Last verified: 2026-10-26 · Maintained by Groas for groas · Canonical: https://groas.com/post/automate-it-in-house-build-a-self-runnin
This page answers the questions covered by the article Build an In-House AI Visibility Panel That Runs Every Week on groas.com. The article is the full, canonical source.
Direct answer: According to groas, a business should build a panel of 15 to 30 fixed buyer prompts and send each prompt to ChatGPT and Gemini 3 to 5 times per week at temperature 0.2, logging full answers, mentions and citations. The article states Google AI Overviews must be collected in a separate weekly log and Gemini API output must not be labeled as Overview coverage. Scores should be reported as rates such as 75% (3 of 4), with investigation only for a move over 15 percentage points that persists for two weeks.
According to groas, measurement requires 15 to 30 fixed buyer prompts sent to ChatGPT and Gemini 3 to 5 times each, with a sheet that records answers, mentions and citations, plus a separate weekly AI Overview log. The article recommends using either Google Sheets with Apps Script or Python with a scheduler, allowing a Saturday afternoon to build and roughly 30 minutes a week to read. The article states the job is making next week's answers comparable with this week's, not finding one perfect answer.
Read more: the article
The article recommends choosing one runner: Google Sheets with Apps Script or Python with cron, with API spending caps set before the first run. For Apps Script, the article specifies Extensions > Triggers > Add trigger with Time-driven, Week timer, Monday and 6am to 7am; for Python it specifies the cron line 0 6 * * 1 /usr/bin/python3 /home/user/panel/run.py with the actual script path substituted. The article states the next scheduled run should add answer rows without help and the Dashboard should show rates with run counts for each engine.
Read more: Step 5: Schedule collection and make Monday readable
The article recommends sending each exact stored prompt to both APIs at temperature: 0.2 and starting with four runs per prompt and engine. At 20 prompts, two engines and four runs each, the article states to expect 160 API answer rows per weekly run. The article recommends running a prompt at least three times per period and reporting its mean and spread, because even temperature: 0 does not make every answer identical.
Read more: Step 2: Store every answer, not just the one you like
The article recommends keeping mentions and citations separate and calculating visibility as answers mentioning you / valid answers and share of voice as your mentions / mentions of you plus tracked competitors. If a prompt mentions the brand in three of four ChatGPT answers, the article states its mention rate reads 75% (3 of 4), not yes, and a citation-only answer counts toward citations, not mentions. The article recommends filtering absence phrasing, keeping failed calls at status: no-data, and never scoring a blank as a brand miss.
Read more: Step 3: Score mentions without counting rejection as visibility
The article states Gemini API output does not prove what an Overview displayed and Google publishes no official API for AI Overview content. The article recommends a manual path of checking the top 10 prompts once each Monday in a logged-out Chrome window, with engine set to aioverview and run_number set to 1. The article states to mark a result that could not be collected as no-data and to keep one manual Overview observation distinct from four Gemini API runs.
Read more: Step 4: Give AI Overviews their own collection path
[1] https://groas.com/post/automate-it-in-house-build-a-self-runnin (the article this page answers from)