Methodology

Prompts: what your customers actually ask

Measurement starts with topics: the products or services your customers look for, such as “domain registration” or “escape room in Denver”. Each topic gets two prompts: a full sentence of the kind people type into a chat, and a short keyword-style query of the kind people type into a search engine.

Your brand name never appears in a prompt. Asking “Is company X any good?” hands the model the answer. It measures whether the model knows your name, not whether it would recommend you to someone who has never heard of you.

• A city appears in a prompt only for local businesses that customers have to visit in person.
• The prompt set stays fixed, so results from different days can be compared.
• You can edit the prompts. If you connect Google Search Console, you also get suggestions based on the searches your site already shows up for.

See in the glossary: Prompt

Collecting answers: the way a user sees them

We ask through the regular chat interface in a browser, not through an API. An API answer can differ from the one in the chat window: a different model, a different web search, no sources. We measure what a person actually sees.

Prompts are sent from an IP address in Poland. Web search inside AI assistants, and Google AI Overviews most of all, picks the market from the IP address, so for broad questions without a location the answers reflect what a user in Poland would see. Prompts that name a city or country are answered for that place. Every prompt starts a new conversation, so earlier answers do not influence the next ones.

• Engines: ChatGPT, Gemini, Perplexity, Microsoft Copilot and Google AI Overviews. The free trial covers the first three.
• Every prompt goes to every engine once a day.
• We save the full answer, its sources (links in the text and in the sources panel) and the search queries the model sent to a search engine, when the engine reveals them (ChatGPT and Perplexity).

See in the glossary: AI brand monitoring, Citation

Analysis: one rule for your brand and your competitors

Each answer is read by a language model. Its job is to list every company named in the answer and, for each one, record its position in the list, how it is presented and the tone used. Your brand is scored exactly like your competitors, from the same list. If your brand got separate, more detailed questions, its score would be inflated compared with everyone else’s.

The model can make mistakes, so code checks its output. A company counts only if its name actually appears in the text. Any tone other than “listed without an opinion” needs a verbatim quote from the answer. If the quote is not in the answer, the rating is dropped.

Different spellings of the same company (“Otterly”, “Otterly AI”, “Otterly.ai”) are merged into one entry. Otherwise one company would have its mentions split across several rows.

See in the glossary: Sentiment, LLM

Visibility Score

Each answer earns from 0 to 10 points. If your brand is not in the answer, it scores 0. If it is, the points come from three parts:

• Position: 4 points for first place, 3 for second, 2 for third, 1 for anything lower.
• Presentation: 3 points when your brand is recommended as the best choice, 2 when it is one of several options, 1 for a passing mention.
• Tone: 3 points for positive, 2 for neutral, 1 for negative.

See in the glossary: Visibility Score, Mention Rate

A five-sample window

The same model answers the same question a little differently every day, so a single answer is a roll of the dice, not a measurement. That is why the score for each prompt and engine pair is the average of the last five answers, none older than 14 days. Each day adds one sample, so the window fills up after five days. Until then, the score is marked as preliminary.

Prompt scores are averaged within each engine, and the engines are combined with weights: ChatGPT 35%, Gemini 22%, Perplexity 17%, Google AI Overviews 14%, Copilot 12%. If an engine has no data, the weights of the others are scaled up proportionally.

Share of Voice

Share of Voice is your share of mentions: the number of answers that mention your brand compared with all mentions of your brand and of the companies on your competitor list, from the same window. Only companies on your competitor list count, not every website named in an answer. A review site or a seller of add-on services would push your share down for no good reason. If you do not consider a company a competitor, you can exclude it with one click, and it drops out of this calculation too.

See in the glossary: Share of Voice

What does not count as zero

Only a valid answer that does not mention your brand scores zero. A failed fetch, a search engine block or a prompt for which Google showed no AI Overview are left out of the score entirely. They are missing data, not proof that AI ignored your brand. If we counted them as zeros, a technical problem on our side would look like a drop in your visibility.

What the measurement does not tell you

Visibility Score summarizes samples. It is not the share of users who saw your brand, and it is not a sales forecast. We ask from accounts with no chat history, while the answer a real person gets may be personalized. The measurement shows direction and how you compare with competitors. To judge the business impact, look at it alongside your website traffic and customer inquiries.