The keyword list you paid for does not describe the queries AI search actually runs against your pages. Gemini generates its own grounding queries before it answers a user, and those machine-written searches look nothing like the data in your SEO tools.

AI search engines write their own long queries to find your content

On September 1, Dr. Pete Meyers published a dataset of 5,333 grounding queries that Google Gemini ran across 1,000 subtopics. Grounding is Google running searches for itself to surface content it might otherwise miss. Meyers calls these queries a machine’s interpretation of the searcher’s intent, noting we still need to add the human perspective.

His data shows these queries run anywhere from 3 to 17 words long, averaging 6.56 words. Over half the dataset was 6 or 7 words. Fully 33% of them mentioned a year, and the system sometimes stacked two years together like “2025 2026”.

Meyers pulled examples like “standard types of auto insurance coverage”, “must have baking tools equipment beginner”, and “extended stay hotel amenities vs traditional hotels”. The systems also build massive term piles. One query was laptop gpu vs desktop gpu performance difference 2025 2026 rtx 40 50 series. The longest ones used search operators, chaining things like "best indie games of 2025" OR "biggest indie game hits of 2025" OR "popular indie games 2026".

Meyers notes his team nudged the system instructions to produce 4 to 6 queries per prompt. He admits they cheated a little bit since default Gemini 3.5 Flash usually returns one or two, and often none. He cautions that AI Mode and actual query fan-out might behave very differently. Generating these queries through the Vertex API currently costs $14 per 1,000.

You will lose if you build a new page for every long-tail variation

A common mistake is trying to spin up hundreds of hyper-specific pages to catch every possible machine-generated query. Google’s own guide to optimizing for generative AI features explicitly tells you not to rewrite content just for AI systems. The guide states you do not have to worry that you don’t have enough long-tail keywords or haven’t captured every variation of how someone might seek your content. It also rules out chunking your content into tiny pieces.

Gianluca Fiorelli pointed this out at the Search Evolution Summit this week when he noted that a million-row keyword export is a terrible way to start a research process. You have to start with the actual entities and taxonomy of your business. As I noted when discussing why AI visibility case studies are mostly superstition, the exact query fan-out changes constantly under the hood. You cannot build a reliable content plan by chasing these temporary variations.

One comprehensive page must answer the entire question space

The solution is to build a single page that answers the whole question space around a topic. The queries are long and specific, so a page that covers only its head term leaves most of that question space uncovered.

This is why every content brief I write names exactly five conversational prompts one article should answer. I design the page to satisfy the whole cluster of intent rather than assigning five separate articles to five slightly different questions. In practice, the page must pass a simple test: a skeptical person who searched gets their intent confirmed in two seconds, their objections answered in ten, and no reason to go back to the results and search again.