AI Prompt Weighting: Formula and Revenue Framework

AI prompt weighting is the process of assigning different importance values to prompts before aggregating their monitoring results. In AI visibility measurement, the weight reflects business relevance—such as buying intent, revenue potential, audience fit, and demand—not token emphasis inside the prompt or the number of times the prompt is tested.

A defensible weighting process follows six steps:

1. Group paraphrases into distinct buyer-need families.
2. Score buying intent, revenue impact, audience fit, and demand evidence.
3. Calculate a normalized priority score for each family.
4. Cap extreme weights so one assumption cannot control the dashboard.
5. Report weighted and unweighted results side by side.
6. Test whether modest changes to the scoring model alter the conclusion.

This guide presents maxaeo’s IRAF framework—Intent, Revenue, Audience, and Frequency—with formulas, spreadsheet fields, governance controls, and a worked example. IRAF is an original decision framework, not an industry benchmark or a claim that a particular weight causes revenue.

What Does “AI Prompt Weighting” Mean?

The phrase has several meanings, and they should not be treated as interchangeable. Someone creating an image may want to emphasize a word inside one prompt. A marketer monitoring AI answers usually wants to make commercially important questions count more in an aggregate visibility score.

Meaning	What is being weighted?	Typical purpose	Is there a universal standard?
Generation emphasis	Words, phrases, or concepts inside one request	Influence an image or response	No; syntax depends on the tool
Instruction priority	System, developer, user, and contextual instructions	Resolve competing directions	No portable numeric weighting system
Soft-prompt weighting	Learned vectors or prompt parameters	Model training or adaptation	No; implementation is model-specific
Monitoring priority	Questions in an AI visibility dataset	Prioritize commercially important results	No; the business defines and documents the method

A notation such as (enterprise security:1.3) only has an effect when the selected application explicitly parses that syntax. It is not a universal command understood by every language or image model.

For general-purpose language models, clear instructions, relevant context, examples, and explicit evaluation criteria are usually more reliable than repeating a phrase or inventing numeric syntax. OpenAI’s official prompt engineering guidance documents these techniques but does not define a portable numeric token-weighting syntax for ordinary prompts.

The remainder of this guide focuses on business weighting for AI search monitoring.

Why Weight AI Search Prompts?

Weighting answers a business question that an equal-weight visibility score cannot: are we visible when commercially valuable buyers are making decisions?

Compare these prompts:

• “What is project management software?”
• “Best project management platform for a 200-person engineering team”
• “Vendor A vs Vendor B for Jira portfolio reporting”

All three are useful monitoring prompts, but they represent different stages, customer profiles, and potential outcomes. Treating them equally can make strong educational visibility conceal weak shortlist visibility.

Prompt weighting is useful when a team needs to:

• Separate general category awareness from vendor-selection visibility.
• Prioritize prompt families connected to valuable customer segments.
• Identify cases where high overall visibility masks weak commercial performance.
• Allocate content, digital PR, product marketing, and reputation work.
• Explain why two prompts with similar mention rates deserve different attention.

It should not be used to:

• Remove prompts where the brand performs poorly.
• increase a headline score after results are known.
• substitute assumptions for missing performance data.
• claim that an AI answer caused revenue.
• make a biased prompt set representative.

Build the prompt portfolio first. A documented workflow such as creating a prompt set for AI brand monitoring helps establish coverage before weights are introduced.

Why Must the Unweighted Baseline Be Preserved?

The unweighted result is the audit baseline; the weighted result is a decision lens. Publishing both prevents methodology changes from being mistaken for visibility improvements.

The two metrics answer different questions:

Metric	Question answered
Unweighted baseline	How often does the brand appear or get recommended across the monitored portfolio?
Business-weighted result	How often does the brand appear or get recommended on prompts the business currently considers most important?

Suppose a brand has a 55% unweighted recommendation rate but only a 34% weighted rate. The 21-percentage-point gap indicates that performance is concentrated in lower-value questions.

If only the weighted result is retained, a team could improve the dashboard by changing coefficients, removing difficult prompts, or adding more variations of a successful prompt. None of those actions demonstrates better AI visibility.

Every report should therefore show:

• The unweighted rate.
• The weighted rate.
• The difference in percentage points.
• The number of prompt families and variants.
• Valid answer counts.
• The active prompt-set and weight versions.
• The date on which weights were approved.

What Should an AI Prompt Weighting Worksheet Contain?

The worksheet should keep business assumptions separate from observed AI results. A prompt must not receive a higher weight because the brand already performs well on it.

