For years, digital competition was relatively easy to describe.
Search engines gave us rankings.
A brand could ask:
«“What position do we have for this keyword?”»
And the answer was measurable.
Position #1.
Position #4.
Page 2.
SEO turned visibility into a measurable market.
Generative AI changes the problem.
Ask an AI system:
«“Who are the best bitumen exporters from Iran?”»
The system does not return a stable ranking.
It generates an answer.
Run the same question again and the answer may change.
Change the wording, language, geography, model, or context—and the recommendation set may change again.
This creates a new question:
«How much of the AI recommendation market belongs to a brand?»
I call this AI Recommendation Share (ARS).
From Search Rank to Recommendation Probability
Traditional search visibility can be approximated by position:
Keyword → Search Engine → Ranking
Generative search looks more like:
Question
↓
AI Model / Retrieval System
↓
Candidate information
↓
Probabilistic generation
↓
Recommendation
The output is not simply a rank.
It is a distribution.
For a given commercial intent, a brand might appear:
- as the first recommendation,
- among the top alternatives,
- as a supplier mentioned in passing,
- only as a cited source,
- or not at all.
Therefore, counting mentions is not enough.
We need to measure recommendation probability and recommendation position.
Defining AI Recommendation Share
Suppose we construct a controlled set of commercial questions:
Q = {q1, q2, ..., qn}
Each question is executed repeatedly across a defined set of AI systems:
M = {m1, m2, ..., mk}
and repeated:
R = {r1, r2, ..., rt}
For each response, we record whether a target brand appears and how it appears.
A simplified metric could be:
[
ARS_b =
\frac{
\sum_{q,m,r} w_q w_m S(b,q,m,r)
}{
\sum_{q,m,r} w_q w_m
}
\times 100
]
Where:
- (b) = brand
- (q) = commercial question
- (m) = AI system
- (r) = independent run
- (w_q) = commercial importance of the question
- (w_m) = market weight of the AI system
- (S) = recommendation score
The important point is that one AI response is not the measurement.
The distribution of responses is.
Mention Is Not Recommendation
This distinction is critical.
Imagine an AI response:
«“The major suppliers include A, B, C, and D. Brand X is also mentioned in several industry directories.”»
Brand X was mentioned.
But was it recommended?
Not necessarily.
A useful measurement system should therefore distinguish between:
AI outcome| Example score
Primary recommendation| 1.00
Top-3 recommendation| 0.70
Shortlist mention| 0.45
Incidental mention| 0.20
No appearance| 0
The exact scoring model should be standardized and published.
The objective is not to create a magical number.
The objective is to create a repeatable measurement protocol.
The Measurement Protocol Is the Product
This is where the opportunity becomes much more interesting.
Anyone can build a script that sends prompts to an LLM.
The difficult part is deciding:
- Which questions should be asked?
- How many times?
- Which models?
- Which countries?
- Which languages?
- Which buyer personas?
- How should recommendations be scored?
- How should citations be treated?
- How should duplicate brands be resolved?
- How should model updates be handled?
- What happens when an AI provider changes its retrieval system?
Without a fixed protocol, two companies can report completely different AI visibility numbers for the same brand.
Therefore, ARS should behave more like a market index than a marketing dashboard.
The Market Question Cell
A useful unit of measurement is a Market Question Cell:
Intent
Product
Market
Language
Buyer Type
AI System
Date
For example:
Intent: Vendor discovery
Product: Bitumen VG 30
Market: GCC
Language: English
Buyer: Industrial procurement manager
AI System: Defined engine
Date: 2026-10-07
This prevents an important statistical mistake:
«Treating fundamentally different commercial questions as if they were the same query.»
“Best bitumen supplier” and “lowest-price bitumen supplier for bulk orders in UAE” may belong to the same product category, but they represent different buying intents.
Why Thousands of Queries Matter
One of the biggest mistakes in AI visibility measurement is taking a single answer seriously.
A single response is a sample.
It is not the market.
Instead, imagine running:
500 commercial questions
×
4 AI systems
×
7 independent runs
×
multiple languages
×
multiple dates
Now we have a distribution.
We can estimate:
- recommendation probability,
- confidence intervals,
- volatility,
- model-specific differences,
- temporal trends,
- competitive movement.
The question changes from:
«“Did ChatGPT mention us?”»
to:
«“What is the estimated probability that AI systems recommend us when buyers ask questions in our market?”»
That is a much more useful business metric.
AI Recommendation Gap
There is another metric hiding inside this idea.
A company can have substantial real-world market share while having very little AI recommendation share.
For example:
Economic Market Share: 18%
AI Recommendation Share: 5%
This creates an:
AI Recommendation Gap
[
ARG = Economic\ Market\ Share - AI\ Recommendation\ Share
]
In this example:
ARG = 18% - 5% = 13 percentage points
This gap could become strategically important.
It identifies companies that have strong real-world businesses but weak representation in AI-mediated discovery.
AI Recommendation Momentum
Share alone is not enough.
Velocity matters.
Suppose:
Current ARS: 11.2%
30-day change: +3.7pp
90-day change: +8.4pp
The brand may not yet dominate the category.
