ModelDrift

When someone asks an AI assistant what software to use, the answer names a handful of companies and describes them, and that description is either accurate or it isn't, whether the company shows up at all or it doesn't. ModelDrift measures that, systematically, for one category, over time.

Starting in August 2026, ModelDrift asks ChatGPT, Claude, Gemini, Perplexity, and Google AI Overviews the same 21 buyer questions about electrical contractor software, each question five times, in a fresh session, per engine. It records which companies get named (presence), what's said about them and whether it's correct (accuracy), and which web pages the answers cite (sourcing). The identical question set runs again in October and December 2026, so changes between runs can be attributed to the models, not to the questions.

Read the full methodology for exactly how questions are asked, how repetitions are counted, and why the question set is frozen for the life of the project.

First results: September 2026

The first run happens August 12, 2026. Nothing on this site yet reflects real findings. There's no data to show. The first essay, covering presence across all five engines for the August run, publishes in September. See the essays page for what's coming and when.

The tracked companies and the full raw dataset (published as CSV and JSON after each run) are listed now so the project is checkable from day one. See the data page.