We measure how AI describes and recommends your brand, then test your current message against a new one under the same conditions. You get the evidence, the comparison, and the limits.
A controlled process. An inspectable result.
A dashboard of findings, with the conditions behind every number. Explore this real report excerpt, then inspect the full studies.
Insta 360
01 · Presence
Mention rate94.9%Wilson 95% CI 89.3–97.7
Share of voice
03 · Perception organic sentiment · n=112
Organic sentiment: 80% positive · 19% mixed · 1% negative. Rounded displayed percentages; n=112.
Message Lift two framings, one partnershipSimulatedDirectional
“…collaborating with Leica to enhance image and video quality”
+0.15 to +0.57
“New product with Leica certified lens and Leica designed filters”
−0.31
Two separate studies against different baselines, not a head-to-head test. One framing of the partnership moved expressed intent. The other did not.
Self-initiated study, not client work. Every figure traces to a stored answer. Inspect the audit example →
Reconstructed report interface using the previously published study figures. The audit and the two simulated tests are separate evidence sets, not a single combined result. Insta360 was not a client and did not commission or endorse this work.
Answers mentioning the newest story
The brand’s 360° camera story was familiar. Its newer positioning was absent from this sample. Being known is different from being understood as intended.
Five prompts, five answers each; single-analyst coding. Independent study: Insta360 was not a client, did not commission the work, and did not endorse it.
Inspect the evidenceFind the gap between the story you tell and the answers buyers receive. Test a candidate message before deciding what to change.
Define the category, competitors, markets and buying situations that matter.
Measure how the brand appears across disclosed prompts and engine modes.
Trace findings to stored answers, sample sizes and explicit limitations.
Compare Current and New in shared contexts. These results are simulations.
A later engagement measures the new state. It is a re-audit, not a subscription.
The metric names matter less than the decisions behind them. Each pillar answers one question a brand team can act on.
From the audit report · Pillar 01
We ask open, unbranded category questions and measure whether AI volunteers your brand.
Share of voice across 180 unbranded answers: of every brand AI named, yours took 24%.
From the audit report · Pillar 02
We compare you with named competitors and measure which brand AI recommends.
Asked to choose between you and each rival 60 times, this is how often AI picks you.
From the audit report · Pillar 03
We examine the qualities, sentiment, and objections AI attaches to your brand.
Nine associations were coded from 25 stored answers. These four carry the decisions.
From the audit report · Pillar 04
We check factual claims against the brand fact sheet and trace citations to their sources.
Every checkable claim AI makes about you is matched against your fact sheet.
AI can compare, describe, and recommend brands before a buyer reaches a site, sometimes without sending a click at all. Your analytics start after that moment.
Traditional attribution
Repeated sampling makes variability visible. Stored answers make results traceable. We report uncertainty where the method supports it, and state what each study cannot establish.
Nothing in the scoring is a black box. A stored answer becomes a distribution you can inspect and reproduce, and the anchor method comes from independent peer-reviewed research (arXiv:2510.08338), which Windtunnel productizes.
Stored answer
A free-text response, immutable and timestamped.
Embedding
The answer becomes a vector: a position in meaning-space.
Cosine vs anchors
Similarity against anchor statements: 5 intent levels × 6 wordings = 30 anchors.
Distribution
A probability mass across the 1 to 5 scale, not a single guess.
Labeled result
Response lift compares mean scores on the 1–5 construct, with n and its simulated label. No confidence interval is invented.
SimulatedEach message is supplied as untrusted context to the same brand-neutral shopping situations. Exactly five ranked recommendations are parsed deterministically. Top-five and top-choice rates describe the result; Shortlist lift compares Current and New in percentage points, with scenario-cluster bootstrap intervals.
Inspect both scoring methods ↗Every Windtunnel figure ships in the Glass Box: labeled measured or simulated, gated by sample size, wrapped in its confidence interval where the math supports one, and traceable to the stored raw response behind it.
Every result shows when it was produced, which model route was tested, how many answers were reviewed, and where the conclusion stops.
GEO and AEO describe work intended to improve how brands appear in AI answers. Windtunnel is the measurement and testing layer for that work: it measures AI visibility, citations, competitive ranking, and brand perception, then tests messages. It does not publish content, build links, or promise rankings. Agencies use the audit to set a baseline and the test to show what their work changed.
Message Lift begins with a stored AI answer from the same audit. One Current message and one New message then go through shared contexts, so only the message changes. Buyer-response and AI-recommendation results are both labeled simulated.
LLM answers are probabilistic. One answer is an anecdote. Repeated answers support a rate, with a confidence interval where the math supports one.
No. Windtunnel uses disclosed API/model routes and settings. Those can differ from a personalized consumer chat session. The site makes no consumer-interface equivalence claim.
A one-off chat check is one anecdote. Windtunnel samples every prompt repeatedly, reports rates with their sample size and a confidence interval where the math supports one, and stores each answer so every figure traces to its source.
Answers are sampled from a distribution, so a single run is an anecdote, not a fact about your brand. Windtunnel reports rates over repeated samples, states the sample size, and timestamps every stored answer.
Market context is part of every prompt. Results are labeled by market, route, and date, so a Singapore answer is never pooled with one from another market.