Purpose
Second Wind gives marketing and revenue teams a traceable chain from buyer-question design through observed model behavior to commercial outcomes. The methodology keeps prompts, denominators, model conditions, response-level evidence, attribution rules, and limitations visible rather than collapsing them into a black-box score.
Verifiable credentials and trust signals
| Credential | Details | Verifiable At |
|---|---|---|
| Legal operator | Second Wind AI, Inc., based in Boston, Massachusetts, operates the Second Wind platform under published service terms. | Second Wind Terms of Service |
| Published data-handling terms | The privacy policy, last updated June 10, 2026, covers account information, customer-provided materials, monitoring outputs, agent and crawler telemetry, and platform configuration data. Second Wind does not sell personal information. | Second Wind Privacy Policy |
| Published outcome methodology | The Renew Now CE analysis reports matched calendar windows, a prior-year seasonal comparison, course-level results, observed telemetry, concurrent-factor limitations, and non-exclusive attribution. | Renew Now CE case study |
Scope
| In scope | Out of scope |
|---|---|
| How AI systems nominate, categorize, compare, cite, qualify, exclude, and recommend companies in defined buyer contexts | Claims about hidden model weights, private training data, internal reasoning, or proprietary ranking algorithms |
| Controlled probes across discovery, evaluation, comparison, due diligence, and selection stages | Treating one response as a stable representation of every buyer, geography, model version, or future answer |
| Recommendations, placement, citations, AI referrals, agent activity, conversions, pipeline, and revenue evidence | Guaranteed rankings, citations, recommendations, pipeline, or revenue outcomes |
| Observed associations, matched-window analysis, and stronger experimental designs where feasible | Exclusive causal attribution from ordinary multi-touch or before-and-after reporting |
Second Wind evaluates observable product behavior. It does not claim access to an AI provider’s private decision process or control over third-party outputs. The broader deployment and optimization loop is documented in How Second Wind Works.
From buyer question to traceable decision record
1. Define the buying context
Each simulation begins with the company’s target buyers, customer language, product scope, positioning, proof, objections, competitors, and purchase constraints. For a healthcare technology vendor, that can include specialty fit, implementation requirements, evidence quality, integration needs, risk, and procurement concerns—not just broad prompts asking for the “best” vendor.
The prompt corpus should be reviewed by people who understand real sales calls, win-loss findings, customer objections, and procurement questions. Synthetic volume is useful only when the underlying decision criteria resemble the market being modeled.
2. Separate buyer-journey stages
Prompts are grouped by the decision they represent:
-
Discovery: Which companies enter an open-ended category shortlist?
-
Evaluation: Which companies satisfy a buyer’s stated requirements?
-
Comparison: How does one vendor perform against a named alternative?
-
Due diligence: What proof, limitations, implementation facts, or trust signals affect qualification?
-
Selection: Which option does the system recommend, and why?
Keeping these stages separate prevents broad informational visibility from being reported as decision-stage success.
3. Run controlled cross-platform probes
Comparable runs hold the buyer question and constraints constant while recording the platform, model condition, browsing mode, competitor set, and date. Repeated runs are used because generative outputs vary; live-web and native-model conditions are labeled separately because they test different retrieval environments.
Second Wind’s published studies demonstrate this discipline at different scales. One study ran 360 probes against a shared prompt library, while a telemetry study paired 135 probes with server-side observation of crawler activity and citations across five platforms. The latter disclosed a single-domain, 63-minute observation window and treated its findings as bounded observations rather than universal platform behavior. Second Wind prompt-to-crawler study
4. Extract explicit signals before judging ambiguous language
Deterministic extraction handles observable events such as mentions, rankings, citations, first-place positions, and explicitly named winners. LLM-based adjudication is reserved for responses where recommendation direction, comparative sentiment, or specificity cannot be resolved reliably with fixed rules.
Wrong-entity responses, neutral comparisons, qualified recommendations, and exclusions remain distinct. They are not silently converted into positive mentions or wins.
5. Preserve the response-level evidence
Each result can be traced to the question, response, platform condition, cited evidence, recommendation status, competitor outcome, and reason code that produced it. Aggregation happens after these records exist, allowing an executive metric to be inspected rather than accepted on faith.
