Case Study · Data-Driven Prompting

The 9% problem: 400 AI prompts, measured against 182 days of real search demand

An agency delivered 400 well-written questions to steer a brand's AI-search content. We matched them against every query real people typed over six months. The clusters the set covered carried roughly 9% of measured demand. Here is what the measurement showed, and what replaced the list.

  • Roughly 5 million impressions and 5,000 distinct queries analyzed.
  • A coverage claim, never a quality claim: the prompts were well crafted, and the covered demand was real.
  • The replacement: a per-page question engine grounded in the queries each page already receives.

Published 24 July 2026 by Creative Data Engineers. Every number is real and traceable to a dated artifact.

~5M
Search impressions in the 182-day Search Console window
5,000
Distinct real-world queries, clustered by topic
~9%
Share of query impressions covered by the 400-prompt set
14
Ready-to-adapt content drafts delivered from the per-page engine
The short version
  • The agency's 400 prompts were framework-generated: persona times intent times topic, reviewed by taste. Not one prompt was tied to a measured query.
  • Measured against 182 days of Search Console data, the clusters the set covered carried roughly 9% of real query impressions.
  • The other ~91% included prices, availability, technical data, and one technology topic with real measured demand and no page anywhere on the client's properties.
  • The fix was a method change: a per-page fan-out of buyer questions grounded in measured queries produced 14 content drafts, every one traceable to demand, plus a visibility dashboard and page-level structured-data recommendations.
The setup

A press portal quietly became the brand's best AI-search asset

Our client's press portal, built for journalists, had become their strongest performer in AI search: server-rendered, ranking on question searches, and regularly cited by AI assistants in its market.

To steer content for AI visibility, the brand's agency delivered its AI-search strategy: hundreds of slides of analysis and a workbook of 400 prompts, the questions they believed buyers ask AI assistants about the brand. The prompts were well written and thoughtfully organized.

The problem

The question opinion could not settle

Are these the right 400 questions? The agency believed yes. Instinct said parts were missing. Nobody had data. The odd part: this discipline has had the answer for twenty years. SEO always started from what people actually type, and Search Console has always been where that data lives. A prompt set is the new keyword list, and the old rule still applies: check the list against measured demand before it steers a page.

We treated the prompt set as a measurable claim instead of a matter of opinion.

The method

What we did

The result

What the data showed

The replacement

A per-page engine instead of a global list

For each high-demand page: take the queries that actually land on that page, fan them out into the roughly 12 sub-questions an AI assistant breaks the topic into when it builds an answer, scan the page for which questions it already answers, and draft fill-content for the gaps using only facts already published on that page.

Six pages scanned, 14 ready-to-adapt content drafts delivered, alongside a visibility dashboard and technical structured-data recommendations specified down to the JSON-LD @graph nodes and named properties. Every draft traceable to measured demand, every fact traceable to the client's own published content.

The outcome

Buy-in, live on the call

The analysis was presented to the client's team the morning it was finished. The measurement framing worked exactly as designed: no debate about whose questions were better. The team member who had commissioned the original question set summarized the result for a new colleague themselves.

And the strongest proof of buy-in: while the presentation was still running, the client's technical lead connected every European market to Google Search Console live on the call, so the same demand analysis can now run for every market. The drafted content packs went straight to the editorial team.

The lesson

Nobody had to be wrong in the room

Generated question sets are hypotheses. Measured demand sets the priorities.

The covered ~9% was real demand, correctly sensed. From here, every question set, anyone's, is ranked against the measured queries before it steers a single page. That is the repeatable method: instrument prompt sets instead of debating them.

About the numbers

What we publish, and what stays with the client

This case study publishes four figures: roughly 5 million search impressions, 5,000 distinct queries analyzed, roughly 9% coverage by the 400-prompt set, and 14 delivered content drafts from 6 scanned pages, all from a 182-day measurement window. The full query-level breakdown stays where it belongs: with the client.

Every published figure traces to a dated artifact: the Search Console pull, the clustered query workbook, and the per-page content scans.

Take this case study with you

The full write-up as a PDF, formatted to share with your team or your agency. No form, no gate: the download is the link.

FAQ

Common questions

What is a prompt set in AI search visibility work?

A prompt set is the list of questions a brand tests against AI assistants like ChatGPT, Perplexity, or Google AI Overviews to measure and steer its visibility. A framework set is written top-down: persona, intent, and topic matrices, reviewed by taste. A measured set starts the other way around, from the queries real people already type about your pages, and ranks every question by demand.

How do you measure whether AI prompts match real search demand?

Pull the property's Google Search Console queries over a long window (we used 182 days), cluster them by topic, then match each prompt in the set to a cluster. The coverage number is the share of total query impressions carried by the clusters the set touches. In this case: roughly 9%. The same matching also surfaces what the set misses, ranked by impressions, which is where the content plan actually comes from.

Why use Google Search Console data to steer AI-search content?

AI assistants break a buyer question into sub-queries that look a lot like what people type into Google, then pick sources that answer them. Search Console is the largest measured record of real demand your pages already receive: real questions, real impression counts, zero guessing. It will not cover everything AI assistants ask, but it turns content prioritization from opinion into measurement, and it is data you already own.

Can you audit our agency's prompt set?

Yes. Any question set can be ranked against your property's measured queries: yours, your agency's, or one we generate. The audit shows the coverage share, the demand clusters the set misses, and the pages where filling a gap pays off first. It runs as part of our AI Search Visibility Baseline Audit, and the point is never to embarrass anyone. Generated sets are hypotheses; the measurement does the ranking.

Want your prompt set measured against real demand?

The Baseline Audit measures where you stand across the AI surfaces your buyers use, ranks any existing question set against your Search Console demand, and delivers a fix list ordered by impact. From €4,900.