Field	What it records	Required control
Prompt ID	Permanent identifier	Never reuse an ID
Prompt version	Exact wording revision	Retain previous wording
Prompt family	Shared buyer need across paraphrases	Deduplicate before aggregation
Prompt class	Non-branded, branded, comparison, support, or reputation	Use for segmentation, not automatic value
Locale and audience	Market, language, role, and customer context	Do not merge materially different audiences
Buying intent	Distance from learning to vendor selection	Score before reviewing results
Revenue impact	Value of the matched segment or use case	Use one consistent financial measure
Audience fit	Alignment with the ideal customer profile	Score desirability, not current win likelihood
Frequency proxy	Evidence that the buyer need recurs	Store source, period, and confidence
IRAF score	Calculated business priority	Formula-controlled
Final weight	Normalized and capped score	Version with the coefficients
Valid-answer count	Eligible observations	Exclude failures under a fixed rule
Outcome fields	Mention, recommendation, rank, sentiment, accuracy, and citation	Store separately from weighting inputs

Prompt class and buying intent should remain separate. “How do I reset my Vendor A password?” is branded but has little acquisition intent. “Best compliance software for a Series B fintech” is non-branded but may have substantial commercial value.

How Should Buying Intent Be Scored?

Score the decision implied by the complete question, not isolated words such as “best,” “review,” or a brand name.

Use a five-level starting rubric:

Intent value	Buyer stage	Observable prompt pattern
0.2	General learning	Requests a definition or category explanation
0.4	Problem exploration	Describes a problem without requesting solutions
0.6	Category evaluation	Asks which type of tool or approach could solve the problem
0.8	Shortlist formation	Requests suitable products with audience or use-case constraints
1.0	Vendor selection	Compares named options, pricing, migration, procurement, or final fit

Examples:

Prompt	Intent	Reason
“What is revenue intelligence?”	0.2	Definition request
“How can we reduce sales forecast errors?”	0.4	Problem exploration
“What tools automate sales forecasting?”	0.6	Category evaluation
“Best forecasting tools for a global SaaS team”	0.8	Constrained shortlist
“Vendor A vs Vendor B for Salesforce forecasting”	1.0	Named vendor selection

Intent is contextual. “Best way to explain forecasting to new hires” remains informational, while “best forecasting platform for Salesforce” is evaluative.

If two reviewers disagree by more than one level, require a written reason and adjudication. Large disagreement often means the prompt is ambiguous or contains multiple buyer jobs that should be separated.

How Should Revenue Impact Be Scored?

Revenue impact estimates the value of winning the customer segment or use case represented by the prompt. It should never depend on whether the brand is currently mentioned.

Choose one financial measure and apply it consistently, such as:

• Expected first-year gross profit.
• Expected contract value.
• CRM-weighted pipeline value.
• Contribution margin for an e-commerce purchase.
• Qualified lead value for a lead-generation business.

When reliable segment data exists:

Raw revenue value = attainable win probability × expected first-year gross profit

Normalize the result to a value from 0 to 1. A logarithmic transformation limits the influence of unusually large enterprise contracts:

R = ln(1 + raw revenue value) ÷ ln(1 + maximum raw revenue value)

Percentile ranks are another defensible option when the underlying values are highly skewed.

Do not alternate between annual contract value, lifetime value, and pipeline depending on which number produces the preferred ranking. Record the financial measure, source period, currency, and data owner in the weight version.

When evidence is missing, use a neutral provisional score such as 0.5, label it low confidence, and schedule a review. False precision is less useful than a visible uncertainty flag.

How Should Audience Fit Be Measured?

Audience fit measures whether the buyer represented by a prompt resembles a customer the business wants and can serve. It is not a prediction that the brand will win or be recommended.

Score four dimensions from 0 to 1:

1. Industry or business model.
2. Company size or maturity.
3. Geography, regulatory environment, or technical environment.
4. Buyer role, job, or use case.

Calculate:

A = (industry fit + company fit + market fit + use-case fit) ÷ 4

A score of 1.0 means the prompt describes a direct ideal-customer-profile match. A plausible adjacent segment might receive 0.5 or 0.75. A market the company cannot legally, operationally, or technically serve should receive 0.

Do not give a perfect audience score solely because a potential contract is large. If the product lacks a required deployment model, certification, integration, language, or service geography, the audience fit is lower regardless of contract size.

How Can Prompt Frequency Be Estimated?

