Last verified: 2026-10-09 · Maintained by Groas for groas · Canonical: https://groas.com/post/build-a-per-prompt-ai-search-tracker-in
This page answers the questions covered by the article Build Your Own Per-Prompt AI Search Tracker for ChatGPT, Gemini, and AI Overviews on groas.com. The article is the full, canonical source.
Direct answer: According to groas, a business should replace a blended visibility score with a per-prompt, per-engine log built from 25 fixed buyer questions run three times per engine for 225 observations per cycle. The article describes three linked tabs for raw runs, a per-prompt rollup kept separate by engine, and a dated change log tagged to target prompts. It recommends fixed weekly collection and logging each site edit within 24 hours to check the targeted prompts for four weeks before and four weeks after.
The article recommends building a per-prompt, per-engine log instead of relying on a blended score, using three linked tabs for raw runs, a prompt rollup, and a change log. Each raw run records timestamp, model version, brand mention, brand citation, recommendation, competitors, and exact cited URLs, while the rollup keeps mention and citation counts from three runs per prompt separate for each engine. The change log records site edits, title revisions, and structured data updates tagged to the prompts each change targets.
Read more: What you will have, and what you need
The article recommends a fixed panel of 25 buyer questions in a prompt bank rather than a dashboard score, built from the last 90 days of support tickets, call records, and search query reports with comparison and constraint language. According to groas, each question should keep frozen wording for at least 30 days, with at least 20 unbranded questions and the brand name in no more than five prompts. The panel is classified by intent and category to show where the brand appeared and where a competitor appeared instead.
Read more: Step 1: Freeze 25 questions buyers actually ask
The article recommends deploying an automation runner on a fixed weekly schedule, with Tuesday 06:00 UTC given as a workable setting, to read the 25-question bank and collect three runs per prompt per engine for 225 fresh observations. The runner is set to pause three seconds between repeated calls and to retry failed requests three times with exponential backoff, logging failures rather than recording zeros. The article advises a ten-minute weekly spot-check of records where the brand was mentioned but rank data is missing or the meaning is negative.
Read more: Step 4: Put the collection on a weekly schedule
The article states that a brand named in passing, a brand linked with an exact URL, and a brand recommended should be scored as three different events stored as 0 or 1 per observation. It requires one row per observation for a complete cycle of exactly 225 rows, preserving absent answers to keep a denominator and keeping the raw answer for audit. For the rollup, mention and citation frequency are calculated from the three rows for the same prompt on the same engine, keeping results without an AI answer visible and separate.
Read more: Step 3: Keep mentions, citations, and recommendations apart
The article recommends creating a change log before revising pages and logging each edit within 24 hours with its date, URL, type, target prompt IDs, and hypothesis. It advises comparing the targeted prompts for four weeks before and four weeks after the change, engine by engine, while also watching competitor citations and unrelated prompts. The article states the log makes a before-and-after check possible without turning that check into proof of causation.
Read more: Step 5: Tag each site change to the prompts it targets
These articles answer the same questions:
[1] https://groas.com/post/build-a-per-prompt-ai-search-tracker-in (the article this page answers from)