Twenty Prompts to Test Your Cafe in ChatGPT and Gemini

Thomas Garner

Oct 7

11 min read

You want to know what ChatGPT and Gemini actually say about your cafe. A random question typed in on a slow afternoon tells you something, but it does not tell you much, because you have nothing to compare it against next month.

Short answer: run the same twenty prompts, the same way, every month. This post gives you that list, grouped so each set of four tests one thing, discovery, comparison, facts, needs, and problems, along with a simple score sheet and a short guide to reading the result. A fixed list beats random questions because it lets you compare month to month instead of judging each answer in isolation.

This is the prompt list and the method. Logging your results in a spreadsheet is its own separate post, and correcting a wrong answer once you find one is covered elsewhere too. Here, the job is just running the test and understanding what the score does and does not tell you.

The Twenty Prompts, in Five Groups

Discovery, four prompts, testing whether your cafe comes up at all without being named first. Is there a good independent coffee shop in {city}? What is the best coffee shop near {neighborhood}? Where should I get coffee in {city} if I want something local? What is a hidden gem coffee shop in {city}?

Comparison, four prompts, testing how your cafe fares against others when directly weighed. How does {cafe} compare to other coffee shops in {city}? What makes {cafe} different from a chain coffee shop? Is {cafe} worth visiting compared to other options in {neighborhood}? Which coffee shop in {city} has the best reputation?

Facts, four prompts, testing whether basic details come back correct. What are {cafe}'s hours today? Does {cafe} have parking? What does {cafe} serve besides coffee? Where exactly is {cafe} located in {city}?

Needs, four prompts, testing whether specific customer requirements are answered correctly. Is {cafe} good for working on a laptop? Does {cafe} have dairy free milk options? Is {cafe} dog friendly? Does {cafe} have outdoor seating?

Problems, four prompts, testing how an assistant handles a less flattering angle. Are there any complaints about {cafe}? Is {cafe} usually busy or hard to get a seat at? Has {cafe} had any recent changes, like new ownership or a menu change? Is {cafe} still open?

The Method: How to Run This Consistently

Run all twenty on the same day, in the same sitting if possible, across whichever assistants you have access to, most commonly ChatGPT and Gemini, and Google AI Mode where it is available to you. For each prompt, record the tool, the exact wording used with your cafe, city, and neighborhood filled in, whether your cafe was named at all, whether the facts given were correct, missing, or wrong, and any source the assistant cited.

Log in state matters and should be recorded, not ignored. Whether you are signed into an account, and where you appear to be located, can change what an assistant returns, since some systems use account history or detected location to shape an answer. Note both for every run, so a month to month comparison is not accidentally comparing a signed in result to a signed out one without knowing it.

Answers vary between runs of the exact same prompt on the same assistant on different days, sometimes even within the same day. A single result is a data point, not a verdict. This is why the list is fixed and repeated monthly rather than run once and treated as final; one month's answer tells you what happened that day, and a pattern across several months tells you something closer to the truth.

Scoring the Twenty Prompts

Use a simple three point scale per prompt: two points if your cafe is named and the information given is correct, one point if your cafe is named but something about the answer is wrong or outdated, zero points if your cafe is not named at all. Total each group of four separately, then sum all five groups for a score out of forty.

This score is a self assessment for your own cafe over time, not a benchmark against other businesses and not a version of any published visibility score. Do not compare it to a formal audit score out of forty from a different scoring method; the numbers can look similar but they are measuring entirely different things, built with entirely different methods, and treating them as comparable would be misleading.

Reading the score by group matters more than reading the single total. A cafe that scores well on facts and needs but zero on discovery has a different problem than a cafe that scores well on discovery but poorly on facts. The first cafe is known but under described once found; the second is found but with wrong details once it comes up. Those call for different fixes.

Reading a Low Discovery Score

If your cafe scores zero or low across all four discovery prompts, meaning it is not named when someone asks a general question about good coffee in your area, that is worth taking seriously but not panicking over. A discovery prompt is the hardest of the five groups for a smaller, less established cafe to score well on, since it is competing against every other option an assistant could plausibly name for that same general question.

