Three conditions, without which the test says nothing
Condition 1. A clean session
Your account knows your earlier conversations and your location. If you have ever discussed your business with the assistant, it knows about it and will name it more readily. Log out or open a private window. That single change usually flips the result.
Condition 2. No business name in the question
Asking "is Clock Tower Bakery any good" always produces an answer containing Clock Tower Bakery. That is not a test, it is an echo. Ask the way a customer asks when they do not yet know where they are going.
Condition 3. Repeats
The model builds its answer fresh every time. The same question two minutes later can return a different set of names. Without repeats you are measuring chance.
How to phrase the questions
A good test question contains a place and a need, and does not contain your business name. Examples across trades:
| Trade | Test question |
|---|---|
| Bakery | where can I get good sourdough near Bedford Ave |
| Dentist | recommended dentist in Park Slope taking new patients |
| Garage | trusted mechanic in Greenpoint for a brake job |
| Barber | good barber in Bushwick, walk-ins |
| Hotel | where to stay in Williamsburg with a dog |
Notice that each one carries an extra condition: sourdough, taking new patients, walk-ins, with a dog. That is how people ask, and it is worth testing that way, because those conditions are exactly what decides whether the model picks you.
Test four engines, not one
ChatGPT, Gemini, Perplexity and Claude answer the same question differently, because they run on different search layers underneath. A business invisible in ChatGPT is sometimes named by Perplexity and the other way round.
Perplexity is the easiest to learn on, because it shows a citation beside every sentence. You can see immediately which page a competitor’s name came from, and work out why yours did not.
What to record
A bare "named" or "not named" is useless a month later. Record:
- the full text of the answer, not just whether your name appeared
- the date and time, because answers drift
- the engine and the run number
- the names that came up instead of yours, because that is your real competition in the eyes of a model
That last one is often the most valuable. It usually turns out the model names entirely different businesses from the ones you consider competitors.
When to hand it over
Twenty questions times three runs times four engines is 240 answers to read and count. By hand that is several hours per measurement, and the measurement has to be repeated to show change.
Our free scan runs the short version: three questions to ChatGPT, full answers with timestamps, result straight away. The four-engine measurement with twenty questions is part of the paid work and is described on the Method page.