Valoh
Measurement Protocol
How Valoh measures brand presence in LLM environments.
This page documents what Valoh measures, how it collects each signal, and what it deliberately does not do. It's intended to be readable by a measurement lead, an MRC-track auditor, or a CFO asking what they're paying for.
What we measure
Five signals, per brand, per model, per day.
Valoh tracks five signals about how a brand appears in conversational AI responses: mention rate, rank position, sentiment, competitor citations, and recommendation strength. These are the same signals AEO tools surface — Valoh's contribution is independence, transparent collection, and auditable retention of the underlying transcripts.
Every reported value can be traced back to a logged conversation. Nothing in the report is inferred; everything is observed.
Prompt Construction
Frozen prompt sets, defined by category.
For each measured category, Valoh constructs a prompt set up front. A prompt set contains 40–60 representative consumer queries spanning awareness, comparison, and purchase-intent stages. Once defined, the set is locked for the measurement window.
- 2.1
-
Category-anchored, not brand-anchored.
Prompts ask about the category, not the brand directly. "What's the best running shoe for flat feet?" rather than "Tell me about Brand X." This is what produces unsolicited mentions. - 2.2
-
Locked for the measurement window.
Clients do not see or influence prompt phrasing during a measurement window. Independence is a methodological requirement, not a courtesy. - 2.3
-
Reviewed by category, on a published cadence.
Prompt sets are reviewed every six months for category drift. Revisions are versioned and noted in the report. Comparable measurement requires comparable prompts.
Model Coverage
Four models. Queried separately. Reported separately.
| Model | Provider | Access | Daily queries / brand |
|---|---|---|---|
| ChatGPT | OpenAI · GPT-class flagship | API, default settings | ~50 |
| Gemini | Google · flagship | API, default settings | ~50 |
| Claude | Anthropic · flagship | API, default settings | ~50 |
| Perplexity | Perplexity · default model | API, default settings | ~50 |
Each model is queried independently. We do not average results across models; presence in ChatGPT and presence in Perplexity are different facts and are reported as such.
Sampling & Frequency
Daily refresh. 200+ queries per brand per day. 90-day retention.
Sampling is calibrated to produce a 95% confidence interval on mention rate at ±3% per model per category. Sample size scales upward where category breadth or low base rates demand it.
Signal Definitions
Each signal has a stable definition, a known unit, and a known method.
| Signal | Definition | Unit |
|---|---|---|
| Mention Rate | The percentage of category-anchored prompts in which the brand is named anywhere in the response. | % / day |
| Rank Position | When mentioned in an enumerated list, the average ordinal position of the brand within that list. | avg rank |
| Sentiment | Polarity of the language used to describe the brand within its surrounding sentences, scored on a continuous −1 to +1 scale. | −1 to +1 |
| Competitor Citations | The set of named competitor brands appearing in the same response as the measured brand, with frequency. | named set |
| Recommendation Strength | Tier 1 (actively recommended) → Tier 4 (mentioned but not recommended), classified by a defined rubric over the recommendation phrasing. | tier 1–4 |
Validation
Sentiment and recommendation tier are double-coded.
Mention rate, rank position, and competitor citations are extracted programmatically and have inter-rater agreement of effectively 1.0 by construction. Sentiment and recommendation strength are subjective signals; both are double-coded — once by an automated classifier and once by a human reviewer on a stratified sample. Disagreements are reviewed and resolved.
- 6.1
- Sentiment. 10% stratified sample of responses is human-reviewed daily. Inter-rater agreement is reported per category per quarter.
- 6.2
- Recommendation strength. 10% stratified sample is human-reviewed against a published rubric. Rubric updates are versioned.
- 6.3
- Audit access. Clients may request the raw transcripts behind any reported signal value, at any time, for the 90-day retention window.
What Valoh deliberately does not do
A measurement firm cannot also be an optimization firm.
This is the line we hold.
We do not tell you what to write, what to publish, how to structure your content, or what to ask the model. We tell you what the model said.
- —
- No content recommendations. Valoh does not generate, suggest, or critique your content. We don't offer "remediation roadmaps" or copy edits.
- —
- No prompt optimization for clients. We don't help you write prompts that elicit better responses. We measure responses to prompts that real users would write.
- —
- No client influence on prompt sets. Clients do not see or revise prompts inside their measurement window. Prompt sets are reviewed publicly, not privately.
- —
- No managed-service optimization. We don't sell remediation engagements alongside measurement. You can use any optimization vendor you like; Valoh measures the result.
- —
- No proprietary aggregate scores. Valoh does not publish a "brand visibility score" of its own design. We report the underlying signals; the buyer decides what matters.
Known Limitations
Things this measurement does not, and cannot, capture.
- 8.1
- Personalized model behavior. Valoh queries via API at default settings. User-personalized behavior (memory, custom instructions, prior conversations) is not captured.
- 8.2
- Model drift. Models update without notice. We log the model identifier returned with each response so version-induced shifts can be attributed correctly, but we cannot prevent them.
- 8.3
- Coverage gaps. Models we don't query are not measured. Coverage expands as the category matures; new additions are versioned and noted.
- 8.4
- Causal claims. Valoh measures presence, not the cause of presence. We do not attribute changes in mention rate to specific brand actions.
Early Access