A low discovery score paired with a decent comparison, facts, and needs score is a reasonable place to be. It suggests that once your cafe is somehow in the conversation, the assistant has decent information about it; the gap is in being the first thing mentioned, which is a harder, slower problem than a factual correction.

Reading a Low Facts or Needs Score

A low score on facts or needs prompts, where your cafe is named but the details are wrong, missing, or outdated, is usually the most fixable category of the five. Wrong hours, an outdated address, or a missing detail about dairy free options or outdoor seating typically traces back to a source the assistant is pulling from, most often your Google Business Profile or your own website, being incomplete or stale.

This is the group most likely to point you toward a concrete, immediate fix: check your Google profile, check your website, and check whether the specific detail that came back wrong is correct and current in both places. Most fixes that matter most come from correcting wrong or missing facts at the source, not from anything about how the prompts themselves are worded.

Reading the Problems Group Carefully

The problems group is the one to read with the most caution, because an assistant's answer here can reflect real information, outdated information, or simply an absence of any information at all, and those three look similar on the surface. If an assistant says it has no information about recent complaints, that is not the same as a clean bill of health; it may simply mean nothing specific has been indexed either way.

Do not treat a vague or hedged answer in this group as either good or bad news on its own. Note what was said, compare it against what you know to be true, and flag any specific factual claim, such as a wrong claim about being closed or under new ownership, for a direct correction rather than a general worry.

What Fixes Tend to Come From Where

In practice, the three fixes that matter most after a full run of twenty prompts tend to come from wrong or missing facts, not from anything about how the prompts are phrased. A cafe finding that its hours are wrong in three different assistant answers has an actual, fixable problem to solve at the source, most likely a profile or website update, rather than a mystery about how AI search works in general.

Resist the temptation to rewrite the prompts themselves looking for a version that flatters your cafe more. The value of this list is that it stays fixed, the same twenty questions, run the same way, month after month, so your own comparison over time stays honest. A prompt list tuned to produce better answers is not measuring anything real anymore.

Running It Once a Month

Pick a recurring day, the first Monday of the month works well, and treat it like any other repeating task on your calendar. Set aside roughly forty minutes: about two minutes per prompt across twenty prompts, plus a few minutes to log the scores and note anything that changed from last month.

Keep every month's results in the same simple log, even a basic spreadsheet, so you can see the trend rather than just the latest snapshot. A single month's low discovery score is a data point. Three months of a flat, unchanged discovery score is closer to a pattern worth acting on. The difference between those two readings is the entire reason to run this monthly rather than once.

Why the Placeholders Matter

The {cafe}, {city}, and {neighborhood} placeholders in the prompt list above are not decoration. Fill them in exactly the way a real customer would type or say them, not the formal legal name on your business license if that differs from what people call your cafe, and not a neighborhood name so obscure that a visitor would not know to use it. The goal is to mirror a realistic question, not a technically precise one.

If your cafe has a common nickname customers use, it is worth running one extra pass of the discovery and comparison prompts using that nickname instead of the full name, and logging it as a separate row rather than folding it into the same score. An assistant may respond differently to a nickname than to a full formal name, and that difference itself is useful information about how your cafe is being talked about.

Keeping the Test Honest Over Time

Resist the urge to skip a month because you already have a rough sense of how things are going. The value of a fixed monthly test comes from the months you run it even when nothing seems likely to have changed, since those unremarkable months are what let you tell the difference between a real shift and normal month to month noise later on.

It is also worth resisting the opposite temptation, running the test more than once in the same month because an answer looked disappointing and you want to try again hoping for a better one. Log the disappointing answer as this month's real result. If next month looks better, that is useful information. If you quietly rerun until you get an answer you like better and only log that one, the whole log stops being an honest record and starts being a collection of your best days.

This is also a reasonable test to hand off, once you have run it yourself a few times and trust the process. A manager or a staff member who is careful and consistent can run the same twenty prompts on the same schedule just as well as an owner can, as long as the method described above, same day, same wording, logged login state and location, stays the same no matter who is at the keyboard that month.

