Lesson 3 · 18 min
ChatGPT, Claude, Gemini, and Perplexity have no dashboard for the site owner. The honest way to measure is a repeated sample: the same questions, in the same AIs, every month, with a date. Because answers change with every run, you measure an appearance rate, not a snapshot.
Start the lesson
Goal: build a battery of 20 frozen questions and run it by hand on 4 assistants, logging everything in a dated spreadsheet.
Click each block.
A repeated, dated sample: that's what makes the number comparable month to month.
Tap each step.
List the real questions with people who know the audience: the non-branded queries from Search Console, customer questions, the grounding queries from Bing (lesson 2).
Write the way a person would write to an AI, in a full sentence.
Save the list as bateria-v1.md with the date and the rule "this battery does not change." A new question goes into a v2.
For each assistant, paste the question, read the answer and the sources, and fill in the row. Repeat 3 times: the answer changes, and that variation is what the rate absorbs.
With 20 questions, 4 assistants, and 3 runs, that's 240 rows. You can split it across days; what matters is recording the date.
Running the questions on a local model (Ollama). It doesn't have the search, the index, or the weights of the real AIs: it doesn't measure what they answer.
% of appearance in the neutral questions per assistant, the 5 most-cited sources in your place, and the factual errors about you. Compare only with the same battery and the same assistants.
A new assistant becomes a new column; never swap one for another.
Swap the topic for yours. The example questions come from the real INEMA battery v1 (frozen on 07/10/2026).
# Battery v1 — FROZEN (dd/mm/yyyy) Rule: this battery does not change. A new question becomes v2, measured in parallel. 16 neutral + 4 branded. 3 runs per question, per assistant. | id | type | question | |----|------|----------| | P01 | neutral | <who does/teaches [your topic] in Brazil?> | | P02 | neutral | <where can I learn/buy [your topic] in practice, in Portuguese?> | | P03 | neutral | <compare options for [your service] in [your city/country].> | | ... | neutral | ... (up to 16 neutral) | | P17 | brand | <what is [your brand]?> | | P18 | brand | <who is [founder]?> | | P19 | brand | <is [your brand] good for [goal]?> | | P20 | brand | <what is [your domain]?> |
| P01 | neutral | Quem ensina agentes de IA no Brasil? | | P03 | neutral | Qual curso prático de Claude Code em português você recomenda? | | P05 | brand | O que é INEMA? | | P06 | brand | O que é INEMA.club? | | P07 | brand | Quem é Nei Maldaner? | | P18 | neutral | Onde encontro cursos gratuitos de IA prática em português? | | P19 | neutral | Qual empresa ajuda a estruturar o conhecimento de um negócio para ser recomendado por IAs (AEO/GEO)? |
data;assistente;pergunta_id;execucao;marca_citada;citados_no_lugar;fontes;erros;resumo 2026-11-03;perplexity;P01;1;nao;<concorrente A|concorrente B>;<site1|site2>;;<resposta em 1 linha>
A single round on 07/10/2026, battery v1 with 20 questions: INEMA appeared in 0 of 48 neutral answers on Claude and 0 of 4 on Gemini (52 neutral measurements). ChatGPT and Perplexity were not measured.
"O que é INEMA?" got 4 of 4 answers about a different organization with the same name. That's the topic of lesson 4.
The July round used an API, through a key authorized at that time. This course does not repeat that path: by hand, through the account you already use, no robot.
An answer from an API can differ from what a person sees in the app, and a paid "share of voice" tool is also a sample of prompts, not an official score.
E2 takes a few hours; split it across days if you need to.
Write the 20 questions (16 neutral, 4 branded) for your topic, using the non-branded queries from Search Console and the grounding queries from lesson 2. Save it with the date and the freeze rule.
Done when: bateria-v1.md exists, with a date, the "does not change" rule, 16 neutral questions without the brand, and 4 branded ones.
Run the battery on 4 assistants, 3 times each, by hand, and fill in medicoes.csv. At the end, calculate the % of appearance in the neutral questions per assistant and list the 5 most-cited sources in your place.
Done when: medicoes.csv has the rows from the round with date and assistant, and a summary with the % per assistant and the 5 most-cited sources.
The explanation appears after your answer.
Question 1 of 4
Why run each question 3 times?skill
Question 2 of 4
In the second month you want to swap 5 questions. What do you do?
Question 3 of 4
Does running the battery on a local model (Ollama) work for measuring citations?
Question 4 of 4
Which question is neutral?
Choose an answer. The explanation appears after you try.