AI Prompts That Miss Real Customers
Tracking AI answers is only useful if the questions you track are the ones your customers actually ask. Many prompt lists are copied from a keyword tool, so they read like search terms: short, generic and stripped of the situation that shapes a real request.
The result is a visibility score that can look healthy while real enquiries go to competitors, or look hopeless for questions nobody asks. This guide shows how to gather questions in customers' own words, turn them into realistic prompts, test how fragile each one is, and tie every prompt to the page that should answer it.
How a mismatched prompt set shows up
Query mismatch is easy to miss because the tracking itself works: prompts run, answers are scored and charts go up or down. The warning signs are in how the figures relate to the rest of the business.
- Good scores, no enquiries. You appear for most tracked prompts, yet new customers never say an assistant pointed them to you and AI referrals stay near zero.
- Prompts that read like search terms. Entries such as "best accountant Leeds" or "garden room prices" are how people type into a search box, not how most people talk to an assistant.
- Only one kind of question. Every prompt asks for a recommendation, and none describe a problem, compare two options or ask how something works.
- No brand or competitor prompts. You cannot see what assistants say when someone asks about you directly, or about you against the business they are also considering.
- Prompts with no owner page. Nobody can say which page on the site ought to be cited for a given prompt, so nobody knows what to improve when it scores badly.
If two or more of these apply, fixing the prompt set comes before any content work, because otherwise you will be improving pages for questions that may not matter.
How questions to assistants tend to differ
There is no reliable public dataset showing exactly how everyone phrases questions to assistants, and the providers do not publish one. What is visible, from your own customers and from the way assistants are designed to be used, is a set of tendencies worth planning around:
| Tendency | Search box | Assistant prompt |
|---|---|---|
| Length | A few words | Often one or more full sentences |
| Situation | Usually left out | Often included: size, budget, location, what they already use |
| Constraints | Implied | Stated: "under £500", "no subscription", "works with Xero" |
| Follow-ups | A new search | A second message in the same conversation |
| Request type | Find pages | Explain, compare, recommend, draft or decide |
A real example of the difference, from a fictional home improvement business, Brookvale Garden Rooms:
Search box: garden room cost uk
Assistant: We want a garden office about 3 by 4 metres in Derbyshire. The garden
slopes a bit. Roughly what would it cost, do we need planning
permission, and what should we ask companies before we get quotes?The assistant prompt contains four separate questions and three pieces of context. An answer to it will draw on pages that cover sloping sites, size-based pricing, permitted development and buying advice. A page that only lists average prices may not be used at all, even if it ranks for the short search term.
Where to find the questions people really ask
The best prompts come from records you already hold. Work through these sources and paste each question into a single list in the words the customer used.
Enquiries and support messages
Emails, contact form submissions, live chat transcripts and call notes are the closest thing you have to real prompts. People write to a business the way they write to an assistant: with context, constraints and several questions at once. Remove names and contact details before the text goes anywhere near a shared spreadsheet.
Sales and first calls
Ask whoever talks to new customers to note the first question each one asked and what they were comparing you with. Competitor names that come up here belong in your comparison prompts.
Search Console question queries
Google Search Console's Performance report can filter queries with a regular expression. Two filters surface the longer, question-style searches that sit closest to assistant prompts:
Queries matching regex (questions): ^(who|what|which|where|when|why|how|can|should|is|are|does|do)\b
Queries matching regex (7+ words): ^(\S+\s+){6,}\S+$These are still searches, not prompts, but they show the problems and situations people bring to the topic.
Your own site search
If your site has a search box, its log shows what visitors expected to find and did not. Failed searches are often good prompts for which you have no page yet.
Public discussions
Forums, community groups and review sites show how people describe the problem to each other. Note the wording and the situations; do not copy the posts themselves.
Ask customers directly
A short question after a sale, such as "if you asked an AI assistant about this, what would you type?", gives you literal prompts. A handful of answers is often enough to show patterns you had not expected.
