An AI-related figure is useful when it clearly describes what was measured and informs an action. Traffic, citation, recommendation and sales belong to different stages. Before using a percentage in planning, check its source, population, period and method.
This applies to vendor studies and internal results alike. Commercial authorship does not invalidate research; its conclusions should be examinable and its scope should accompany the figure.
Seven checks before using a figure
| Check | What to ask for | Error avoided |
|---|---|---|
| Original source | The document that produced the data | Counting republication as independent research |
| Unit | Clicks, sessions, answers, URLs, mentions or recommendations | Treating citation growth as revenue |
| Denominator | Observation count and inclusion rules | Presenting a subset as all demand |
| Period | Collection dates and comparison window | Confusing publication and execution dates |
| Surface | Product, browser or API, market and language | Treating different experiences as equivalent |
| Repetition | Questions, rounds, failures and common universe | Reading a fluctuation as a trend |
| Application | The decision the data informs | Applying another sector’s average directly to your business |
Consider a claim about “citation share”. A dataset limited to social sources cannot be relabeled as all sources. English-language questions do not become evidence about Brazilian buyers simply because the report is translated.
Separate visibility from traffic
A page can be cited without receiving a visit. A brand can be recommended without a link. A visit attributed to an AI product may never become a commercial opportunity.
Track these measures together using their respective instruments: captured answers for presence and SOV, analytics for sessions and conversion, and CRM data for qualification, opportunities and sales. Connecting them requires an attribution rule and may remain incomplete.
Write “recommended in X of Y eligible answers”, “Z sessions attributed to the channel” or “N qualified opportunities”. Clear units prevent claims extending beyond what was measured.
Auditing brand detection
A word search may count the name already present in the question, a namesake or a sentence denying that the company exists. Classification needs context.
murmur’s August 6, 2026 Brazilian campaign, conducted before this article collection was published, documented this issue: 88 text matches across 800 captures occurred in questions already containing the name; judged presence was zero, including 0 of 712 outside those questions. This is a historical measurement-validation example, not a current performance statement. The questions were in Portuguese.
The practical lesson is to retain answers and evidence, separate branded questions from discovery, and review positive and negative samples. Version the name list so newly discovered competitors do not disappear from analysis.
Reading external research
The academic paper introducing GEO evaluates interventions in a benchmark; SparkToro’s consistency study examines answer variation. They address different questions.
Vercel’s crawler study uses another unit: requests and access behavior. It informs technical decisions, not hiring rates. The review arXiv 2607.14035 helps examine evidence design before generalizing results.
Frequently asked questions
Can I use vendor research?
Yes. Check method, source, scope and limits. Commercial interests should be disclosed, but evidence quality depends on more than who published it.
Does a larger number mean better performance?
Only with comparable units, periods and populations. More mentions can coexist with fewer recommendations; more visits can coexist with lower qualification.
What should a monthly report include?
Baseline, current results, denominators, failures, breakdowns by intent and engine, completed actions and next steps. Connect SOV to business outcomes without conflating their metrics.
How murmur works
murmur works through a SWAS model, combining proprietary measurement software with specialists who diagnose, plan and execute GEO to grow Share of Voice (SOV). Eligible contract models may include an SOV-growth guarantee, with the metric, baseline, period, scope and conditions agreed in writing. This does not guarantee an individual answer, a fixed position or sales.
Author: Mateus Gomes, murmur operator. This article explains our approach and discloses our commercial interest.