AI Search Queries Your LLM Fires Mid-Answer

The new keyword research lives inside the model

Traditional keyword research starts with what humans type into a search box. AI search has produced a new layer above that: when a user asks ChatGPT, Claude, or Gemini a natural-language question, the model often fires its own web searches mid-answer to ground its response. Those internal queries are not visible in Google Search Console. They never reach your traditional analytics. But they are the highest-quality buyer-intent signal currently available, because they were generated by a model that already understood the user's intent and translated it into the query most likely to surface a useful answer.

The user said "is Profound worth it for a small SaaS". ChatGPT translated that to "Profound vs Lumos vs Peec AI" and went looking for grounding. The translated query is the one you need to be findable on. Optimizing for the user-facing query is optimizing for the old SEO surface. Optimizing for the translated query is optimizing for the new one.

What you see inside Lumos

The preview above is a snapshot of the queries ChatGPT and Claude fired mid-answer for prompts about Lumos and GEO this week. Signing up runs the same capture on your prompt library.

• User-facing prompt → translated query. Each row pairs the original natural-language question with the web query the AI fired to ground its answer. Both sides matter — you need to know what users say and what AI translates it to.
• Frequency and buyer-intent score. Queries that repeat across multiple user prompts get higher signal. A query fired once is noise; a query fired across 14 different user prompts is a fixed asset of your category.
• Source pages returned. Lumos captures the actual URLs each query returned, so you can see which pages AI is grounding on for that query.
• Engine breakdown. ChatGPT and Claude fire different queries for the same user prompt because their tool-use prompts and grounding strategies differ. Lumos shows both so you can optimize for the full surface.

How Lumos captures the queries

The mid-answer query is the artifact of the model's tool use. Lumos runs your prompt library through each major engine with tool-use enabled and captures:

1. The user's natural-language prompt.
2. The mid-answer web queries the model fired (sometimes one, often three to five).
3. The URLs returned for each query.
4. How the final answer cited or paraphrased those URLs.
5. The model and engine version, so you can compare apples to apples week over week.

The raw query log is stored alongside your normal scan data. You can filter, search, and export it. The aggregation layer rolls up duplicates and ranks by frequency × buyer-intent score.

How to use the captured queries

Build content that matches the translated form. AI is looking for exact-match titles like "Profound vs Lumos vs Peec AI". If your content is titled "Why we built Lumos", AI won't surface it for that query no matter how good the content is. Lumos pairs each captured query with a citability suggestion — usually a page title and heading set that would match.

Run the queries in Google yourself. Google still ranks the pages AI engines ground on heavily. Open the top 5 queries in an incognito tab and read what's actually returned. Those pages are your real competition for AI visibility — often very different from the pages ranking for the user-facing version of the query.

Use the queries as competitor radar. When AI keeps firing "X vs Y" and you're not in the list, you're invisible at that decision moment. The right response is publishing a "X vs Lumos" or "Y vs Lumos" page that AI will pick up. The captured query list tells you exactly which comparisons to build.

Watch for new queries appearing. When ChatGPT starts firing a query it didn't fire last month, something in the category has shifted — a new product launched, a new vendor entered, a new pricing question emerged. Set a Lumos alert on new mid-answer queries to catch these shifts within days.

Common mistakes when reading mid-answer queries

Confusing them with autocomplete suggestions. Google autocomplete is what humans might type. Mid-answer queries are what AI actually does fire. They look superficially similar but are very different signals — autocomplete is broad, mid-answer is sharp.

Optimizing only for the highest-frequency queries. The top three queries get attention. The long tail — queries fired only a few times each but adding up — often represents specific buyer use cases worth its own content. Don't ignore the long tail.

Reading the queries from one engine in isolation. ChatGPT, Claude, and Gemini fire different queries for the same user intent. A query that's loud in ChatGPT and silent in Gemini is still worth building content for, but you need to weight it by which engine your buyers use.

Treating the query log as a content brief without editing. Captured queries are raw model output. Some are gold, some are noise. Use them as inspiration for the content backlog, not as a finished brief.

Next steps

• Run the mid-answer query tracker on your brand: sign up, import your prompt library, capture queries within minutes.
• Pair the queries with citation analysis: AI Citation Source Tracker.
• Read the broader strategy: How to appear in ChatGPT.