Prompt frequency should estimate how often a buyer need occurs—not how many times the monitoring tool executes the prompt. Monitoring run count is a sampling decision, not evidence of buyer demand.

Useful demand signals include:

• Repeated themes in sales and discovery-call transcripts.
• Search query families in Google Search Console or paid search.
• On-site search and chatbot questions.
• Support, community, and customer-success conversations.
• RFP requirements and win-loss interviews.
• Structured customer or prospect surveys.

Google Search Console reports activity from Google Search, as defined in its official Performance report documentation. Its impressions are useful demand evidence for related search needs, but they are not a census of prompts entered into ChatGPT, Gemini, Perplexity, Claude, or other answer engines.

Normalize each source before combining it

Do not add 10,000 search impressions directly to six enterprise RFP observations. The units and evidentiary value are different.

For source s and prompt family k, normalize within that source:

xₛₖ = ln(1 + family countₛₖ) ÷ ln(1 + highest family countₛ)

Then combine normalized sources:

Fₖ = Σ(cₛ × xₛₖ) ÷ Σcₛ

Where cₛ is a reliability coefficient assigned before AI performance is reviewed.

For example, recent CRM-linked discovery calls may receive greater evidentiary confidence than an undated keyword list. Store the source, observation window, sample size, normalization method, and coefficient so the frequency score can be reproduced.

Keywords are useful starting evidence, but they must be rewritten as natural buyer questions. This SEO-keyword-to-AI-prompt workflow explains how to preserve search intent without treating keywords and conversational prompts as identical data.

How Is the IRAF Score Calculated?

IRAF combines buying intent, revenue impact, audience fit, and frequency using an additive model. The default coefficients are a transparent starting hypothesis, not universal economic truth.

Use:

IRAF score = 100 × (0.35I + 0.30R + 0.20A + 0.15F)

Where:

• I = buying intent from 0 to 1.
• R = normalized revenue impact from 0 to 1.
• A = audience fit from 0 to 1.
• F = normalized frequency proxy from 0 to 1.

The default model gives intent and revenue the greatest influence while preserving strategic-fit and demand signals.

An additive formula is preferable to multiplying all four inputs. In a multiplicative model, one missing or provisional zero can erase the value of an otherwise important prompt. Multiplication also magnifies scoring noise and makes disagreements harder to diagnose.

Convert the score into a portfolio weight

Normalize each IRAF score around the portfolio average:

Raw prompt weight = IRAF score ÷ average IRAF score

Then cap the result:

Final weight = min(1.5, max(0.5, raw prompt weight))

The 0.5 floor preserves category-learning, reputation, support, and emerging-demand signals. The 1.5 ceiling prevents a small number of high-value assumptions from controlling the dashboard.

Organizations may choose different caps, but the values and reasons should be recorded before performance is examined.

How Should Paraphrases and Prompt Families Be Weighted?

Assign business importance to the buyer need first, then allocate it across wording variants. Otherwise, a large paraphrase family can dominate the metric without representing more demand.

These prompts may test the same underlying need:

• “Best analytics software for startups”
• “Which analytics platform should a startup choose?”
• “Top startup analytics tools”

Use one of two valid methods:

1. Family-first aggregation: Calculate each variant’s result, average the variants into one family rate, and apply one family weight.
2. Allocated prompt weighting: Divide the family weight by the number of valid variants and apply that share to each variant.

Family-first aggregation is generally easier to audit:

Family rate = Σ variant rates ÷ number of valid variants

Weighted portfolio rate = Σ(family weight × family rate) ÷ Σ family weights

Do not merge prompts when language, geography, buyer role, company size, or product requirement materially changes the decision. “Best CRM for a five-person agency” and “best CRM for a regulated global bank” are not simple paraphrases.

What Spreadsheet Formulas Can Implement the Method?

A basic spreadsheet can calculate IRAF without hiding the assumptions inside a proprietary score.

Use these columns:

Column	Value
A	Prompt or family ID
B	Prompt family
C	Intent score
D	Revenue score
E	Audience-fit score
F	Frequency score
G	IRAF score
H	Final weight
I	Valid answers
J	Recommendations
K	Recommendation rate

For row 2:

G2 = 100*(0.35*C2+0.30*D2+0.20*E2+0.15*F2)

H2 = MIN(1.5,MAX(0.5,G2/AVERAGE($G$2:$G$101)))

K2 = IFERROR(J2/I2,"")

Portfolio formulas:

Unweighted rate = AVERAGE(K2:K101)

Weighted rate = SUMPRODUCT(H2:H101,K2:K101)/SUMPRODUCT(H2:H101,--ISNUMBER(K2:K101))

The weighted denominator should include only rows with a valid outcome rate. Preserve the raw inputs and formulas rather than pasting calculated values over them.

