Which AI Prompts Should a Local Business Track? Build Coverage, Not a Wish List

Ask a local business owner which prompts they want to track in ChatGPT and you will have a list inside ninety seconds. Best plumber in Leeds. Top rated dentist Manchester. Cheapest accountant near me. It is a fast, confident list, and it is the wrong shape, because it describes the answers the owner would like to win rather than the range of ways a customer arrives at the question. Nobody can know the exact sentence a stranger typed last Tuesday. What you can do is cover the space of reasons people ask, and then find out whether the wider web backs you up across it. This post is about how to build that coverage: where the raw material comes from, what shape it takes, why it is different for an emergency plumber than for an orthodontist, and what the results are actually telling you.

Start with what people actually type

This is no longer a niche behaviour you can reason about from your own habits. About half of US adults now report using AI chatbots, up from a third two years earlier.¹ More to the point, the largest study of what they use them for, run over a representative sample of real ChatGPT conversations, found that Practical Guidance and Seeking Information sit alongside Writing as the three dominant categories, together accounting for close to eighty per cent of conversations.² Practical guidance is exactly the shape of who should I call about this.

Now put that next to how people search. Decades of usability research put the typical search query at two or three words, reaching for the familiar term rather than the clever one.³ A question to an assistant has no such ceiling, so it carries the things a search box forced people to drop: the situation, the deadline, the budget, the constraint. When researchers wanted to study real assistant use they had to go and collect a million consented conversations, precisely because prompts written by researchers do not resemble what people actually send.⁴ That is the same trap as the ninety-second list. A list written from your own head is a list of what you think the question looks like.

The left column is not a strawman, it is what most tracking lists look like on day one. Each entry is a keyword wearing the costume of a question. They have no situation in them, no constraint, and two of them have no place at all, so an assistant either hedges, answers about the wrong city, or produces a general essay about choosing a plumber. Meanwhile a superlative invites exactly the response you cannot use: it depends what you need, here are some things to look for. You run a list like that, find yourself missing from most of it, and conclude you have a visibility problem. You may well have one. What you definitely have is a coverage problem, and until that is fixed you cannot see past it.

Where the raw material comes from

You already own everything you need to write the right-hand column. It arrives in two streams, and the second one is the one almost nobody uses.

The keywords you already track

If you run local rank tracking, that list has already been filtered by reality: somebody picked those terms because they bring work in, they have survived contact with your Search Console data, and the services behind them are ones you want more of. Keep the ones with money behind them and cut the purely informational. What remains are the topics, and every topic becomes a small ladder of questions rather than a single line.

Your own first-party record of how customers talk

This is the good stuff, and it is sitting in systems you already pay for. Your customers have been describing their problem to you, in their own words, for years. Nobody has to invent anything: it needs collecting and tidying, not imagining.

The conversion is mechanical. Take what the customer said, drop the pleasantries, keep the constraint, and write it as one sentence a stranger could send cold. The constraint is the valuable part and the part an invented list never contains, because you do not think to invent before Friday, or with parking, or who takes NHS referrals, or someone who has done a listed building before.

Why coverage, and not guesswork

Here is the part that makes the method make sense, and it is the reason we draft questions for you rather than handing you an empty box.

When someone asks an answer engine a real question, the engine does not run that one sentence and stop. Google describes AI Mode as using a query fan-out technique, issuing multiple related searches concurrently across subtopics and multiple data sources, then bringing the results together into one response.⁵ ChatGPT does its own version, turning your request into searches of its own and assembling a reply from what comes back.⁶ We have a full guide to query fan-out because the implication is large: the sentence a customer types is the start of a retrieval process, not the whole of it. Tracking one phrasing tests one narrow slice of that process.

A tracked query is not a keyword you rank for. It is a probe into a space, and coverage is how many parts of that space you have probed.

