How we find what ChatGPT says about you, and fix it.
Seven steps. Each one leaves a dated record. That record is what separates this from an opinion, and it is what you should ask every other vendor to show you.
The seven steps.
Signed out, every time.
We ask the buyer's question on ChatGPT, and on Google AI Mode for client readings, with no login, one fresh chat per question. No account history shapes the answer, so what we record is what a stranger types in and gets back.
Questions with no brand in them.
A buyer who has not picked a store yet does not type your name. So the main battery is open questions, with no store named. A second set names the brand. The open set is where a store is found or skipped.
Every answer saved, with its date.
Verbatim, stamped with the day it was asked. Each question is asked twice, a day apart, because these engines drift: in our September census, 1 answer in 6 flipped a store in or out overnight. One reading is an anecdote. Two dated readings are a measurement.
A control store.
In the published study, a store we are not working is read on the same questions, on the same days. If every store moves, the engine moved. If only ours moves after a fix, we say so, with both dates.
Find the records the answer leans on.
Each answer names or quotes outside records: review sites, directories, data aggregators. In the September census, ChatGPT quoted CarEdge in 146 of 168 open answers and DealerRater in 101. That is where the answer gets decided, so that is where we work.
Fix them, and log every fix.
We correct the outside records and write down each fix the day it goes live. In the study, the store approves every fix first. The log is the store's to keep.
Read it again, same questions.
The same battery, the same conditions, on a schedule. The before and after carry a date on both sides, so anyone can line them up.
Ask any vendor for three things.
A dated before and after, which I can show you from Huntsville and Temecula. A control store and a log of what was fixed and when, which my published study puts in public. If a vendor cannot show a dated before and after, you have your answer. My published study is built to show all three, with the store's name on it, and I publish it whatever it shows.
The method, on the record.
ChatGPT, signed out, 48 franchised stores, 502 answers. 84 of 168 open answers named the store. 18 of the 48 came up in none. See the record.
Nine wrong directory listings, one email. Eight gone within a week, still gone two months later. See the record.
A transmission shop outside Google's top three map results on two checks, then #2 and #3 on September 28. See the record.
We do not sell a ranking. Google re-sorts the map every day and the engines keep re-reading their sources. We sell the work and a record you can check, which is why the readings keep running after the fix.