People do not ask assistants one question before hiring a contractor. They ask a few, in a rough order, and the order tells you something about where they are in the decision.
These are the eight shapes that cover almost all of it. AIgnored runs on the same idea with 8 broad phrasings of its own, listed in every scan's why drawer, and the last section says which of these shapes they cover. Eight is enough to see a pattern and few enough to run cheaply, and each shape pulls a different set of names.
Swap the trade and the city and they work for plumbers, roofers, electricians, HVAC, dentists, lawyers or anything else people hire locally.
1. The comparison question
Who is the best {trade} in {city}?
The one everybody assumes is the important one. It is the hardest to win and the least commercially urgent, because the person asking is orienting rather than buying.
It pulls the broadest candidate set: the businesses that appear on "best of" listicles and directories, which is a popularity contest decided mostly by who is on those pages. A brand new business will almost never appear here. A twelve year old business with no directory presence will also never appear here, which surprises people.
Useful as a benchmark. Not where the money is.
2. The decision question
I need a {trade} in {city}. Who should I call first?
Same topic, different framing, and it produces different answers more often than you would expect. "Who should I call first" pushes the model to commit to a single name and justify it, which changes what it selects for: it favours businesses that come with a reason attached rather than businesses that appear on the most lists.
This is where quotable facts earn their keep. "Call X first, they do 24 hour callouts" is a sentence the model can write. It cannot write "call X first, they are professional". Why a reason matters so much is covered step by step in how to show up in ChatGPT recommendations.
3. The trust question
Recommend a reliable {trade} in {city}.
The word "reliable" pulls review language. Answers to this shape lean harder on rating, review count and review themes, and they are the ones most likely to quote a number back at you.
If a business has volume of reviews but no theme in them, this is the question where that shows: the model has the count but nothing to say about it, so it names someone with a weaker rating and a clearer story.
4. The urgency question
Best emergency {trade} in {city}?
The most winnable question in most categories, and the most valuable.
Urgency questions have a much smaller candidate set, because most businesses do not say anywhere that they do emergency work. Saying it plainly, on your own page and in the services list of your Google Business Profile, is often enough to get into that smaller set.
The person asking this question is not comparing. They are about to call someone.
5. The service-specific question
Who does {specific service} in {city}?
Water heater replacement, flat roof repair, consumer unit upgrades, clear aligner treatment. Whatever the concrete job is.
This shape is where a small business can beat a large one outright. The large competitor's pages talk about the trade in general; the specialist's page talks about the exact job. Models composing an answer for a specific service prefer the page that names that service.
If you sell three services, this is three separate winnable questions, and most of your competitors have written a page for none of them.
6. The proof question
Top rated {trade}s in {city} with good reviews?
Explicitly asks for ratings, so it produces the most number-heavy answers, and we expect it to name more businesses per answer than the other shapes. That is easy to check: count the names.
Useful for auditing because a business that appears here but nowhere else has a good profile and no other presence. That is a very specific diagnosis with a very specific fix.
7. The price question
Affordable {trade} in {city}?
The shape most likely to get a hedged answer, because assistants are cautious about implying a business is cheap. Often the model will answer with advice about getting three quotes instead of names.
Worth including anyway, for two reasons. Some categories do produce names here, and the businesses that appear are almost always the ones who published price ranges. Publishing a price range is uncomfortable and it is one of the few things that reliably differentiates an answer.
8. The personal question
Which {trade} in {city} would you use yourself?
An odd phrasing that people genuinely use. It pushes the model toward a single confident recommendation and away from a hedged list.
It also produces the most variance run to run, which is a useful reminder: if a single answer is your evidence, you have no evidence. This shape shows you the noise floor.
Why the spread matters more than the count
Run all eight and the pattern is usually not "invisible everywhere" or "visible everywhere". It is lumpy.
The shape we expect, and the one worth checking for on a business you are auditing, is a business that scores zero on the broad comparison questions where the assistant leans on directories, and picks up its only mentions on the specific-service questions where its own pages give the model something to quote.
We are not going to put a number on how common that is until we have run enough scans to publish one. When we do, it will be here with the sample size next to it.
That shape difference is the actual story you tell a client. "You went from zero to four" is a number. "You now appear for emergency and water heater questions, which are the two people call about, and you are still absent from the comparison question, which we can work on next" is a plan.
Running the set properly
Three rules, and none of them are optional if you want the numbers to mean anything.
Fresh conversation for every question. Earlier turns become context. Ask about a business and then ask who is best, and you have told the model the answer.
Logged out or memory off. Otherwise you are measuring your own history.
Record every name in every answer, not just the one you care about. The competitor list is the more useful output. It tells you who occupies the slots and, when you look at their profiles side by side, usually why.
Ask across at least two assistants. They read different sources and disagree often enough that a single-model audit overstates its own certainty. The 15-minute manual audit turns these rules into a spreadsheet you can fill in, one row per question.
Turning eight answers into one number
Count how many of the eight answers named the business, divide by the answers that completed, and round. Named in three of eight is 38. Named in none is 0.
Do the same for the competitors that came up most, and you have a table with a gap in it. That table is the whole argument, and it needs no adjectives.
One caution worth repeating: models are not deterministic, so re-running the set next week will move the numbers. A business at 38 might read 25 or 50. A business at 0 tends to stay at 0, which is why the zeros are the finding you can act on with confidence. The methodology page goes through this and the other limits in detail, including the difference between querying the APIs and using the consumer apps.
If you want to run this at volume
Eight questions across two models is sixteen answers per business. By hand that is about a quarter of an hour, by our own estimate. For one business, fine. For every plumber in a city, it is a week.
That is the part AIgnored automates, with one difference worth knowing before you rely on it. A scan does not ask the eight shapes above. By default it asks 8 broad recommendation questions per model: who is the best, who to call first, who is top rated right now, who is reliable for an urgent job, who locals recommend, who is most trusted, who does the highest quality work, and a list of the best. Those cover the comparison, decision, trust, urgency and proof shapes. They do not ask the service-specific, price or personal questions, so those stay a manual check. Every scan lists its exact questions in the why drawer, and the methodology page explains how they are asked. Names are matched back to real businesses with the uncertain matches shown rather than guessed, and every business in the city is scored in one pass so you can see who is invisible and who is taking the slots. Scans run on your own provider keys, so the cost is whatever the provider charges. The pricing page has a free plan that covers a single city.