The second reason is what you are ultimately measuring. Research on how models learn found a strong relationship between a model's accuracy on a fact and how many documents supporting that fact appeared in its training data.⁷ Things the web says often are known well; things the web barely mentions are known badly, and scaling the model up helps far less than you would hope. Retrieving pages at answer time softens this but does not remove it,⁸ because retrieval still has to find something to bring back. For a local business this translates cleanly. Your visibility in an answer is downstream of what the wider web says about you, how consistently it says it, and in how many places. That is the same prominence idea Google has documented for local results for years.⁹

Which is why the point of a coverage plan is not to guess wording. It is to probe that consensus from enough angles to find where it is thin. If you are named confidently for one service and absent on the neighbouring one, you have not discovered a prompt-writing failure. You have found the edge of your own footprint, and the fix lives in entity clarity, citation consistency and the rest of your local signals rather than in the question you asked.

One keyword, a ladder of queries

Every keyword can be asked from several distances. Someone standing in a flooded kitchen asks differently from someone with three quotes on the table, who asks differently again from someone idly wondering what a job like this costs. Those are rungs on a ladder, and how many rungs you take is a budget decision, so it is worth knowing what each buys you.

The ladder changes shape by sector, and that is the whole trick

Here is where generic advice falls over. The rungs are not the same for every business, because the reasons people ask are not the same. What sits behind a question is what we call the angle: urgency, price, credentials, occasion, whether you can start this year. A dentist and an emergency plumber both have a decision rung and a comparison rung, but the contents have almost nothing in common, and a plan that treats them identically ends up measuring the wrong half of one of them.

Three worked examples, then two more in brief. Read them for the pattern rather than the words: what changes between them is the order of the angles, and the order of the angles is what your sector decides.

Two more, compressed. The skeleton is the same, three rungs at the point of decision, but watch what happens at the comparison stage: for a builder the real second question is quotes and risk, and for a professional firm it is whether you have done this for someone like them.

Your own name belongs in the coverage plan

Service questions tell you whether you get discovered. They do not tell you what the web has concluded about you, and that is a separate and highly concentrated signal. Somebody who has your card in their hand, or who was given your name by a friend, goes and asks about you directly. What comes back is the consensus, summarised.

1. 1

   Ask what you are known for

   Tell me about [your business] in [your town]. Read the answer for what it gets wrong as much as what it gets right. Wrong opening hours, a service you stopped offering, a merged description of you and a similarly named firm: each is an entity problem with a specific fix.
2. 2

   Ask whether you are any good

   Is [your business] any good, what do people say about them? This one runs almost entirely on reviews and third-party mentions, so the answer is a direct readout of your reputation footprint rather than your website.
3. 3

   Ask how you compare

   How does [your business] compare with other [service] in [town]? The businesses named alongside you are the peer set the web believes you belong to, which is often not the peer set you would have chosen. See where those mentions live.

Two or three brand questions per location is plenty, and they are the cheapest diagnostic in the set. A thin, hedging or wrong answer to a direct question about your own name is the clearest possible evidence that the consensus problem is upstream of anything you write. It is also the fastest thing to fix, because the sources feeding it are ones you can go and correct today.

Breadth beats depth, and it is not close

Suppose you can afford forty checks a month. You can spend them as twenty keywords with two questions each, or as four keywords with ten. The second option feels more thorough and is close to worthless, because ten rephrasings of the same question mostly agree with each other. You will have bought ten slightly different views of one thing and no view at all of the sixteen services you did not check.

The variance that matters sits between keywords, not between phrasings of one. A business is rarely uniformly visible. It gets named consistently for the service it is known for, and is invisible for the two service lines it has been quietly trying to grow, which is exactly what you would predict from the way coverage in the training data and in retrievable sources tracks how well a thing is known.⁷ You only find that by asking about all of them.

Put the place in the query, or you are measuring someone else’s market

A real customer never has to say where they are. Their phone, their account and their history supply it, and Google resolves location at the moment of the search.¹⁰ Assistants carry their own version of that context, including what they remember about the person asking.¹¹ When you send a question through a tracking tool, none of that exists, so if you do not write the place into the question, the engine answers for everywhere and nowhere. This is the single most common reason a tracking setup produces results that look inexplicably wrong.

City or town level is almost always right. Go finer, to a postcode or a street name, and the model has very little to work with and starts guessing. Go broader, to a county or a country, and you are measuring a national conversation dominated by national brands, which is not the market you compete in.

Ask every engine the same set

ChatGPT and Google’s AI Mode do not work the same way and will not give you the same answer. One turns your request into its own searches and assembles a reply from what comes back;⁶ the other fans your question out across subtopics inside Google’s own stack.⁵ A business can be well covered in one and thin in the other, and that gap is genuinely useful information about where your work is landing.