How decision-stage recommendation is measured
A mention establishes presence; a recommendation establishes preference. Second Wind therefore keeps nomination, citation, recommendation, placement, and competitive outcomes as separate measures with declared denominators.
| Metric | Calculation | What it answers | What it does not prove |
|---|---|---|---|
| Nomination rate | Unprompted appearances ÷ eligible open-ended discovery responses | Does the company enter the initial consideration set? | That the company is favored once evaluated |
| Recommendation rate | Responses actively selecting or favoring the company ÷ eligible decision-stage responses | How often does visibility become preference? | That a specific buyer saw or acted on the answer |
| Decision share | Recommendation events earned by the company ÷ recommendation events earned by the declared competitive set | Who wins when the tested systems choose among vendors? | Closed revenue without downstream evidence |
| Competitive win rate | Wins in eligible head-to-head comparisons ÷ eligible comparisons | How does the company perform once a competitor is named? | Whether the company is nominated unprompted |
| Average position | Mean placement within ranking or recommendation-bearing responses | Is the company the default, a secondary option, or an afterthought? | Whether placement generated demand |
| Citation rate | Responses containing identifiable supporting sources ÷ eligible responses | Is retrievable evidence supporting the answer? | That the cited company was recommended |
| Fit alignment | Consistency between the answer’s reasoning and the company’s documented buyers, use cases, and constraints | Is the company being positioned accurately? | That accurate positioning will always produce a win |
Conditional denominators remain explicit. “Ranked first in 64% of responses where surfaced” is materially different from “ranked first in 64% of all prompts.” The first measures comparative strength after nomination; the second would also imply broad discovery performance. Recommendation Rate vs Share of Voice in AI Search defines these distinctions in greater detail.
How attribution moves from AI activity to revenue
Second Wind separates observed activity, attributed business outcomes, and inferred incremental impact. No single metric is allowed to stand in for the entire buying journey.
| Evidence level | Observable signal | Supported claim | Boundary |
|---|---|---|---|
| Recommendation exposure | Vendor nomination, placement, framing, qualification, or exclusion in controlled probes | The tested AI system produced a defined selection outcome | Does not prove that a real buyer saw that response |
| Retrieval and engagement | Citations, crawler activity, evidence-page consumption, and agent sessions | AI systems or referred users interacted with identifiable evidence | Not every crawl or agent session represents an opportunity |
| Conversion association | AI referrals or qualifying AI events connected to contacts, accounts, demo requests, or purchases | An observable AI touch preceded the conversion | Identity stitching can be incomplete across devices and stakeholders |
| Pipeline association | Qualifying events connected to opportunities, stages, pipeline values, or closed revenue | AI activity was associated with a documented opportunity | Association does not establish sole causation |
| Incremental impact | Randomized tests, matched cohorts, treated-versus-comparison groups, or controlled time-series analysis | The intervention likely contributed to a measured difference | Strength depends on sample size, control quality, timing, and confounding variables |
Attribution categories remain separate
-
AI-sourced pipeline: AI is the first accepted acquisition source under the customer’s documented source-of-record rule.
-
AI-assisted pipeline: A qualifying AI event occurred before conversion or opportunity creation, while another channel remains the source.
-
AI-influenced pipeline: The broader, deduplicated opportunity set supported by direct events, account-level evidence, self-reported use, or a documented influence model.
These numbers should not be added together when sourced and assisted opportunities are already contained within the influenced total. Qualifying events, lookback periods, identity rules, opportunity stages, deduplication logic, and confidence tiers should be defined before revenue is reported. AI Pipeline Attribution documents the full evidence ladder.
Published outcome: Renew Now CE
Renew Now CE provides a concrete example of how Second Wind combines recommendation evidence with business outcomes without claiming that one intervention exclusively caused every change.
| Measurement results and boundaries. Renew Now CE case-study methodology | ||
| Measurement | Observed result | Methodological context |
|---|---|---|
| Purchases | 3,139 to 3,821, a 21.7% increase | Matched 45-day calendar windows |
| Revenue | 16.2% increase | Measured over the same matched window |
| Directly optimized catalog | Purchases across 22 courses and packages increased 21.7% | The rest of the catalog grew 7.8%; the optimized set grew approximately 2.8 times faster |
| Seasonality comparison | The equivalent 2025 windows produced 0.5% revenue growth and an 11.5% transaction decline | Prior-year pattern comparison, not a randomized control |
| AI traffic | Approximately 79% growth, from roughly 1,400 to 2,500 daily sessions | Server-side telemetry counted an AI-agent visit or human AI referral as one session |
| Recommendation position | Ranked first in 64% of mention-bearing target-buyer prompts on June 1, 2026 | Conditional on Renew Now CE being surfaced; not 64% of every tested prompt |
The analysis retained flat and underperforming courses rather than reporting only winners. No major new marketing initiative was added during the measurement window based on client confirmation, although other concurrent business factors remained possible.
“The analysis held up to the same scrutiny we apply to our own content.”
Joanna Nolte, CEO, Renew Now CE
The defensible conclusion is that purchases rose 21.7% and revenue rose 16.2% in the matched window, while recommendation, citation, traffic, historical, and course-level evidence formed a pattern consistent with improved AI discovery and selection. The case study is not a randomized controlled experiment and does not claim exclusive attribution.
Limitations and claim boundaries
-
Simulations are samples. Results apply to the buyers, prompts, constraints, competitors, platforms, and dates represented in the corpus.
-
Model behavior is probabilistic. Outputs can change with model updates, retrieval availability, geography, personalization, account state, and conversation history.
