Somebody left fingerprints in our Search Console.
Sixteen months of data. 114,814 queries and 2.67 million impressions, almost all typed by people. This is the story of the 3,476 that were not: repeated prompt templates that revealed which buyers our competitors simulate, which rivals they benchmark, which markets they run in, and what they count as a win.
Read one row like a research brief.
Every choice in the sentence is a research decision: the buyer, the market, the rival set, the evidence source, the shape of the answer.
One prompt, taken apart
Each highlighted piece reveals a decision in the tracking setup
... what do users say about
as a vp of marketingThe role being simulated.
at a mid-size company in north america, operating in the construction sectorCompany size, market and industry become separate test dimensions.
hubspot aeo vs profound vs semrush ai toolkitThe competitor set the system has chosen to measure.
for ai visibility tracking and executive-level reportingTwo product jobs used as evaluation criteria.
on g2?A source constraint.
you must provide a forced ranking from best to worst.The answer must be decisive. It does not say which brand should win.
It did not miss a day.
Three impressions in 209 days, then 217 consecutive days from 13 January 2026, weekends included, with 99.3% of impressions from desktop.
Seven months of nothing, then every single day
One prompt family, daily impressions, weekends included
18 Jun 2025 to 17 Aug 2026Five languages, one country.
The German run served 90.2% of its impressions in the United States. Query language and execution market are separate settings.
One template, five languages
Forced-ranking impressions by query language
A prompt bank is a strategy document.
Brand mentions inside the forced-ranking family map the market somebody believes they are in.
The market they chose to measure
Competitor mentions inside forced-ranking prompts
Now read yours.
Somebody has been scoring us on voice search and local SEO. We do neither. Seven repeated prompts compared Trakkr with Profound and Otterly on categories that do not describe what we do. 721 impressions, no clicks, and a confident weekly number against the wrong product. Nothing in a score tells you the brief behind it is wrong. You find that out by reading the prompts.
Check your own Search Console
Four narrow filters, each aimed at a specific fingerprint
forced ranking|ranking forzado|erzwungene rangliste|classement forcé|ranking forçado^(as an? |i am an? |about me:)|my main motivations:|my main pain points:^context: location: .+\(not for language\)\[person_name\]|\[company_name\]|_(states|kingdom|america|plus)- 01Open Search Console, go to Performance, then Search results.
- 02Add a filter on Query, switch the match type to Custom (regex), and paste the line above.
- 03Use the longest available date range. Compare cadence, device, country and clicks before classifying anything.
What the Search Console data showed
- Sixteen months of Google Search Console data for trakkr.ai: 114,814 distinct queries and 2.67 million impressions, almost all typed by people.
- One repeated row stood out: a prompt that simulates a marketing buyer, names three competitors, constrains the evidence to G2 and demands a forced ranking from best to worst. 1,237 impressions, zero clicks.
- The row is a template. 3,476 distinct structured queries follow the same design, built from role, industry, company-size and region slots.
- One persona family produced three impressions in 209 days, then ran for 217 consecutive days from 13 January 2026, weekends included.
- The structured set earned 61,966 impressions and zero clicks, against roughly 147 expected at the ordinary click-through rate. 99.3% of impressions came from desktop, and the same template ran in five languages, with the German run serving 90.2% of its impressions in the United States.
- Brand mentions inside the forced-ranking family map the competitor set somebody chose to measure: HubSpot, Profound, Otterly, Scrunch, Rankscale, Semrush, Rankshift, AthenaHQ, Ahrefs and Peec AI.
- Several patterns match public product designs, which identifies a design, not an operator. Readers can run the same regex filters on their own property with the free AI query scanner.
Run this analysis on your own Search Console.
We ran these filters across ten authorized properties, and eight showed the same syntax, the same cadence and the same silence where the clicks should be. Scan sixteen months of your own queries for the evidence that separates machines from people.
Method, sources and every figure: the full written study.