But it is gaining rapidly.
This suggests another metric:
AI Recommendation Momentum
The objective is to measure the rate at which a brand's AI recommendation share changes over time.
Now the dashboard becomes more than a visibility report.
It becomes a market dynamics system.
The Real Moat Is Historical Data
The dashboard itself is easy to copy.
The historical dataset is not.
Imagine continuously storing:
Timestamp
Prompt
Prompt family
Language
Market
AI system
Raw response
Brand entities
Recommendation position
Citation sources
Confidence
Scoring version
After one year, you have a historical record.
After three years, you potentially have something much more valuable:
«A longitudinal map of how AI systems construct commercial recommendations.»
You can ask:
- Which brands consistently gain recommendation share?
- Which sources influence recommendations?
- Which industries are becoming more AI-mediated?
- Which types of evidence correlate with visibility?
- How quickly does a new brand enter the recommendation graph?
- What happens after a certification is published?
- How stable are recommendations across models?
This is no longer simply SEO analytics.
It is AI market intelligence.
From Measurement to Attribution
The next layer is even more interesting.
Suppose:
ARS increased from 8.4% → 13.2%
The client does not only want to know that it increased.
They want to know:
«Why?»
A future system could maintain an evidence graph:
Brand
├── Product pages
├── Technical documentation
├── Certifications
├── Industry directories
├── News
├── Reviews
├── Expert sources
└── Third-party citations
Then changes in AI recommendation visibility can be investigated against changes in the evidence environment.
This does not mean claiming simplistic causality.
It means building a structured system for evidence attribution and hypothesis testing.
That distinction matters.
The Performance Contract
This is where measurement becomes a business model.
Instead of selling:
«“We will publish 100 articles.”»
sell:
«“We will increase your AI Recommendation Share from 5% to 20% under a predefined measurement protocol.”»
For example:
Baseline ARS: 5.1%
Target: 15.0%
Stretch target: 20.0%
Measurement window: 90 days
The commercial contract can then be partially performance-based.
But there is an important condition:
The measurement protocol must be locked before the result is known.
The client should not be allowed to choose only questions where the brand performs well.
The agency should not change the scoring system after seeing the results.
The AI systems should be predetermined.
The prompt portfolio should be versioned.
The baseline should be recorded.
And major changes in an AI provider's underlying system should be treated as measurement events.
This makes the metric auditable.
Measurement Before Optimization
This leads to a simple architecture:
AI MARKET
│
▼
┌─────────────────┐
│ Measurement │
│ Protocol │
└────────┬────────┘
│
▼
┌─────────────────┐
│ ARS Benchmark │
└────────┬────────┘
│
▼
┌─────────────────┐
│ Evidence │
│ Engineering │
└────────┬────────┘
│
▼
┌─────────────────┐
│ ARS Change │
└────────┬────────┘
│
▼
┌─────────────────┐
│ Performance │
│ Contract │
└─────────────────┘
This creates a closed loop:
Measure → Diagnose → Improve → Measure Again
Why This Could Become a New Category
SEO created an enormous industry around a relatively simple question:
«“Where does my website rank?”»
Generative AI introduces a different question:
«“When buyers ask AI what they should choose, how often does my brand become part of the answer?”»
That question will become increasingly important as AI systems move from information retrieval toward decision assistance.
The competitive battlefield may shift from:
Search Ranking
to:
Recommendation Probability
And that creates a new measurement problem.
Whoever defines the measurement standard has an opportunity to define the category.
But There Is a Serious Risk
AI recommendation share should never become another vanity metric.
It must not be manipulated through:
- fake reviews,
- fabricated citations,
- spam networks,
- misleading claims,
- synthetic authority,
- entity stuffing,
- retrieval manipulation.
The goal should be to improve the evidence environment surrounding a legitimate business, not to trick an AI system into recommending it.
Otherwise the metric becomes meaningless.
The strongest version of this idea is therefore not:
«“How do we make AI mention our brand?”»
It is:
«“How do we make the information ecosystem around a legitimate brand sufficiently clear, authoritative, verifiable and accessible that AI systems can accurately represent it?”»
That is a much more defensible proposition.
The Bigger Idea
The deeper shift is this:
Search engines gave businesses rankings.
Social networks gave businesses engagement metrics.
Marketplaces gave businesses conversion metrics.
Generative AI may create a new commercial metric:
«Recommendation Share.»
And if AI becomes part of how buyers discover suppliers, compare products, evaluate vendors and make purchasing decisions, then recommendation share could become an economically meaningful layer between information and transaction.
The opportunity is therefore not to build another AI SEO dashboard.
It is to build the measurement infrastructure for the AI recommendation economy.
The Thesis
«SEO measured where you ranked.
Generative AI changes who gets recommended.
AI Recommendation Share measures the market between the question and the recommendation.»
The companies that learn to measure that market early may not simply optimize for AI.
They may help define how AI-mediated markets are measured at all.
And that is a much bigger opportunity than another SEO tool.
Created by Seyed Alireza Alhosseini Almodarresieh
Top comments (0)