How Panoramma measures
When someone asks an AI assistant about your industry, does it name you or your competitors?
This page explains, in plain language, where the numbers in the report come from, why they can be trusted and where their limits are. It is written so that anyone in your company can read it and decide how much weight to give each figure.
What happens every week
The measurement, step by step
We define the questions with you
They are questions a real buyer would ask an AI before deciding: “which office furniture manufacturers are there in Monterrey?”, “what's the best option to furnish a 50-person office?”. You know your customer better than anyone, so the list is reviewed with you and can be extended or trimmed at any time.
Fixed rule: no question contains your brand. If we asked “is Your Brand's furniture any good?”, the AI would be forced to name you and that data point would be worthless. We want to know whether it names you when nobody has mentioned you first.
We ask four AIs, with the internet switched on
ChatGPT, Claude, Gemini and Perplexity. We query them in their web-search mode, which is how they answer today when a user asks about companies or products: they search, read pages and reply citing where they got it.
A brand can be measured on fewer than four: it is chosen question by question. The columns in your report are always the AIs that brand actually measures, never one more — so if you only see three there, only three are being asked.
The question is asked from your country and city, so the AI answers the way it would answer one of your customers rather than someone in the United States. Gemini is the exception: Google doesn't yet allow setting the location, and we say so instead of pretending otherwise.
Every Monday, every question, on every AI
Measurement is weekly and always under the same conditions: same question, same day, same configuration. That makes each week comparable with the previous one. If an AI fails on Monday, it is retried on Tuesday and Wednesday and the result is stored as part of that same week.
The full answer is stored, with its sources
We don't store a “yes” or a “no”. We store the complete text the AI returned and the list of web pages it cited. Any figure in the report can be traced back to the answer that produced it, and that answer can be read.
We check who shows up
A program searches each answer for your brand (and its variants: with or without accents, with or without the legal suffix, the abbreviations people use) and does exactly the same for each competitor, in the same answer. Nobody reads or interprets it: either the name is there or it isn't.
What each figure means
The numbers in the report
- CoverageMention rate
- Of all the answers the AIs gave to your questions, the percentage in which your brand appears. It is calculated per AI and then averaged across them, so that one AI failing for a week doesn't sink the figure without your visibility having changed.Example: 20 questions × 4 AIs = 80 answers in the week. If you appear in 24, your coverage is around 30%.
- Share of VoiceShare of mentions
- Of all the times the AIs name any brand in your sector (you or your competitors), the percentage that is yours. This is the figure for comparing yourself with the competition: everyone is measured in the same answers, by the same yardstick.
- CitationsSources
- How many times an AI used a page from your website as a source for its answer. It's the most direct signal there is: it isn't that the AI “remembers” you, it's reading you in order to answer.
- TrendWeekly chart
- Coverage week by week, yours and that of the five most-named competitors. It's the most valuable part of the report: a single data point moves around, a two-month trend doesn't lie.
Why you can trust it
What the measurement guarantees
Every figure has an answer behind it. Nothing is estimated or interpreted by hand. If a number surprises you, ask us for the AI's answer and we'll show it to you in full, with date and sources.
The AI doesn't know we're measuring you. The questions don't carry your name and there is no instruction nudging it towards you. What comes out is what a buyer would get.
It's repeated the same way every week. Same question, same day, same configuration, same location. What changes from one week to the next is the AI, not how we measure.
A technical failure never counts as a zero. If an AI doesn't answer, it is retried. If it still doesn't answer, that data point is left blank and stays out of the calculation. A 0% in the report means you were asked about and not named, not that the system went down.
Your competitors are measured by the same yardstick. They are searched for in the same answers, at the same time, with the same method. If the comparison favours you or hurts you, it's because the AI said so.
The history is never retouched. Answers are stored as they came. Last month's report matches the one you were sent last month.
And this matters too
What the measurement is not
AIs don't answer the same way twice. It's their nature: the same question can return a slightly different list one Monday and the next. That's why nothing is judged on a single answer. The reliable reading is the whole set of questions across several weeks, and that's why the report draws no conclusions about a question until it has at least 6 measurements of it.
It's a sample, not a census. We measure the questions agreed with you, not everything people ask AIs about your sector. If the questions are well chosen, the sample is representative; if the business changes, the list needs reviewing. It's an open conversation, not a closed list.
It's each AI's “neutral” answer. We query each company's official connection, with no history or personal account. Upside: it measures what the AI knows about you, not what one particular user has told it. Limit: a person with months of conversations may see a somewhat different answer.
It doesn't measure traffic or sales. It measures whether you are named and whether you are cited. That is what decides today whether a buyer who asks an AI ever finds out you exist. What happens afterwards is measured with your own tools.
We don't claim causes we can't prove. When we publish content for you, the report says what happened afterwards to the questions that content was aimed at. It doesn't say “this article achieved this”: with several AIs changing on their own, that would be a lie. The exception is citations: if an AI cites your article, it is using it as a source, and that is direct.
Reading the report
Three rules
Look at the trend, not the week.
A one-week jump can be AI noise. Three weeks in a row in the same direction is a signal.
Compare yourself with your competition, not with 100%.
Nobody appears in every answer. The useful question is whether you show up more or less than those competing with you for the same customer.
A 0% can mean two things.
If the AI names your competitors and not you, that's a gap with proven demand: that's where the work goes. If the AI names nobody, that question measures nobody's visibility and gets replaced. The report tells the two apart.
Technical sheet
The specifics
- Frequency
- Weekly, anchored to Monday. Retries on Tuesday and Wednesday.
- AIs queried
- ChatGPT, Claude, Gemini and Perplexity, in their web-search mode. Each brand can measure on fewer; its report shows which ones.Each company's general-purpose models, through their official connections: gpt-5-mini, claude-sonnet-5, gemini-3.6-flash and perplexity/sonar.
- Query location
- The client's country and city on ChatGPT and Claude; country only on Perplexity. Gemini doesn't support it yet.
- Instruction to the AI
- None that steers it towards a brand. ChatGPT and Perplexity are asked to answer directly, with specific names and in the language of the question, because without that instruction they ask for clarification instead of answering.Perplexity gets one extra sentence: search the web before answering. Without it, it answers from memory — and then we would not be measuring what a user sees.
- What is stored
- Full text of every answer, cited pages, date and AI. The entire history is kept.
- Brand detection
- Match of the name and its variants in the text, ignoring case and accents. Same for each competitor.
- Coverage
- Mentions over valid answers, per AI; average of the AIs with data. Technical errors are excluded from the denominator.
- Report window
- The last 30 days, unless another is requested.
- Minimum to judge a question
- 6 valid measurements.
- Gemini citations
- Google only reports the source's domain, not the page. They count towards the domain, not a specific article.
Want to check? Pick any figure in the report and we'll show you the exact AI answer that produced it, with its date and the pages it cited.