How to Measure Your Brand's AI Visibility
Short answer: AI visibility is not a Google ranking. Measure how often your brand is named correctly in machine answers (ChatGPT, Gemini, AI Overview, Perplexity, and peers) for questions your buyers actually ask — plus position, sentiment, and which sources get cited. Without a frozen question set and repeated reads, you only have anecdotes.
Definition: AI visibility
AI visibility is how often a brand is named, or has a page cited by URL, in AI engine answers to a fixed set of prompts that leave the brand out. SOLEDAD, White Wood’s AI-monitoring engine, keeps the two apart: for whitewood.ai, 10 prompts on six engines from 18 September to 18 October 2026 gave 2 named cells and 8 cited cells out of 60.
A cell is one prompt on one engine over the measurement period. Each cell lands in one of four states, and the four are never added up into a single score:
- Absent: the brand is not named and none of its pages is cited.
- Named: the brand appears in the answer’s brand list at least once in the period.
- Cited by URL: a page on the brand’s domain appears in the answer’s sources at least once.
- Top-3: the brand is named with a median position of 3 or better, counted over the answers that name it.
Prompts that contain the brand or a sub-brand still belong in your prompt set, as advised in the method section below, but are left out of this scorecard. They test whether an engine recognises you, not whether you turn up for a buyer who doesn’t know your name yet.
What to measure (not vanity)
- Share of voice in answers — on prompt X, how often you appear vs competitors
- Position / order of mention — first or buried
- Accuracy — name, category, claims extracted: right or wrong
- Cited sources — which pages the model leans on (owned, media, third-party lists)
- Gaps — important prompts where you do not appear at all
Dual Legibility Tension (White Wood): marketing teams often read “we’re viral on LinkedIn”; machines read whether a citable page exists for an unbranded question.
A repeatable method
Freeze a question set. 20–40 prompts in the buyer’s language (not internal jargon). Mix: category-without-brand, comparisons, “services in Jakarta,” and branded asks.
Run across engines. One engine is a sample. Six engines show a pattern. Date every capture; answers drift.
Read sources, not just logos. If a competitor wins because of a third-party list or their FAQ page, that tells you what to build — not which ad channel to buy more of.
Report honest numbers. Don’t average in prompts you already “own.” Separate prompts where work can still change the outcome.
The SOLEDAD scorecard, using our own numbers
This scorecard is whitewood.ai’s real data, not a mock-up. We use our own brand as the example because it is the only brand whose data we can publish. It covers 18 September to 18 October 2026: 10 unbranded prompts, six engines, 1,676 answers read.
| Engine | Answers read | Absent (of 10 prompts) | Named | Cited by URL | Top-3 |
|---|---|---|---|---|---|
| ChatGPT | 298 | 9 | 1 | 0 | 1 |
| Gemini | 296 | 9 | 0 | 1 | 0 |
| Perplexity | 283 | 7 | 0 | 3 | 0 |
| Copilot | 269 | 10 | 0 | 0 | 0 |
| Google AI Overview | 247 | 7 | 1 | 3 | 0 |
| Google AI Mode | 283 | 9 | 0 | 1 | 0 |
| Total (60 cells) | 1,676 | 51 | 2 | 8 | 1 |
Three things stand out.
Being cited and being named rarely happen together. Of the 8 cells that cite a whitewood.ai page, only one also names White Wood: Google AI Overview on “cara agar brand muncul di jawaban AI”. Our only Top-3 cell has no link at all. ChatGPT put us at position 2 on a podcast recommendation prompt, in 2 of 26 answers.
The cited page is often not the one you would pick. The whitewood.ai home page was cited 21 times, all by Google AI Overview. /id/agensi-ai-marketing-jakarta/ was cited 8 times, by Gemini and Google AI Mode. This page is not on the list yet: on “cara mengukur visibilitas AI”, whitewood.ai was absent on all six engines.
“Named” can mean almost never. That ChatGPT cell counts as named because we appeared in 2 of 26 answers. That is why we always print the answer count next to the status instead of a tick.
To read your own scorecard, start with the absent column. A prompt absent on every engine is a gap. A prompt cited on one engine only is thin: a single change on that engine can take it away.
What this cannot tell you. The scorecard records whether a brand showed up, not why. The export holds no answer text, so we cannot see which part of a page an engine used, and tier and location per engine are not recorded. Sentiment, listed above as worth watching, is not in it. The numbers cover one brand over 30 days; yours may look very different.
Declared interest: White Wood sells this work and measures it with SOLEDAD, our own AI-monitoring engine.
Sources: SOLEDAD baseline for whitewood.ai, 10 unbranded prompts on ChatGPT, Gemini, Perplexity, Copilot, Google AI Overview and Google AI Mode, 18 September to 18 October 2026 (1,676 answers).
Quick checklist
- Written prompt list (version + date)
- Baseline before the content push
- Columns: appears / not · position · quote · source URL
- A decision: which page gets fixed this week
Straight answers
Do social impressions equal AI visibility?
No. Social can move humans; machines cite pages they can retrieve again.
How often should you measure?
Weekly is enough to see direction; monthly is too slow while you’re fixing foundations.
Do you need paid tools?
Tools speed it up. What’s mandatory: a question set, reading discipline, and page decisions — not a dashboard with no action.