What Does a Worked AI Prompt Weighting Example Show?

The example below shows how a healthy-looking baseline can conceal weak visibility during vendor selection.

The dataset contains six fictional ExampleCo CRM prompt families. Each has 20 valid answers collected on the same platform and schedule. The recommendation observations are constructed solely to demonstrate the calculation; they are not maxaeo customer data or an industry benchmark.

Prompt family	I	R	A	F	IRAF	Weight	Recommended
What is CRM software?	0.2	0.2	0.5	1.0	38.0	0.56	15/20
Best CRM for a 50-person B2B SaaS team	1.0	0.8	1.0	0.8	91.0	1.35	4/20
CRM with SOC 2 support and easy Salesforce migration	0.8	0.9	1.0	0.5	82.5	1.22	9/20
ExampleCo CRM reviews for B2B SaaS	0.9	0.8	1.0	0.4	81.5	1.21	6/20
ExampleCo CRM vs MarketSuite	1.0	0.7	0.8	0.5	79.5	1.18	7/20
How to import contacts into a CRM	0.3	0.1	0.4	0.7	32.0	0.50	17/20

For the second prompt:

IRAF = 100 × [(0.35 × 1.0) + (0.30 × 0.8) + (0.20 × 1.0) + (0.15 × 0.8)] = 91

The unweighted recommendation rate is:

58 recommendations ÷ 120 valid answers = 48.3%

After applying the capped prompt weights:

Weighted recommendation rate = Σ(weight × prompt rate) ÷ Σweights = 40.6%

The −7.7 percentage-point gap reveals the business problem: ExampleCo performs well on broad educational and support questions but poorly when qualified buyers form a shortlist.

How Should Coefficient Sensitivity Be Tested?

A weighting model is not decision-ready until the team knows whether reasonable coefficient changes reverse its conclusion.

The same six-prompt dataset produces these results:

Model	Intent	Revenue	Audience	Frequency	Weighted rate
Balanced	25%	25%	25%	25%	42.0%
Default IRAF	35%	30%	20%	15%	40.6%
Intent-first	50%	20%	15%	15%	40.3%
Revenue-first	20%	50%	15%	15%	40.6%

All four models remain well below the 48.3% unweighted baseline. The exact weighted percentage changes, but the diagnosis does not: commercial prompt performance is weaker than general visibility.

That stability is more important than defending 35% instead of 30% as the “correct” intent coefficient.

Investigate the model when:

• A five- or ten-point coefficient change reverses the priority list.
• One prompt repeatedly reaches the weight ceiling.
• Provisional revenue data determines the conclusion.
• One reviewer’s scores materially change the result.
• The weighted metric moves sharply while raw prompt results remain stable.

In those cases, improve the evidence or report multiple scenarios instead of presenting one score as objective truth.

How Should Confidence Affect Prioritization?

Evidence confidence should affect the decision to act, but it should not be hidden inside the headline visibility weight.

A practical confidence rubric is:

Confidence	Evidence condition
1.00	Recent first-party evidence from at least two independent sources
0.75	One strong first-party source or several consistent proxies
0.50	Modeled or third-party proxy with limited validation
Below 0.50	Evidence too weak for high-cost action

Use confidence in the opportunity calculation:

Opportunity = final weight × max(0, target rate − current rate) × evidence confidence

This keeps two questions separate:

• How important is this prompt family if the assumptions are correct?
• How confident are we in those assumptions?

A high-value, low-confidence family may justify research before content investment. A high-value, high-confidence family with a large recommendation gap can move directly into the execution backlog.

How Should Weighted AI Visibility Be Reported?

Report prompt-level rates before aggregation so extra monitoring runs do not silently create extra business importance.

For prompt or family i:

pᵢ = recommended valid answers ÷ all valid answers

Then calculate:

Unweighted rate = average of all pᵢ values

Weighted rate = Σ(wᵢ × pᵢ) ÷ Σwᵢ

This is a macro-average: each prompt family’s influence comes from its documented business weight, not its run count.

Keep outcomes separate:

• Brand mentioned.
• Explicitly recommended.
• Included in a shortlist.
• Ranked first, second, or later.
• Described accurately.
• Supported by a visible citation.
• Discussed positively, neutrally, or negatively.

