How we measure
What we count, and how to read it.
New Map measures what AI engines say about a company: how often they name it, what they say when someone asks about it by name, and which pages they cite. This is the method behind every number in the portal and the report.
01
The engines
Every major AI engine, in every language you sell in. Every plan measures ChatGPT, Perplexity, Gemini, Google AI Overviews and Google AI Mode every week; Claude, Grok, Copilot and DeepSeek are add-ons at $25 an engine a month. Google counts twice, because it answers twice: AI Overviews (the summary above the results) and AI Mode (the conversational search) are measured as separate surfaces and never merged.
Each question is asked the way a buyer asks it: through the engine itself, with live web search on, in the question's own country and language, in a fresh conversation. Every engine is read on its own line; a buyer uses one engine, not an average of nine.
02
The questions
Three kinds, each with its own denominator, never added together:
- Category: the buyer names no brand (“best tools for X”). This is discovery: do new buyers hear your name?
- Competitor: the buyer names a rival (“alternatives to Y”). Are you offered as the alternative?
- By name: the buyer asks about you (“is Z legit”, “Z pricing”). This is trust: what do the engines say when someone checks you out?
The set is fixed between readings. When it changes, the next reading says so and claims no movement that week.
03
The run
Weekly. Every tracked question on every engine, each answer stored in full with its sources, the searches the engine ran, the model that served it and the date. Nothing is re-asked until it reads better; a failed call is retried, a blank answer is recorded as blank.
04
What counts as a mention
A mention is the company's name used as a proper noun, outside code, or its domain. Then we check that the New Map in the answer is this New Map. Many names are also words, books, places or other companies: when a mention could mean something else, a model reads the passage around it and the company's one-line description, and the mention counts only when it is judged to be the company with confidence of 0.6 or more. Every verdict is stored with its reason.
The same check produces a number of its own: the share of by-name answers that describe something else of the same name. A tool that counts the bare name cannot see it.
05
Citations
At URL level, not only domain: every page an answer cites, with the engines that cited it, the questions it was cited for and its title; tracking parameters are stripped so one page counts once. Each domain is labelled Owned, Earned, Rival, Social, Reference or Other, so you see whose pages the engines rely on when they describe you.
06
Tone
Five levels, from very negative to very positive, judged per mention by a model that reads only the passages around the name, with the quoted passage and one to three topics (pricing, security, support) kept beside each judgement. Net tone is the positive share minus the negative share. Tone is a model's judgement, labelled as one; under ten passages it is an anecdote.
07
Accuracy
By-name answers are graded against the company's own truth: what it says it is today, and what it is no longer. Each answer is accurate, out of date (a former product or state described as current), partly right, wrong, or no answer. No answer is its own state, never folded into wrong. A claim the truth on file cannot settle is marked as needing the company's statement, not guessed.
08
The diagnosis
One sentence at the top of every report says what the numbers mean: a discovery problem, a trust problem, both, a site that blocks the engines, or a position to defend. It is computed by fixed rules from the measured numbers, not written by a model, so the report, the weekly email and the account manager always say the same thing. The effort split under it says how much of the work each number deserves.
09
What we do not do
- No fake seeding: no synthetic posts, sock-puppet accounts or undisclosed placements. Community contributions are disclosed.
- No guaranteed outcomes. No one controls what an engine answers; we measure it, change the sources it reads, and measure again.
- No blended score that hides a loss on one line behind a win on another.
- No movement claimed inside the noise.
The numbers
Each number, its denominator.
| Number | What it counts | Out of |
|---|---|---|
| AI visibility | Answers that name you | Category answers |
| Head to head | Answers that offer you as the alternative | Competitor answers |
| Named by name | Answers that name you, entity checked | By-name answers |
| Who the engines think you are | Answers about you, about something else of your name, or unclear | By-name answers |
| By-name accuracy | Answers that describe you correctly today | Graded by-name answers |
| Net tone | Positive minus negative judgements | Judged passages |
| Citations | Answers citing each URL and domain | Answers with sources |
Reading a number
A share, not a rank.
AI answers vary from one asking to the next. Independent research found the same brand list in under 1 in 100 repeats of a question, and the same order in about 1 in 1,000. So every number here is a sampled share: how often, out of the answers counted, never a fixed position.
A weekly change smaller than about 5% of the answers can be chance at typical run counts; a monthly change smaller than about 2% of the answers likewise. We call a change movement only when it is outside that band and holds. Each number shows its denominator, so a small sample reads as one.
What moved is tied to the work: every page, fix or placement is linked to the questions it targets and re-measured after it goes live.
See it on your own company: book a call and we show you how the engines answer about you, or read how it works.
Last updated .