What This Test Cannot Tell You

This test tells you what an assistant said, on a given day, in response to a given prompt. It does not tell you why the assistant said it, and it does not tell you how many real customers ever asked anything similar or ever saw a similar answer. Treat it as a window into one specific, repeatable check, not as a measure of your cafe's overall reputation or actual customer volume.

It also will not, on its own, tell you whether a wrong answer today will still be wrong next month. Some corrections take effect quickly once a source updates; others take longer to be reflected in an assistant's answer, for reasons outside your control. That lag is a real limit of this test, and it is another reason a single month's result should be read as a data point rather than a final verdict.

Frequently Asked Questions

How do I test my cafe in ChatGPT?

Run the twenty prompt list in this post, grouped into discovery, comparison, facts, needs, and problems, using your cafe name, city, and neighborhood filled into each placeholder. Log whether your cafe is named, whether the facts given are correct, and any source cited. Run the same list monthly, on the same day, so results are comparable over time rather than a one off snapshot.

Should I run these prompts on ChatGPT, Gemini, or Google AI Mode?

Use whichever tools you can keep checking every month. ChatGPT and Gemini are the two this post is built around, and Google AI Mode is a sensible third where it is available to you. For every prompt, log which tool answered, the exact wording with your cafe, city, and neighborhood filled in, whether you were signed in, and where you appeared to be located, since results differ between systems and between login states.

What does the score out of forty mean?

It is a self assessment for your own cafe to compare month to month, not a national benchmark and not the same scale as any formal visibility audit score. Two points per prompt for named and correct, one for named with an error, zero for not named, summed across all twenty prompts. Read it by group, not just as one total.

Why does my cafe score low on discovery prompts but fine on facts prompts?

That pattern suggests your cafe is not usually the first thing named for a general question about good coffee nearby, but once it is somehow part of the conversation, the assistant has decent, accurate information about it. That is a different, generally slower problem to address than a factual error, and it calls for a different fix.

What should I fix first after running this test?

Start with anything in the facts or needs groups that came back wrong, missing, or outdated, since those usually trace directly to your Google Business Profile or your own website being incomplete or stale, and are the most fixable of the five groups. Discovery gaps take longer to move and are a lower priority for an immediate fix.

Does a single low score mean something is broken?

Not necessarily. Answers vary between runs of the same prompt on the same assistant, sometimes within the same day, so one result is a data point rather than a verdict. Read the score by group, since each set of four tests something different out of forty total points. A low discovery score is common for smaller cafes, and a pattern across two or three consecutive monthly runs is the more reliable signal.

Does being signed into an account change the answers?

It can. Whether you are signed in and where you appear to be located may shape what an assistant returns, since some systems use account history or detected location. Record your login state and location for every run so a month to month comparison is not unknowingly comparing a signed in result to a signed out one.

How should I read the problems group of prompts?

Read it carefully. An assistant saying it has no information about complaints is not the same as a clean bill of health; it may simply mean nothing has been indexed either way. Note the exact answer, compare it against what you know, and flag any specific wrong factual claim, such as being incorrectly reported closed, for a direct correction.

How long does running all twenty prompts take?

About forty minutes. That is roughly two minutes per prompt across all twenty, plus a few minutes to score each answer on the two, one, or zero point scale, total the five groups out of forty, and compare against last month. Pick a fixed recurring day, such as the first Monday of the month, and keep every month in the same simple spreadsheet so the trend is visible.

Can I change the prompts to get better answers about my cafe?

No, that defeats the purpose. The value of the list comes from running the same twenty prompts, the same way, every month, so the comparison stays honest. A list tuned to flatter your cafe stops measuring anything real. The same goes for rerunning a disappointing prompt in the same month: log the first answer as the real result. Fix wrong facts at the source, usually your Google Business Profile or website, instead.

Work with Hillcane

Twenty fixed prompts, five groups, one score sheet, run the same way every month so this month's answer actually means something next to last month's.

If you want this run and logged for you every month instead of squeezed into a busy afternoon, reach us at hillcane.co/contact or (256) 384-2449.