A brand may be mentioned as a poor fit, cited without being recommended, or recommended without a visible citation. Combining these outcomes into one “visibility” flag removes information needed for diagnosis.

Report platforms separately

Prompt importance, platform importance, and sample size are different variables.

Apply the same prompt-family weights separately within ChatGPT, Gemini, Perplexity, Claude, Copilot, Google AI Mode, AI Overviews, or other monitored surfaces. If customer research supports a platform-mix adjustment, calculate it as a second, explicitly named layer:

Cross-platform rate = Σ(platform-mix weight × platform portfolio rate) ÷ Σ platform-mix weights

Do not embed platform usage assumptions inside the prompt weight. More monitoring runs improve measurement confidence; they do not make a prompt or platform more commercially valuable.

Monitoring cadence is a separate design choice. The AI search monitoring frequency guide explains when daily, weekly, monthly, or quarterly observations are appropriate.

How Can the Method Remain Auditable?

Auditability requires immutable raw observations and versioned business assumptions. A historical dashboard should be reproducible from the prompt text, raw answers, scoring rubric, coefficients, and weight table active at that time.

Use this control sequence:

1. Freeze the exact prompt. Store its ID, wording, family, locale, audience, and version.
2. Retain the raw answer. Record platform, displayed model label, timestamp, citations, and eligibility status.
3. Define valid observations. Document how refusals, timeouts, missing answers, and unavailable features are handled.
4. Fix the evaluation rubric. Define mentions, recommendations, ranks, citations, and factual errors before scoring.
5. Score business inputs independently. Do not expose performance results to weight reviewers.
6. Publish both metrics. Retain the unweighted baseline beside the weighted result.
7. Version every change. Record the owner, evidence, coefficients, approval date, and reason.
8. Run an overlap period. When weights change, calculate results under both versions before replacing the official series.

Never overwrite historical data with a revised methodology. If old observations are recomputed under a new rubric or weight table, label the series as restated and preserve the original.

How Should Weighted Results Prioritize GEO Work?

Prioritize prompt families that combine business importance, a measurable performance gap, sufficient evidence, and an issue the team can influence.

Set the target using a documented reference:

• The brand’s historical best result.
• A leading competitor’s observed rate.
• An approved business threshold.
• A controlled pre-change baseline.

Then diagnose the answer rather than publishing generic content:

Observed failure	Likely workstream
Brand absent from relevant shortlists	Category positioning, comparison coverage, and third-party corroboration
Brand mentioned but ranked low	Use-case differentiation and proof
Recommendation contains incorrect facts	Entity consistency and authoritative corrections
Brand recommended without citations	Concise, retrievable evidence supporting the recommendation
Brand cited but not recommended	Product fit, reputation, pricing, or competitive proof
Performance varies sharply across paraphrases	Entity ambiguity, weak topical consistency, or insufficient sampling
Performance fails only in one market	Locale-specific evidence, availability, or regulatory fit

Translate validated diagnoses into an AI visibility backlog prioritized by recommendation impact. High weight alone is not an instruction to produce another article.

How Can the Metric Be Connected to Revenue Responsibly?

A weighted AI visibility score is a leading indicator, not proof of revenue attribution.

Use three evidence layers:

• Visibility: recommendation rate, rank, description accuracy, and citation presence on high-weight prompts.
• Behavior: self-reported AI research, identifiable AI referrals, branded search movement, and sales-call mentions.
• Commercial outcomes: qualified opportunities, win rate, sales velocity, and revenue in the relevant segments.

For a stronger evaluation:

1. Pre-register which prompt families a planned change should affect.
2. Select similar untouched families as a comparison group.
3. Keep platforms, locales, run schedules, and evaluation rules stable.
4. Measure both groups over the same period.
5. Review raw answers for unrelated model or market changes.
6. Report contribution and uncertainty rather than claiming perfect attribution.

If targeted families improve while comparison families remain stable, the intervention becomes a more plausible contributor. It still does not prove that an AI answer caused a purchase.

Which AI Prompt Weighting Mistakes Distort Results?

The worst mistakes let the measurement design reward itself or allow weights to change after performance becomes visible.

