A Practical Way to Find the Questions Your Customers Ask AI (And Your Keyword Tools Miss)
/ 8 min read
Summary
I always start here, because it's the source most teams overlook, and it's usually the best one. Sales calls, discovery calls,. The practical question is what this changes for SEO, content quality, and AI search visibility.
For years, the standard operating procedure for content planning has been to plug a seed keyword into a tool, sort by volume, and pick the winners. But there is a growing gap between what people type into a search box and how they actually describe their problems. As AI assistants become the primary interface for complex queries, the most high intent questions are moving away from traditional search and into conversations. This connects with Working Framework when the same signal needs a clearer operating decision. A useful companion note is AEO Tool Stack I Would Actually Start, because it looks at a nearby part of the same system.
The problem is that these conversational queries rarely show up in keyword tools. If you only optimize for "shopify migration seo," you miss the nuance of a business owner asking why their rankings dropped six months after a move and whether their collection pages are cannibalizing each other. To capture this traffic, you have to look where the conversations are actually happening.
Mining Your Own Sales Calls and Support Tickets
The most overlooked goldmine for content is your own internal data. Sales discovery calls, support tickets, live chat logs, and pre demo inquiries are essentially raw transcripts of your customers describing their pain points in their own words. This is the exact phrasing they use when prompting an AI assistant.
Consider a furniture retailer who tells a consultant that they moved to Shopify years ago, saw a temporary recovery in rankings, but then dropped again. They aren't asking for a general guide on SEO; they are asking a specific question about the timeline of migration recovery and internal competition between pages. A keyword tool would condense this into a generic term, but the actual question contains the platform, the timeline, and the specific technical fear.
Expert Interpretation: This matters because it shifts your focus from "traffic" to "intent." The tradeoff is that this data is qualitative and small scale compared to a million search volume keyword. However, the decision you should inspect here is whether you are prioritizing volume over conversion probability. A question asked by five qualified leads is often more valuable than a keyword searched by five thousand strangers.
Scaling the Extraction Process
You don't have to manually read every transcript. If you use a call recording tool like Fireflies or any service that exports text, you can use an AI assistant to categorize these questions at scale.
I recommend feeding the last three months of transcripts into an AI with a prompt that asks it to list every implied or explicit question. Crucially, tell the AI to rewrite these questions as if the user were asking a chatbot, preserving all constraints like budget, company size, and previous attempts to fix the problem. I also ask the AI to group similar questions and count how many times they appear.
Counting the frequency is the most important part. When a specific scenario appears in five different calls, you have found a pattern. This frequency acts as a proxy for demand data, providing a much stronger justification for a content piece than a mere hunch.
A few warnings: always strip out personally identifiable information before uploading transcripts to a third party AI, and ensure your recording consent covers this type of analysis. since sales calls lean toward bottom of funnel leads, make sure to include support tickets to capture the questions people ask earlier in their journey.
Using Forums to Map the Full Customer Journey
Internal calls only tell you what people who already found you are asking. To see the rest of the market, you need to go to the forums. While Reddit is the most prominent, don't ignore Quora, industry specific Slack or Discord channels, Facebook groups, or review platforms like G2 and Trustpilot.
The real value in these forums isn't just the original post, but the thread that follows. The opening post usually outlines the constraints and the specific problem. The replies often contain the follow up questions that an AI assistant would need to answer to provide a complete solution.
For example, if you search Reddit for "shopify migration rankings dropped," you will find store owners discussing redirect errors or the specific amount of time they waited before panicking. By copying these threads and running them through the same extraction prompt used for sales calls, you can normalize the data and group forum based questions with your internal sales data. The same pattern also shows up in What AI Says About Your Locations, where the practical question is how the signal becomes visible.
Expert Interpretation: The risk here is "noise." Forums are filled with outliers and edge cases that may not represent your target customer. The decision to make is how to filter for quality. Focus on threads with high engagement or those where the user's profile matches your ideal customer profile (ICP) to avoid wasting resources on low value content.
Analyzing How AI Assistants Deconstruct Questions
Once you have a list of real world questions, you need to understand how an AI actually processes them. Google uses a "query fan out" technique for AI Overviews and AI Mode, meaning it doesn't just perform one search; it issues multiple related searches to synthesize a single, detailed answer.
