”), five should be qualifying prompts (“plumber in Tempe that finances a water heater replacement”), and five should be the specific objections your phone actually gets (“plumber in Tempe who works on 1970s cast iron”). Those last five matter most. They are where you win, because they are too specific for a directory to answer well.
Open a spreadsheet with six columns: prompt, date, engine, recommended (yes or no), position in the list, and cited domains. Run each prompt three times per engine in a fresh session, spaced across the day, and log every run. Twenty prompts times three runs times two engines is 120 rows. You do it on a Wednesday morning with coffee. Score it at the end: your recommendation rate is the count of yes divided by total runs, per engine.
Two rules keep the data honest. Use a fresh incognito session or a signed-out window every time, because a logged-in account with your own business history will flatter you. And do not change your prompts between weeks. The whole value is comparing week 6 against week 1 on identical inputs. Add prompts to the bottom of the list, never edit the ones already there.
Then set a floor. If your recommendation rate on the five objection prompts sits under 30%, that is your work queue for the month. If your discovery rate is fine and your qualifying rate is near zero, you have a content problem, not a profile problem.
Search Console just handed you the only first-party number
On August 31 Google finished rolling out its AI performance reports in Search Console to every account worldwide. The documentation now carries a line noting the control went global that day. The report breaks out impressions from AI Overviews and AI Mode with pages, countries, devices, and dates. It is first-party data from Google itself, and it costs nothing.
Be clear about the tradeoff before you build a monthly report around it. The AI report gives you impressions and no click data. So it tells you which pages Google pulls into generative answers, and it does not tell you whether anyone came to you afterward. That is a real limit, and it lines up with what Search Engine Land reported the same week: calls and clicks keep sliding as Maps becomes the destination rather than the doorway. Impressions up and calls flat is now a normal quarter, not a broken one.
Use the report for one job, and use your prompt log for the other. Search Console tells you which of your pages Google is willing to ground an answer on. Your log tells you how often that grounding turns into a recommendation for your business by name. Those are different questions and only the second one pays your bills.
Google also shipped a toggle in the same release that blocks your site from appearing in generative AI features. Google says sites that opt out get no traffic and no impressions from those features, and that the setting is not a ranking signal for regular search. For a local service business the answer here is easy. Leave it on. Opting out of the surface where your customers now ask their questions is not a strategy, it is a retreat. The toggle exists for publishers with a licensing fight to pick. That is not you.
Fix the source the models actually reach for
The same Steady Demand data explains where the work goes. Nearly 60% of Gemini’s local citations pointed at businesses’ own websites, more than directories, review platforms, and forums combined. Reddit came second at 13.7%, beating the entire local-service directory category including Angi, Thumbtack, and HomeAdvisor. Your site is the primary source. That is good news, and it means the old instinct to spend the marketing hour on a directory profile is now backwards.
So write the pages the models need. One page per service, with the specifics a model can quote: price ranges, service area by name, what you will and will not work on, hours, response time, financing, warranty terms. Put the objection answers from your prompt log directly on the page in the customer’s own words. If “do you work on 1970s cast iron” comes up in your log and no page on your site says yes, no amount of profile tuning fixes it.
One more finding should change how you read any AI visibility tool you pay for. Running those same 1,487 queries through ChatGPT, the two engines cited the same domains only 8% of the time and recommended the same top business 4.2% of the time. Gemini leaned on business websites. ChatGPT leaned on Reddit and directories. Winning one tells you nothing about the other, so track both columns separately and expect the fixes to differ. A tool that gives you one blended AI visibility score is averaging away the only thing you needed to know.
The signal to watch over the next 30 days is whether AI report impressions in Search Console start moving independently of your Maps calls. If impressions climb while calls stay flat, the answer is capturing the customer and you need a reason for them to leave the answer and reach you. That reason is usually a price, an availability window, or a promise nobody else on the page is making. Start writing it down this month.
Atlas Unchained builds the sites and measurement setups that put local operators in the answer instead of next to it. If you want the prompt log template and the service page outline we use, subscribe and reply to the first email.
