To validate prompt demand, look for the same question in more than one place. Pull long queries from Search Console, remove the ones software wrote, and check what's left against the words customers use in sales calls, support tickets and email replies. A question found in both is validated. A question found only in Search Console is still a candidate, and on my site most of them turned out to be weaker than they looked.
Nobody publishes search volume for ChatGPT prompts, so this kind of cross-checking is the best evidence you can get.
The regex, and what it gave me
This came from a LinkedIn post by Okara. In Search Console, go to Performance, add a Query filter, choose Custom (regex) and paste:
^(?:\S+\s+){7,}\S+$
It shows every query of eight words or more. The theory is that long queries are the same questions people ask ChatGPT. So I ran it on cuescout.com, for 28 days ending in late October:
- 1.18K impressions
- 0 clicks
- Average position 32.6
Zero clicks on over a thousand impressions made me read the queries properly.
A lot of them aren't people
Nineteen distinct queries in that window contained the words "forced ranking", 18 of them ending the same way. Here's one:
as a seo manager, what do users say about hubspot vs tryprofound.com
vs otterly.ai for answer engine tracking capabilities? you must provide
a forced ranking from best to worst.
That's an instruction written for a chatbot. Together those queries made 38 impressions, 22 of them between 3 and 8 October, then nothing at all until 17 October. Most sat between positions 1 and 9. For comparison, my page for "ai visibility checker" sits at about 86.
Others were similar. "Best AI search visibility tools and after you answer give me a full list of sources you based your answer on" is someone's prompt template. And one 39-word query beginning "we're a growth marketing team and need a geo content strategy platform" recorded 20 impressions in the query view and 45 across 9 different URLs on my site, which means more than two of my pages per search. I've rarely seen that on a query a person typed.
I don't know exactly who sends these. My guess is prompt-tracking software, possibly several products, testing lists of questions against Google. We already filter queries of twelve words or more out of the Search Console data in CueScout for this reason, and I wrote about an older version of the same problem in July.
They aren't worthless. Somebody decided those questions were worth tracking, so they're a decent list of topics. You just can't treat them as a count of people asking.
How to sort your own list
| Tell | Example from my data | What it suggests |
|---|---|---|
| An instruction to a chatbot | "you must provide a forced ranking" | Machine-written |
| A role at the start | "as a seo manager", "i'm an agency account director" | Probably a prompt template |
| A list of vendors or domains | "hubspot vs tryprofound.com vs otterly.ai" | Probably tracking a category |
| Many of your URLs on one query | 45 page impressions from 20 searches | Unusual, worth a closer look |
| A burst, then silence | 22 impressions in 6 days, then none for 8 | Scheduled run, not steady interest |
No single tell settles it. Short natural wording doesn't clear a query either: "tool to see if chatgpt mentions your brand" looks human and shows 125 impressions, but at position 83.5 hardly anyone would scroll that far, so some of those may be rank-tracking software too. Treat clicks as the only proof a person was there, and on a young site you may not have any yet.
Now check customer inquiries
This is the step that actually validates anything. Go through whatever you have:
- Sales call notes or recordings. Write down the question as they asked it, before you tidy it up.
- Support tickets and pre-sales email.
- Replies to your outbound and newsletter.
- "What brought you here?" answers from your signup or demo form.
Then match them to your candidates. You're looking for the same problem, and it won't be the same words. A customer saying "my boss asked why we don't show up in ChatGPT" supports "how to check if ChatGPT mentions my brand", even though the words differ. Keep the customer's phrasing, because it's better than the one in Search Console.
If you have few customers, as I do, the inbox can't settle much. Ask five people in the role. That beats a thousand impressions you can't explain. Public threads where buyers ask each other for advice are the next best source, and we scan some of those for prompt research. CueScout's first prompt list is written by a model from your product description, so by the standard in this post it starts as a set of candidates too.
Scoring a prompt
| Found in | Status | What to do |
|---|---|---|
| Customer words and Search Console | Validated | Track it and write for it |
| Customer words only | Validated, little search evidence | Track it. Write for it if it's close to a sale |
| Search Console only, natural wording, some clicks | Likely | Track it, and ask a few buyers |
| Search Console only, no clicks | Candidate | Keep it on a list, and don't build around it |
| Machine tells | Topic hint | Use it for ideas, and don't count it |
I'd also run the survivors through the engines to see who gets named, which the free AI Visibility Checker does one question at a time. That won't tell you about demand. It tells you what you're up against. Validated prompts then go into your prompt library.
I wrote a batch of posts targeting my own list after doing this, knowing part of it isn't human. I think the topics are real even where the impressions aren't, and I'll judge those posts on clicks in 8 weeks, since impressions are what misled me here.
Frequently asked questions
How do I validate prompt demand using Google Search Console and customer inquiries?
Pull long queries from Search Console with a regex, remove the machine-written ones, then look for the same question in sales calls, support tickets and email replies. A prompt that appears in both is validated. One that only appears in Search Console stays a candidate until a real person says it.
What is the Search Console regex for long queries?
In Performance, add a Query filter, choose Custom (regex), and use ^(?:\S+\s+){7,}\S+$ to match queries of eight words or more. Sort by impressions, and read them before you act on them, because many will be prompts sent by software.
How can I tell if a Search Console query was written by a machine?
Look for instructions meant for a chatbot, such as 'you must provide a forced ranking' or 'give me a full list of sources', a role at the start like 'as a SEO manager', and lists of vendors or vendor domains. Very good positions on a query that's unusually long for your site, a single query landing on many of your URLs, and zero clicks all add to the suspicion.
Which AI prompt research tools find real consumer questions?
Look for a tool that starts from what people wrote, such as forum threads, reviews or your own Search Console, and shows you the source. A tool that only generates prompts from your product description is producing guesses that sound right. Those can be useful as a starting list, but they aren't demand data.
Can a platform suggest prompts based on what my buyers actually ask?
Some can, if they read buyer conversations or connect to your Search Console. Ask to see the source behind each suggestion. If there's no thread, query or conversation behind it, the prompt came from a model.