Methodology
How Basiro measures what AI says about a brand
Every number Basiro shows comes with how many answers it rests on, how sure it is, and the answers themselves. This page explains the method; inside the product every figure links to the answers behind it.
The method
- Questions: the questions your customers ask, per market, customer type and language. Only these questions are sent to the AI assistants, never customer data.
- AI assistants: each question goes to the assistants people use (ChatGPT, Google AI Overviews and AI Mode, Gemini, Claude, Perplexity, Copilot and others) with web search on, the way a customer asks.
- Samples: assistants answer differently each time, so each question is asked once or several times on each assistant, depending on your plan (the count is shown above). A share is the answers that name you divided by all real answers.
- Questions back: a reply that asks the user something, or is empty, is not an answer. We keep and show it, but never count it as "not named".
- Reading: whether you are named, your rank, the tone, the rivals and the claims are read by fixed rules, not by another AI, so the same answer always reads the same. A different business with your name is left out.
- Intervals: every share has a 95% interval (Wilson). Repeated answers to one question are related, so the interval is widened by how much the share differs between questions: at worst, each question counts as one answer.
How many answers a number needs
For a share near 50%, n answers give a figure good to about this many points either way:
- 30 answers: ± 18 points
- 100 answers: ± 10 points
- 400 answers: ± 5 points
- 1000 answers: ± 3 points
19 of 20 and 190 of 200 are both 95%. Only the second is evidence: its interval is 91%–97%, the first one's is 76%–99%.
API and screen
An API answer and what a person sees on the assistant's own screen can differ: the screen adds its own defaults, location and product cards. Where we capture both, we show the gap.
How accurately we read answers
We test the reading on example answers whose right reading is marked by hand: the same reading the product uses runs on them and is compared. This runs on every release; a release that reads worse than the last one is stopped.
So far 0 of 46 test answers are labelled by people. The rest are hand-written examples that check the reading works; they are not yet proof for real answers. Our target is at least 100 labelled real answers per language.
| What we read | Arabic | Turkish | English |
|---|---|---|---|
| Brand named | 100%80.6%–100% · n = 16 | 100%79.6%–100% · n = 15 | 100%79.6%–100% · n = 15 |
| Same-name business (F1) | 100%n = 1 | 100%n = 1 | 100%n = 1 |
| Rank | 100%74.1%–100% · n = 11 | 100%74.1%–100% · n = 11 | 100%74.1%–100% · n = 11 |
| Tone | 100%74.1%–100% · n = 11 | 100%74.1%–100% · n = 11 | 100%74.1%–100% · n = 11 |
| Rivals named (F1) | 100%n = 4 | 100%n = 4 | 100%n = 7 |
| Claims found (F1) | 100%n = 5 | 100%n = 7 | 100%n = 8 |
| Claim verdicts | 100%56.6%–100% · n = 5 | 100%64.6%–100% · n = 7 | 100%67.6%–100% · n = 8 |
| Answer or question back | 100%80.6%–100% · n = 16 | 100%79.6%–100% · n = 15 | 100%79.6%–100% · n = 15 |
F1 combines what was found and what was missed; n is how many items the labels hold. Shares carry a 95% interval.
Measurement health
Each brand has a measurement health screen: which AI assistant answered in the last 24 hours, which stopped because of credit, a daily quota or an error, and when each last answered. An empty screen always says why.