AI Answer Accuracy Audit
This instrument asks ChatGPT, Gemini, Perplexity and Claude the questions a shopper asks about specific vehicles on your lot, then grades every answer against the price, fee, incentive and availability facts captured from your own vehicle page at the same moment. Every raw API response is stored and hashed. The question scripts are fixed and hashed into each run so nobody can tune them after the fact.
One dealer, up to six vehicles, four assistants, one fresh conversation per vehicle per assistant.
For each vehicle the operator records what the VDP shows right now: advertised price, dealer fee and whether it is already inside the price, incentives included, conditional offers excluded, and whether the unit is still live. The record is hashed and the VDP is submitted to the Wayback Machine.
Loading the fixed question scripts…
A deterministic rule flags a price as wrong when it misses the ground truth by more than $100 or 1%. A recorded LLM grader extracts the claims (price, fee treatment, availability, incentives) with a fixed rubric. Humans can override any grade; overrides are labeled.
Attribution rate: how often the dealer is named. Misquote rate: how often an answer carries a material error. Exposure = attribution × misquote, the share of answers that are both about you and wrong. Plus median and worst dollar miss, phantom availability count, and the visible-vs-correct quadrant.
Operator console
Runs are created by the Dealer AI Visibility team. Finished runs are public at their report link so dealers, vendors and journalists can inspect every answer.