While you can't see these internal searches in a standard chat session, you can approximate them using the Gemini API. By using "Grounding with Google Search," the API provides a field called webSearchQueries. This reveals the specific sub queries the model ran to find the information needed to answer the prompt.
If you input a complex question about Shopify migration, the AI might generate sub queries about recovery timelines, redirect audits, and keyword cannibalization. Each of these sub queries represents a critical section that your content must cover to be seen as a high quality source by the AI.
You don't need to be a developer to do this. There are ways to run this through tools like Screaming Frog using custom JavaScript snippets and an API key from Google AI Studio. Instead of running it on your existing H1s, run it on the list of customer questions you've collected.
Expert Interpretation: This is a shift from "keyword research" to "entity and topic research." The tradeoff is that AI responses can vary between runs. The key decision is to look for the commonalities across multiple runs. If the AI consistently searches for "redirect audits" when answering a migration question, that topic is non negotiable for your content.
Validating with Search Console
Google Search Console (GSC) is not a discovery tool for conversational queries, but it is an excellent validation tool. It allows you to see if these long tail, conversational questions are already driving traffic to your site, even if they have negligible volume in traditional tools.
To find these in the Performance report, use the Custom (regex) filter. You can use a pattern to isolate queries of eight words or more, or a pattern that filters for queries starting with question words (who, what, where, why, how).
Compare these GSC results against your list from calls and forums. If a specific migration scenario appears as a long tail query in GSC, you have empirical evidence that real people are using Google to ask that exact question. If it doesn't appear, it could be a coverage gap you can exploit, or the query might simply be too rare for Google to report. Either way, it helps you prioritize which questions to answer first.
Expert Interpretation: Remember that GSC data is anonymized. You will never see every single query, only a sample. The decision here is to use GSC as a "sanity check" rather than a definitive list. If it's in GSC, it's a priority. If it's not, it's still a potential opportunity based on your sales and forum data.
Building a Practical Content Plan
At this stage, you have a collection of questions from sales calls, forums, AI sub queries, and GSC. To make this actionable for a content team, organize it into a structured sheet. Each row should represent a specific question or scenario, including the source of the question and the frequency with which it appears.
Prioritize these based on a few proxies: the business value of the lead, the evidence that the question is being asked, the size of the current coverage gap, and your ability to provide a specific, expert answer.
You will likely find that these questions cluster around a few core situations. These clusters are your actual content plan. Rather than creating dozens of thin pages for every single variation of a question, focus on strengthening existing high value pages. Updating a complete guide to include these specific, conversational nuances is usually more effective and faster than publishing new, fragmented posts.
Expert Interpretation: The temptation is to create a "FAQ" page. Avoid this. Instead, integrate these answers into the narrative of your main service or educational pages. The decision is between "breadth" (many small posts) and "depth" (few authoritative posts). Depth almost always wins in the current AI driven search landscape.
Understanding the Limitations of This Approach
This method is less tidy than sorting a spreadsheet by search volume. You won't have a neat number telling you exactly how many people search for a term, which means you are relying on judgment and frequency rather than raw data.
There are also inherent biases. Your sales calls only reflect people who already know your brand. The Gemini API provides an approximation of Google's behavior, not a perfect mirror. Search Console only shows a fraction of the total queries due to privacy filtering.
Despite these limitations, the approach is more grounded in reality than relying on keyword tools alone. When presenting this to stakeholders who are used to seeing volume numbers, frame the output as a "well evidenced shortlist" based on actual customer behavior rather than a theoretical traffic forecast.
Putting the Process Into Action
If you want to start today, the lowest hanging fruit is your own data. Export three months of support tickets or call transcripts, anonymize them, and use an AI to extract the core questions. Even a small sample of 20 calls will likely yield more high intent opportunities than a standard keyword export.
Take the five most frequent scenarios you find and run them through Gemini with grounding enabled to see which sub queries the AI generates. Finally, compare those sub queries against your existing content. Identify the gaps and update your strongest pages first to ensure you are providing the specific answers that AI assistants are looking for when they synthesize responses for your future customers.
Comments
Comments are reviewed before they are published. Links are not allowed inside comments.