Mining sales calls and support tickets
Sales calls, discovery calls, support tickets, live chat logs, and messages sent before booking a demo provide the best data. These channels contain people describing their problems in their own words. This is close to how they would describe a problem to ChatGPT or Gemini. I look at a recent discovery call with a furniture retailer to see this in practice. The prospect explained they moved to Shopify two years ago and lost a big chunk of rankings. Some of that ranking returned, but they dropped again this year. They asked if the migration was the cause or if their own collection pages were competing with each other. They also asked if they needed completely separate content for France and Germany since they sell the same products in both countries. A keyword export would only show "shopify migration SEO" or "international SEO." Those terms tell me nothing about the specific platform, the two-year timeline, the two possible causes, or the decision regarding international content.
To find these at scale, I take three months of sales and discovery call transcripts and run them through an AI assistant. I use a prompt to list every question the prospect asked or implied, rewrite each one the way they would phrase it to an AI assistant, keep their constraints like company size or budget, and group similar questions by frequency. I strip out all names and company details before putting transcripts into a third-party tool. I must also ensure my call recording consent covers this use. If the same scenario appears in five separate calls, I have demand data for a question with no search volume.
Forums and the Gemini API
Forums fill the gaps left by internal data. Reddit is the obvious choice, but Quora, industry Slack and Discord communities, Facebook groups, and review sites like G2 or Trustpilot also work. The value sits in the opening post where someone explains their constraints, and in the replies where others ask follow-up questions. I copy relevant threads into a document and use the same AI extraction prompt used for call transcripts. Once I have a set of questions, I determine what an AI assistant would search for to answer them. Google uses a query fan-out technique, meaning AI Overviews and AI Mode issue multiple related searches to build one response. You cannot see these searches in a real AI Mode session, but the Gemini API provides a reasonable approximation.
When I use Grounding with Google Search, the API response includes a "webSearchQueries" field. This field lists the searches the model ran to answer the prompt. If I feed in the migration question, I expect to see sub-queries covering 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. Go Fish Digital published a walkthrough that uses Screaming Frog with a custom JavaScript snippet and a Google AI Studio API key to automate this process. I treat this output as a coverage checklist rather than a definitive record of what AI Mode ran for users. For ChatGPT or Perplexity, I ask the question manually, note the sub-topics it covers, and check which sources it cites.
Validation and the content plan
Search Console serves as a validation tool to see which conversational queries already find their way to a site. I use the custom regex option in the Query filter to find queries of eight words or more or those starting with a question word. I compare these results against the questions I pulled from calls and forums. If a migration scenario shows up as a long query, it confirms people bring it to Google. If it does not show up, it is a coverage gap or the queries are too rare for Google to show. Google anonymizes rare queries, so conversational queries are often hidden. Suganthan Mohanadasan found that 57.7% of his impressions sat in the anonymized pool over a 59-day period. I check the generative AI performance report to see which URLs appear in AI Overviews and AI Mode. This report does not have query data yet, and sites with few AI impressions might not see it at all.
I build my content plan in a single sheet to turn messy lists into a usable format. I use business value, frequency, and the coverage gap to set priority. I prioritize by strengthening existing pages because it is quicker than publishing new ones and usually performs better. I accept that this method lacks the volume figures found in keyword tools. I also know that my call data is biased toward people who already know my business. The Gemini fan-out is an approximation of Google behavior, and I would not present it to a client as anything more definitive than a coverage checklist. Search Console only provides a partial picture because Google anonymizes rare, conversational queries. Suganthan Mohanadasan found that 57.7% of his impressions sat in the anonymized pool over a 59-day period.
| 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 | 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 |
To start, I export three months of call transcripts or support tickets, anonymize them, and use an AI extraction prompt. I then run five frequent scenarios through Gemini with grounding to see the sub-queries. Finally, I check those scenarios against my strongest existing pages to fix any coverage gaps.