It is only useful if the questions are identical. Run one set on one engine and a different set on the other and you can never tell whether the difference in front of you is the engine or the question, which makes the comparison worse than not running it. Same ladder, every engine, every time.

One thing not to do: never read the order businesses appear in as a ranking. An answer engine is writing a sentence, not publishing a results page, and the name it happens to mention first is a feature of prose. There is no position to track. What you can track is how often you appear at all, and how much of the available mention space is yours against the competitors who keep coming up.

What to measure once the questions are running

Read the trend rather than any single check, and understand why that is not a hedge. Identical requests to a model can return different text run to run, largely because the load on the server changes the batch a request is processed in, which changes the arithmetic slightly.¹² That variation is in the infrastructure, not in your setup and not in your marketing. One check is an anecdote. Eight weeks of repeated checks is a measurement. Keep the responses word for word too, because a number without the reply behind it cannot be audited, and a number you cannot audit is one you should not put in front of a client.

Once a question is running, stop editing it

This one catches people out. Rewriting a question that has been running for two months does not improve your history, it ends it. Whatever trend line you had was keyed to that wording, and the new wording starts from nothing. The discipline is the same one that makes rank tracking meaningful: hold the ruler still, so the change you see is a change in the world rather than a change in the measurement.

So add rather than edit. New service, new keyword, new ladder. Want to test a different angle, add it as a rung and leave the existing ones alone. Review the whole set once a quarter, or whenever your service list genuinely changes, and expect most of it to stay exactly as it was.

How this is built into SearchOps

Everything above works with a spreadsheet and a lot of patience. We built it into the product because the patience is where it falls apart. Someone sets up a careful ladder in month one, adds four keywords in month three, never goes back to expand them, and the new services stay invisible in the reporting for a year before anybody notices.

In ChatGPT Visibility, the ladder is called Coverage, and it is stored against the location rather than rebuilt each time. You choose the depth once, ready to buy, comparing, or full funnel, and the questions are drafted from your tracked keywords using the angles that suit your sector, resolved from your business category with an override if we get it wrong. Because the plan is stored, adding keywords later is not a trap: next time coverage is read, new keywords are brought up to the depth you already chose, whichever route they arrived by. Brand evaluation is a switch on the same plan, so the questions about your own name run alongside the service ones. Our agent then improves the drafted wording using what it knows about your business, and the same set runs against every engine you have switched on, so the Gemini numbers and the ChatGPT numbers are actually comparable.

Two things it deliberately does not do. It does not claim to guess the exact sentence your customer will type, because that is not knowable and pretending otherwise would be selling you a story. And it will not quietly redraft questions that are already running, for the reason in the section above. What it does instead is make sure the space is covered, keep it covered as your keyword list grows, and show you the parts of it where the web has not yet made up its mind about you.

The audit, in one page

• Every question traces back to a keyword you already track or to something a customer actually said, in a call, a form, a chat or a review.
• The qualifiers your customers repeat, the deadline, the constraint, the worry, appear in the question set rather than only in your notes.
• Every question names a real place you trade in, qualified if the town name exists in more than one country.
• Every question is written the way a customer talks, demands included, rather than in keyword shorthand.
• Most questions sit at the point of decision. At most one per keyword is an early-journey question.
• Two or three questions ask about your business by name, so you can see the consensus directly.
• Breadth before depth: every service line that matters has at least two questions on it.
• The same question set runs against every engine you report on.
• You are reporting visibility rate, share of voice and cited sources, and never a position or a place in a list.
• Responses are kept in full, and the headline is the trend across repeated runs rather than the latest single check.
• Wording of running questions is frozen. New angles get added as new rungs, not as edits.
• Keywords added since setup have been expanded to the same depth as the originals.

None of this is exotic. It is the same discipline that made rank tracking useful twenty years ago: measure the things people actually do, hold the ruler still, and read the trend rather than the reading. What is new is that the thing people do is now typing a sentence rather than a fragment, that the engine turns that sentence into a search of its own, and that whether your name survives the process depends on what the rest of the web has been saying about you. Cover the space, and you can finally see which part of that is missing.