Mistake	Why it fails	Better control
Giving every paraphrase a full weight	Large families dominate without representing more demand	Average variants within a family or divide the family weight
Using monitoring runs as frequency	Sampling effort is mistaken for buyer demand	Use external demand evidence
Scoring after seeing results	Coefficients can be tuned to improve the headline	Freeze inputs before measurement
Giving low-intent prompts zero weight	Discovery, reputation, and emerging demand disappear	Use a documented floor
Letting one enterprise prompt dominate	A fragile revenue estimate controls the report	Normalize revenue and cap weights
Mixing platform and prompt weights	Model changes become indistinguishable from strategy changes	Report platform mix separately
Showing only the weighted score	Methodology changes can resemble growth	Preserve the unweighted baseline
Combining mentions and recommendations	Negative or incidental mentions look like success	Maintain separate outcome metrics
Treating search volume as AI prompt volume	Different behaviors and platforms are conflated	Label search data as a proxy
Updating weights too frequently	The baseline moves faster than performance can be interpreted	Use a scheduled governance cycle

How Can a Team Implement IRAF in 30 Days?

A four-week rollout is sufficient for a defensible first version when the team prioritizes governance over artificial precision.

Week 1: Define the prompt portfolio

• Deduplicate prompt families.
• Separate branded, non-branded, comparison, reputation, and support prompts.
• Confirm coverage across discovery, evaluation, shortlist, and selection stages.
• Record important audiences, locales, and use cases.
• Identify missing buyer needs before weighting begins.

Prompt count should follow coverage, not an arbitrary quota. The guide to how many AI search prompts to track provides a portfolio-based approach.

Week 2: Score independently

• Have marketing score intent and audience fit.
• Have sales validate buyer-stage assumptions.
• Have finance or revenue operations provide the commercial measure.
• Hide AI performance results during scoring.
• Investigate differences greater than one intent level or 0.25 on a normalized input.

Week 3: Add demand evidence

• Map first-party signals to prompt families.
• Normalize each evidence source separately.
• Record source periods and sample sizes.
• Assign confidence levels.
• Mark provisional values instead of inventing precision.

Week 4: Establish and challenge the baseline

• Collect sufficient valid observations.
• Calculate outcome rates at the family level.
• Publish weighted and unweighted results.
• Run balanced, intent-first, and revenue-first sensitivity tests.
• Freeze version 1 only after reviewers understand which assumptions drive the result.

Review observations as often as the monitoring program requires. Review business weights quarterly or after a material change to the ideal customer profile, product, pricing, economics, market availability, or strategy.

Frequently Asked Questions About AI Prompt Weighting

Does repeating a word give it more weight in every AI model?

No. Repetition may influence some outputs, but it is not a reliable or standardized numeric control. Tool-specific syntax only works when that application explicitly supports it. For language models, use clear instructions, context, examples, and evaluation criteria rather than assuming repeated words receive a predictable weight.

Should low-intent prompts receive a weight of zero?

Usually not. Low-intent prompts can reveal category discovery, reputation errors, support friction, and emerging demand. A floor such as 0.5 preserves those signals while limiting their effect on the weighted total. Exclude a prompt only when it is outside the documented scope.

How many prompts are needed before weighting is useful?

There is no universal minimum. Weighting becomes useful when the set represents important buyer needs, stages, audiences, and markets. Start with distinct families and add paraphrases only when they test meaningful wording or context differences. More prompts cannot repair a biased sample.

How often should prompt weights be updated?

Review evidence quarterly and revise weights after a material change in the product, ideal customer profile, economics, or market strategy. Keep this governance cycle separate from daily or weekly monitoring. Calculate an overlap period under both weight versions before changing the official series.

Can third-party AI prompt-volume estimates be used?

Yes, but only as disclosed proxies. Record the provider, geography, period, collection method, and confidence level. Compare estimates with sales calls, customer research, and related search behavior. Do not present modeled prompt volume as a complete census of answer-engine usage.

Is weighted AI share of voice the same as revenue impact?

No. Weighted AI share of voice shows how often a brand appears relative to competitors on prompts the business considers important. It can guide prioritization, but it does not prove that an AI response generated a sale.

The Decision Rule

Use AI prompt weighting to prioritize decisions, not to rewrite history. Define the prompt portfolio first, score business importance before seeing results, cap extreme weights, preserve raw observations, and publish the weighted result beside an unweighted baseline.

Pay particular attention to the gap between the two metrics:

• A negative gap means broad visibility may be masking weakness on commercially important questions.
• A positive gap indicates strength near the buying decision but may expose limited category discovery.
• A rapidly changing gap with stable raw results usually points to a methodology or portfolio change.
• A conclusion that reverses under modest coefficient changes needs better evidence before investment.

IRAF turns an abstract visibility score into a testable operating model. Its value does not come from pretending the coefficients are universally correct. It comes from making the business assumptions explicit, challengeable, and reproducible.