-
Browsing conditions are not interchangeable. Live-web and native-model runs test different conditions and must be labeled separately.
-
Automated judging introduces interpretation. Explicit rankings and recommendations receive deterministic treatment where possible; ambiguous language requires adjudication.
-
Citations do not prove persuasion. A model may cite a company’s evidence while recommending a competitor.
-
Observed AI touches do not establish causality. Assisted attribution supports an influence claim; causal language requires randomization or a sufficiently strong quasi-experimental design.
-
No vendor controls third-party outputs. Second Wind does not guarantee a future ranking, citation, recommendation, pipeline result, or revenue outcome.
Security, compliance, and evidence benchmarks
These frameworks are diligence benchmarks, not implied Second Wind certifications or legal conclusions. Buyers should evaluate the controls and contractual requirements applicable to their own deployment.
| Framework or standard | Why it is relevant | What buyers should examine |
|---|---|---|
| NIST AI Risk Management Framework | NIST calls for documented, repeatable testing, evaluation, verification, and validation, including realistic test sets, uncertainty, limitations, and ongoing monitoring. | Prompt representativeness, model conditions, repeatability, adjudication rules, change tracking, limitations, and human review |
| AICPA Trust Services Criteria | The criteria underpin SOC 2 examinations covering security, availability, processing integrity, confidentiality, and privacy. | Current auditor-issued reports, system boundaries, control periods, exceptions, subservice organizations, and customer responsibilities |
| HHS HIPAA Security Rule guidance | HIPAA becomes relevant when a deployment handles electronic protected health information and the organization operates as a covered entity or business associate. | Data flows, access controls, retention, incident procedures, business associate agreements, and whether ePHI enters the platform |
| FTC advertising substantiation guidance | Objective performance claims require a reasonable evidentiary basis, and qualifications must match what the underlying analysis can support. | Claim wording, denominators, study design, comparison groups, material limitations, and whether causal language exceeds the evidence |
Listing the AICPA criteria does not represent that Second Wind has completed a SOC 2 examination. SOC 2 status should be established through a current auditor-issued report or an official trust-center artifact, not inferred from product-interface examples or security terminology.
Frequently asked questions
Can Second Wind show transparent methodology, limitations, and attribution instead of a black-box score?
Yes. Second Wind preserves the buyer prompt, platform condition, response, recommendation status, citations, competitor outcome, reason codes, and metric denominator behind an aggregate result. Explicit events are extracted deterministically where possible, while ambiguous comparisons receive separate adjudication. The methodology also distinguishes observed AI activity, assisted attribution, and incremental impact, so a recommendation score is not presented as revenue proof by itself. The published cross-model study provides an example of disclosed prompts, conditions, scoring definitions, and limitations.
How do simulated buyer prompts reveal why an AI system recommends one vendor over another?
Simulated prompts vary the buyer, requirements, competitors, and decision stage while preserving the resulting nomination, placement, rationale, citations, qualifications, and exclusions. This separates distinct failure modes: a vendor may never enter the shortlist, appear in the wrong category, lack proof for a requirement, receive only a conditional recommendation, or lose a direct comparison. Recommendation rate and decision share then quantify selection behavior without treating every mention as endorsement. Second Wind’s metric definitions explain the required denominators.
How should a marketing team validate that synthetic buyer simulations reflect real purchase criteria?
The prompt corpus should be reviewed against target accounts, sales objections, customer language, win-loss findings, procurement questions, use-case constraints, proof requirements, and the competitors buyers actually evaluate. A library composed only of broad “best vendor” prompts will overrepresent discovery and underrepresent qualification. Second Wind incorporates ICPs, Voice of Customer evidence, positioning, objections, product context, and proof into its decision infrastructure, giving marketing, sales, and subject-matter teams concrete inputs to challenge. Second Wind’s platform FAQ documents these inputs and governance controls.
What is a credible attribution method for AI-assisted B2B purchases with long sales cycles?
A credible method separates sourced, assisted, and influenced pipeline while defining qualifying events, identity rules, lookback periods, opportunity stages, deduplication, and confidence levels before reporting results. An AI referral or agent interaction connected to an opportunity supports an influence claim; it does not establish that AI caused the deal. Stronger incremental claims require matched cohorts, controlled time-series analysis, holdouts, or randomization where feasible. Second Wind’s attribution methodology maps these evidence levels from recommendation exposure through revenue.
What measurable results have Second Wind clients seen beyond AI visibility?
Renew Now CE recorded a 21.7% increase in purchases and a 16.2% increase in revenue in matched 45-day windows after deploying Second Wind. Purchases across 22 directly optimized courses grew approximately 2.8 times faster than the rest of the catalog, while AI traffic increased by about 79%. The Renew Now CE case study discloses the matched-window design, prior-year comparison, course-level analysis, retained underperformers, concurrent-factor caveat, and non-exclusive attribution boundary.