How we calculate
The rules every number follows: per-assistant averaging, the date window, trends and when a change counts as real.
The same answers can be averaged in several ways, and each gives a different number. If you recompute a figure yourself and get a different result from the app, one of the rules below is almost always the reason.
Per chatbot first, then across chatbots
AI visibility, share of voice, average position, sentiment and sources per answer are built in two steps:
- For each AI chatbot, the metric is computed over that chatbot's answers in the selected period — for example, the share of ChatGPT's answers that name you.
- Across chatbots, those values are averaged without weighting into the single number you see.
Each AI chatbot is a separate place a buyer might ask, so one chatbot that answered more often shouldn't drown out the others. A chatbot that answered but never named you counts as 0%. A chatbot with no answers in the period is left out.
Example. ChatGPT names you in 12 of 40 answers (30%) and Gemini in 4 of 38 (10.5%). AI visibility is (30% + 10.5%) ÷ 2 = 20.3%. Pooling all answers would give 16 ÷ 78 = 20.5% — a different number from the same data.
All AI chatbots run at the same time, so their answer counts differ only slightly — for example when an answer fails, or when a chatbot is added to your plan during the period. The difference between the two calculations is therefore usually small.
Different versions of the same AI chatbot count as one chatbot.
A topic is not the average of its prompts
A topic row pools each chatbot's answers across all prompts in the topic, then averages across chatbots. It is not the mean of the prompt rows beneath it, because prompts in a topic run a different number of times.
Example. Two prompts in one topic: the first is named in 50% of 10 answers, the second in 10% of 90 answers. The topic figure is 14 ÷ 100 = 14%, not (50% + 10%) ÷ 2 = 30%.
The brand-wide headline on each page is built the same way over all counted prompts, and it matches the "Brand-wide average" row in the prompts table.
What counts toward the headline
- Only active prompts and search queries. Disabling one removes it — including its past answers — from the headline figures. Enabling it again brings its history back.
- Brand-profile prompts are excluded from AI visibility, share of voice and average position. They ask about you by name, so they would always report close to 100%. Their answers feed the Brand Profile instead.
- Prompts without a topic are included, shown as "Uncategorized".
- Answers you corrected as not mentioning your brand stop counting as mentions.
- A search run counts only when the engine returned results. Failed or empty searches are left out of search metrics.
The date window
Every page shows the past three weeks (21 days) by default, and you can pick any other range. An answer belongs to the day it was collected, and both the first and the last day of the range are included.
A shorter window reacts faster to change but rests on fewer answers, so single answers move the percentages more. A longer window is steadier but slower to show change.
Trends
The trend beside a metric compares the selected period with the 21 days before it.
The Brand Profile compares the selected period with the immediately preceding period of the same length. A trend is shown only when the earlier period has at least 5 answers across at least 3 days, and at least a quarter as many answers as the current period.
"What changed" on the Overview page
The Overview page compares the last 7 days with the 7 days before. A change is reported only when it passes three gates:
- Both weeks are comparable. Each week has at least 5 observations, and neither has less than half as many as the other. Otherwise the page says the weeks can't be compared, and why.
- The change is bigger than normal noise. For shares, such as visibility, a statistical test checks that the difference is larger than chance would produce at that number of answers (90% confidence). For averages, such as share of voice or sentiment, the value must change by at least 20% in relative terms and at least 1 point. For position, by at least 0.3 places.
- Then it is labelled by size. A change of at least 50% in relative terms, at least 10 points, or a fall to zero is "notable". Anything smaller is "slight".
At most three changes are shown, largest first, each with the answer counts of both weeks. When nothing passes, the page says that nothing moved beyond normal week-to-week noise.
Minimum data for a results summary
Pages show a results summary only when there is enough data behind it. This applies only to the summaries — on the Overview page and in the results summary at the top of a page. In the prompts table and on a prompt's detail page you see the numbers from the first answers on.
| Page | Needs at least |
|---|---|
| AI Visibility | 10 answers (and 10 per topic before a topic is judged) |
| Search Visibility | 20 search runs (and 20 per topic) |
| Brand Profile | 5 answers |
| Overview — where you stand | 10 AI answers or 20 search runs, and 2 competitors with data |
Below these thresholds the summary says it is still collecting.
Why your own calculation may differ
If a number you work out yourself doesn't match the app, check these first:
- You pooled all answers instead of computing per assistant and then averaging.
- You averaged prompt rows to get a topic figure.
- You included paused prompts or brand-profile prompts.
- Your date range differs by a day at either end.
- You counted answers where the app counts mentions — one answer can mention your brand more than once, which matters for sentiment.