TL;DR
FACT: AI answers change from run to run, so a single check tells you almost nothing. Tracking means a fixed set of questions, a fixed schedule, verbatim saved answers, and a score you can compare across runs.
FACT: We run this loop on BackTier itself. On 4 October, two search-enabled models answered 16 of our field-study questions and cited 58 links — none to backtier.com. On 9 October, a stand-in run produced 12 answers and 35 cited links — again none to us. Zero is a measurement. It is where tracking starts, not where it ends.
Step one: fix the questions
Pick the questions a buyer would actually ask an assistant: the problem you solve, the category you are in, the comparison a prospect makes before calling you. Ten to thirty questions is enough. Write them down and do not edit them between runs, or you are measuring your editing, not your visibility.
Our own set includes the market questions our field studies cover — a solo counselor's local visibility, AI in deathcare records, a Central Florida insurance agency — plus direct questions like "What is a bounded AI stack?", a term we coined.
Step two: run on a schedule and save everything
Run the same questions against the same models on a schedule you can keep — monthly is a reasonable start. Save every answer word for word, with the date, the model, and every source link it cited. Screenshots or a dated evidence file both work; memory does not.
FACT: Every run we publish on our Experiments page is backed by a dated file with the verbatim answers. That is what makes the results inspectable by anyone, including you.
Step three: score four things
For each answer, score: Were you selected — named at all? Were you cited — linked as a source? Where did you appear — first recommendation or footnote? And was the entity accurate — did the answer describe the right business?
INFERENCE: of the four, entity accuracy is the one most brands skip, and it is the one that corrupts the rest. An answer that names you but describes a different company is worse than no answer, because it teaches the model the wrong fact.
Step four: change one thing, then re-run
A tracking loop only earns its keep when it drives a change: a definition page for a term you own, a corrected fact on a page the models read, a citation earned on a site the models already trust. Change one thing, wait, re-run the same questions, and compare.
INFERENCE: expect weeks, not days. Search engines and AI tools revisit pages on their own schedule, and a single re-run is still a small sample. The trend across runs is the signal.
Canonical definitions used in this post
AI visibility and related terms: [What is BackTier](/what-is-backtier) and [GEO, AEO, AIO, HEO and SEO, defined together](/answers/geo-aeo-aio-heo-seo). Bounded AI stack: [the canonical definition](/bounded-ai-stack). Other terms: [the glossary](/glossary).