In my last article, I made the case that search volume is screening out your best content opportunities. The short version is that the highest-intent questions your buyers ask are increasingly being said to an AI assistant, and they rarely show up in a keyword tool. Near the end of that piece, I listed a few places to find those questions instead: People Also Ask, forums, your own sales calls and support tickets, and the assistants themselves. That list was only four bullet points, and the most common reply I’ve had since is some version of “OK, but how do I actually do that?”
So, this is the practical follow-up. I’ll go through how I collect these questions, how I use Search Console to check which ones are already reaching a site, and how I turn the whole lot into something a content team can work from.
Start With Your Own Calls And Tickets
I always start here, because it’s the source most teams overlook, and it’s usually the best one. Sales calls, discovery calls, support tickets, live chat logs, and the messages people send before booking a demo are all people describing their problem in their own words. That’s very close to how they’d describe it to ChatGPT or Gemini.
Here’s a real example from a recent discovery call of mine (anonymized, with the details changed enough that you couldn’t identify the business). The prospect was a furniture retailer, and this was roughly how they opened the call:
“We moved to Shopify a couple of years ago and lost a big chunk of our rankings. Some of it came back, and then we dropped again this year. Is that still the migration, or are our own collection pages competing with each other?”
A few minutes later, they asked a second question:
“We sell exactly the same products in France and Germany, but those sites barely get any traffic. Do we need completely separate content for each country?”
Neither of those would ever appear in a keyword export. The closest you’d get is something like “shopify migration SEO” or “international SEO,” which tells you almost nothing about the actual situation. But if you look at those two questions as a content brief, they’re incredibly specific: a platform, a timeline, two possible causes, and a decision to make.
This is exactly the kind of question a buyer takes to an assistant, and it’s the kind of question your content should be able to answer.
How I Pull These Out At Scale
If you record your calls (I use Fireflies, but most call recorders produce a transcript you can export), I’d take the last three months of sales and discovery calls and run them through an AI assistant with a prompt like this:
“From these call transcripts, list every question the prospect asked or implied. Rewrite each one the way they would phrase it to an AI assistant, and keep their constraints, such as company size, budget, timeline, platform, and what they’ve already tried. Group similar questions together and note how many calls each group came up in.”
The final instruction here is critical. If the same scenario comes up in five separate calls, that’s the closest thing to demand data you’ll get for a question with no search volume, and it’s a lot more persuasive in a meeting than “I think people are asking this.”
You want to be careful with a couple of things here. Strip out names, company details, and anything identifying before transcripts go into a third-party tool, and make sure your call recording consent covers this kind of use. Also, sales calls skew toward people who are already close to buying, so I’d add support tickets and pre-sale inquiries into the mix to capture earlier-stage questions too.
Use Forums To Fill In The Rest Of The Situation
Your own calls only tell you about people who already found you. Forums help with everyone else.
Reddit is the obvious one, but Quora, industry Slack and Discord communities, Facebook groups, and review sites such as G2 or Trustpilot can all be just as useful depending on your market. The value usually lies in the opening post, where someone explains their constraints in full, and in the replies, where other people ask the follow-up questions an assistant would also need answered before it could give a decent response.
Sticking with the migration example, a search like site:reddit.com shopify migration rankings dropped should bring up plenty of threads where store owners describe the same problem in their own words, often with details my prospect didn’t mention, such as redirect issues or how long they waited before panicking.
I copy the most relevant threads into a document and run the same extraction prompt over them. That way, forum questions and sales questions end up in the same format, and you can group them together later.
See also: The Review Gap: Finding Client Opportunities In Competitor Feedback
See How The Assistants Break A Question Down
Once you have a set of real questions, the next step is working out what an AI assistant would actually search for when answering them.
Google is open about the fact that AI Overviews and AI Mode may use a query fan-out technique, issuing multiple related searches to build a single response. You can’t see those searches for a real AI Mode session, but you can get a reasonable approximation through the Gemini API.
When you use Grounding with Google Search, the API response includes a field called webSearchQueries, which lists the searches the model ran to answer your prompt. Google’s own documentation describes it as useful for understanding the model’s reasoning process.
If you feed in the first migration question from earlier, you’d expect to see sub-queries covering things like how long rankings take to recover after a Shopify migration, how to check redirects after a re-platform, and how to spot keyword cannibalization between collection pages. Your own output will vary from run to run, but each sub-query is effectively a section your page would need to cover to be useful as a source.
You don’t need to be technical for this. Go Fish Digital published a walkthrough that runs it through Screaming Frog with a copy-and-paste custom JavaScript snippet and an API key from Google AI Studio. Their version sends each page’s H1, but swapping in your own list of questions is straightforward.