The fix, step by step
Step 1
Collect questions in customers' own words
Gather raw questions from enquiries, support messages, call notes, site search logs and question-style queries in Search Console, and paste them into one list without tidying the language. Keep the context people gave, such as team size, budget, town or the software they already use, because that detail is what shapes an assistant's answer. Remove personal details as you go. Aim for a long list at this stage; you will cut it down in later steps.
Step 2
Turn each question into a realistic prompt
Rewrite each raw question as something a person would type into an assistant, keeping their situation and constraints but removing anything that identifies them. Where several customers ask the same thing, merge them into one prompt that carries the most common context. Write a short version and a fuller version for your most important questions, because real users do both, and give each prompt an ID so its history can be tracked once measurement starts.
Step 3
Label prompts by intent and stage
Tag every prompt with its intent, such as explain, problem, compare, recommend, price or brand, and with how close the person is to buying. Count the tags. A set made almost entirely of recommendation prompts misses the earlier questions where people first meet a brand, while a set with no brand or comparison prompts hides what assistants say when someone is deciding between you and a rival.
Step 4
Test wording variants before you commit
For each important prompt, write two or three rewordings that a real customer might use, and run all of them several times in each assistant you track. If the brands named stay broadly the same across versions, one prompt represents the group well. If small wording changes produce completely different answers, keep more than one version in the set, because tracking only one would give you a misleadingly narrow picture of that question.
Step 5
Pair every prompt with the page that should answer it
Next to each prompt, write the single URL on your site that you would most want an assistant to cite. Then read that page with the prompt in mind: does it address the situation, the constraints and each of the questions inside the prompt, in plain text near the top? Mark each pairing as covered, partly covered or missing. A prompt with no suitable page is a content gap, not a visibility problem, and should be recorded as such.
Step 6
Close the gaps the answers reveal
Run the new set and read the answers for prompts where you are invisible. Note which sources are cited and what they cover that your page does not: a situation you never mention, a comparison you avoid, a price range you do not state, or a follow-up question you leave unanswered. Add that material to the paired page, or create a page where none exists, rather than writing a new page for every rewording of the same question.
Step 7
Refresh the set on a schedule
Once a quarter, compare the tracked prompts with the latest enquiries and support messages. Questions that customers have stopped asking can be retired, and new ones added as a dated batch, so existing prompts keep their history. Never edit the wording of a prompt that already has results; retire it and add the new version with its own ID. Record every change in the same log you use for site changes.
A prompt map you can copy
A prompt map is a plain table that connects each tracked prompt to a page and a status. It turns the tracker's findings into a to-do list, because every invisible result points to a specific page and a specific gap.
| ID | Prompt | Intent | Page that should be cited | Coverage | Gap |
|---|---|---|---|---|---|
| P-01 | We want a 3 x 4 m garden office in Derbyshire on a sloping garden. Roughly what would it cost? | Price | /garden-offices/prices | Partly | No mention of sloping sites or groundwork costs |
| P-02 | Do I need planning permission for a garden room used as an office? | Explain | /guides/planning-permission | Covered | None found |
| P-03 | Brookvale Garden Rooms or Oakridge Cabins for a year-round office? | Compare | None | Missing | No page compares insulation and heating options |
| P-04 | What should I ask a garden room company before getting a quote? | Problem | /guides/before-you-buy | Partly | Answer is split across several sections |
Keep the map beside your tracking data. When a prompt moves from invisible to cited, you can see whether the cited page is the one you paired with it, which is the best evidence that the change you made was the one that counted.
Why wording sensitivity matters for measurement
Assistants respond to the exact words in a prompt. Adding a town, a budget or the name of a competitor can change which sources are retrieved and which businesses are named. This has two practical consequences for tracking.
- One wording can overstate or understate you. If your tracked prompt happens to include a phrase that appears on your own page, you may score far better on it than on the ways most customers phrase the same need. The reverse can also happen.
- Context is part of the question. "Best accounting software" and "best accounting software for a sole trader who already has a business bank account and invoices by email" are different questions with different answers. If most of your customers bring the second kind, tracking the first tells you little about the enquiries you are missing.
The fix is not to track every possible wording. It is to test a few variants of each important question once, keep the ones that behave differently, and group them so that a single report line summarises the whole question.
Edge cases
Follow-up questions
Many conversations start broad and narrow down: "what are my options?" followed by "which of those works for a small team?". Most tracking runs single prompts, so fold the most common follow-up into a single, more specific prompt rather than trying to script whole conversations.
Several countries or languages
Prompts in another language, or naming another country, are separate questions. Track them separately with their own page pairings, and do not assume an English-language result carries over.
Seasonal questions
Some questions only matter at certain times of year. Keep them in the set all year so the history stays continuous, and read their results in season.
Questions you should not win
Some real customer questions are better answered by someone else, such as official guidance on a legal requirement. Being named as a sensible next step can still be useful, but do not count the absence of a citation there as a failure.
Keeping prompts and customers in step
- Make question-gathering routine. Ask the people who handle enquiries to forward interesting questions to one shared place, stripped of personal details.
- Review the intent mix each quarter. If one type starts to dominate the set, rebalance it with a new batch.
- Keep the prompt map current. When a page is merged, moved or removed, update every prompt paired with it the same day.
- Check new products and services. Each launch needs its own prompts before it goes live, so there is a baseline to compare against.
Once the prompt set is sound, make sure the way you run and record it is equally consistent; fixing AI visibility tracking covers that part.
1. Can I use my SEO keyword list as AI prompts?
You can use it as a starting point, but not as it stands. Keyword lists are built from short search terms, while many people write to assistants in full sentences that include their situation and constraints. Take each important keyword, find the real questions behind it in your enquiries and support messages, and rewrite it as a prompt a customer would type. Keep a short version too, because some people do use assistants like a search box.
2. How do I find out what people ask ChatGPT about my industry?
OpenAI and the other providers do not publish what individual users ask, so there is no direct report to look at. The closest evidence is your own: enquiry emails, support messages, call notes and site search logs show how customers phrase questions. Question-style and long queries in Search Console add the situations people bring to the topic. Asking recent customers what they would type into an assistant is simple and often surprisingly revealing.
3. Should prompts include my brand name?
Some should and most should not. Brand prompts show what assistants say when someone asks about you directly, which reveals accuracy problems and whether you are recognised at all. Comparison prompts naming you and a competitor show how you are described at the point of decision. The bulk of the set, though, should be unbranded category and problem questions, because those show whether you are suggested to people who have not heard of you yet.
4. How specific should a tracked prompt be?
As specific as your typical customer. If most enquiries mention a town, a budget or a type of business, put that context in the prompt, because it changes which sources an assistant uses and which businesses it names. Avoid adding detail that only you would think of, such as your product names, since that tilts the result in your favour. For important questions, keep a short and a detailed version and track both.
5. Why do small wording changes give completely different AI answers?
Assistants that search the web build a search from the words in the prompt, so adding a place, a price or a competitor can retrieve a different set of pages. The model then writes a fresh answer from those pages, with some randomness in the result. That combination makes some questions very sensitive to phrasing. Testing a few rewordings of each important prompt shows which questions are fragile, so you can track more than one version of them.
6. How often should I change my tracked prompts?
Rarely, and never by editing prompts that already have history. Changing wording breaks the comparison with earlier results, so a figure that moves may only reflect the edit. Review the set about once a quarter against recent customer questions, retire prompts nobody asks any more, and add new ones as a separate dated batch with their own IDs. That way the long-running prompts keep a clean history while the set still reflects real demand.
7. What if no page on my site answers a tracked prompt?
Then the prompt has found a content gap, and no amount of technical work will get you cited for it. Record it as missing in your prompt map and decide whether the question is worth answering. If it is, add a clear answer to the closest existing page or create a new page when the topic is distinct. Read the sources assistants cite for that prompt first, so you know what a useful answer needs to cover.
The AI Visibility Tracker runs your target prompts in ChatGPT, Claude, Perplexity, Gemini, Copilot and Brave Leo, records whether your brand is cited, mentioned, recommended or invisible for each, and identifies citation gaps against competitors.
Open AI Visibility Fixes →