I want to be clear about what this output is, though. It shows you how a Gemini model expands a question, which is a fair proxy for Google’s AI features, but it isn’t a record of what AI Mode ran for an actual user. I treat it as a coverage checklist, and I wouldn’t present it to a client as anything more definitive than that.
For ChatGPT, Perplexity, and the other assistants, the manual route still works fine. Ask each one your question, note which sub-topics it decides to cover and which sources it cites, and compare that against the page that’s supposed to answer it.
See also: The 2003 Framework That Was Already Doing Query Fan-Out
Check Search Console For What’s Already Reaching You
Search Console is better as a validation step than a discovery tool for this, because it shows you which of these questions are already finding their way to your site.
Long conversational queries do turn up in the Performance report. Suganthan Mohanadasan has already covered four ways to surface them in a lot of detail, so I’d recommend checking that out if you haven’t already. For a quick look, I use the Custom (regex) option in the Query filter. This pattern returns queries of eight words or more:
^([^ ]+ ){7,}[^ ]+$
And this one returns queries that start with a question word:
^(how|what|why|which|should|can|do|does|is|are|when|where|who)\s
Then I compare what comes back against the questions from calls and forums. If the migration scenario shows up as a long query, that’s good confirmation that real people are bringing it to Google. If it doesn’t show up, it could be a coverage gap, or the queries may simply be too rare for Google to show you.
That last point is significant because Google anonymizes rare queries, and conversational queries are rare almost by definition. Suganthan found that 57.7% of his impressions sat in the anonymized pool over a 59-day period, so anything you find here is an undercount.
If you have access to the generative AI performance report, the pages view is worth a look too. It shows which of your URLs are already appearing in AI Overviews and AI Mode. There’s no query data in it (yet), and sites with few AI impressions may not see the report at all, but it does tell you which pages Google is already comfortable using as a source.
Turn It All Into A Content Plan
By now, you’ll have a messy list of questions from several sources. To get it into a shape a content team can actually use, I put everything into one sheet. Here’s how a single row might look, using the migration example:
| Column | Example |
|---|---|
| Question | “We moved to Shopify, lost rankings, recovered some, then dropped again. Is it the migration or are our collection pages competing?” |
| Source | Discovery call, plus several Reddit threads |
| Frequency | For example, 3 of the last 20 calls |
| Situation | Ecommerce, Shopify, migrated around two years ago, second drop this year |
| Sub-questions | Recovery timelines, redirect checks, collection page cannibalization |
| Existing page | Shopify SEO service page |
| Coverage | Partly answered (covers migrations, not cannibalization) |
| Priority | High |
For the priority column, I use the same proxies from my previous column: business value, evidence that the question is being asked, how big the coverage gap is, and whether you can answer it with real specifics. The frequency column feeds straight into the second one.
Once the sheet is filled in, you’ll usually notice that most of the questions cluster around a handful of situations. Those clusters are your content plan. In my experience, most of them can be handled by strengthening a page you already have, which is quicker than publishing something new and tends to perform better too.
What This Won’t Tell You
I’m aware this is a lot less tidy than sorting a spreadsheet by volume, so I think it’s worth being up front about the limitations.
You still won’t have a volume figure next to any of these questions, which means you’re prioritizing on judgment and frequency. Your own calls are biased toward people who already know you exist, so questions from audiences you don’t reach yet will be underrepresented. The Gemini fan-out is an approximation of Google’s behavior, and Search Console only shows you part of the picture because of anonymization.
However, I don’t think any of that is a reason not to do it. It just means you should present the output as a well-evidenced shortlist, particularly to stakeholders who are used to seeing a number next to every target or goal.
Over To You
If you want somewhere to start right now:
- Export three months of call transcripts or support tickets, anonymize them, and run the extraction prompt above. Even 20 calls will probably give you more usable questions than your last keyword export.
- Take your five most frequent scenarios and run them through Gemini with grounding switched on to see which sub-queries come back.
- Check those five against the page that should answer each one, and fix the gaps on your strongest existing page first.
Keyword tools are still useful for understanding what people type into a search box. But a growing share of the questions that decide whether a buyer picks you are being asked in a conversation with an assistant, and the good news is that plenty of them are already sitting in your inbox and call recordings.
More Resources:
- AI Answers About Your Locations Are Often Wrong – Check Before Customers Do
- The AI Conversations Leaking Into Your Search Console
- AI Visibility Measurement: What To Track & What To Ignore
Featured Image: kmzway/Shutterstock