Share of Model:
The AI-Era Metric Every B2B Team Should Track

A board member forwards a screenshot. They asked a popular AI assistant which platforms a company like theirs should evaluate in your category, and it named three vendors with reasons. You were not one of them. The question in the email is short: "Why is AI recommending them and not us, and how would we even know if this is happening at scale?"

Most marketing dashboards have no answer. They report sessions, rankings, share of voice, leads, and pipeline. None of those metrics can see the moment a model narrows your buyer's shortlist before a form is filled or an SDR is alerted. The measurement system was built for a buying journey that started on your website. That is no longer where it starts.

This article is about the metric that closes that blind spot. Not another argument that AI matters, and not a definitions post about AEO. A practical, board-grade way to measure whether the machines that now advise your buyers actually recommend you, and a framework for improving it.

Why This Matters Now

For two decades, B2B demand ran on a familiar physics. A buyer had a problem, searched, clicked, landed, and entered a funnel you could instrument. Marketing optimized for rankings and clicks; sales optimized for the conversation that followed. The model assumed the buyer would eventually arrive on a property you controlled.

That assumption is breaking. Forrester's Buyers' Journey Survey (2025) found that twice as many B2B buyers named generative AI or conversational search as their most meaningful source of information than any other channel, ahead of vendor websites, product experts, and sales reps. Gartner (2026) found that most B2B buyers now prefer a rep-free, digital, self-service research experience. The early research that used to land on your site increasingly runs through a model first.

I wrote earlier about how your buyer's AI is already researching you before sales ever speaks. That piece described the demand side, what the buyer experiences. The AI Visibility Architecture describes the supply side, what you build so the machine has something accurate to say about you. This article is about the third leg: measurement. You cannot manage what you cannot see, and right now most revenue teams cannot see the most consequential moment in their buying cycle.

Why Mention-Counting Fails the Board

As soon as AI visibility became a concern, a market of tools and agencies rushed in to "measure share of model." Most of them count the same thing: how often your brand name appears across a list of prompts, expressed as a percentage against competitors. It is easy to compute, it produces a chart, and it is almost useless as a management metric.

The problem is that a mention is not a recommendation, and a recommendation on a question no one asks is not pipeline. An engine can name you in a throwaway list of "other options to be aware of" and your mention count goes up. It can describe you inaccurately and your mention count still goes up. It can recommend a competitor decisively on the one question that precedes every real deal in your category, and a raw share-of-model number will barely move. A metric that rises when nothing of revenue value happens is a vanity metric.

This is the same failure pattern I described in the Revenue Execution Gap: organizations measure the activity that is easy to count rather than the outcome that actually compounds. The fix is not a better tool. It is a better definition of what is being counted.

The Recommendation Ladder

Before you can measure Share of Model honestly, you have to recognize that not all AI presence is equal. When an engine responds to a buyer's question, your company can occupy one of five positions, each closer to a purchase decision than the last. I call this the Recommendation Ladder.

The ladder matters because effort spent moving from Mentioned to Cited feels like progress on a mention-count dashboard, but the deal does not move until you reach Recommended. I have watched a vendor with strong brand recognition sit happily at Rung 3, appearing in every comparison, and lose consistently because the model's actual recommendation always tipped to a rival with sharper proof. Their share-of-model chart looked healthy. Their win rate did not. Most brands optimize to be mentioned. Buyers only act on what is recommended.

Qualified Share of Model: The Metric, Defined

Once you accept the ladder, the metric writes itself. Qualified Share of Model is the percentage of high-intent buyer questions where AI engines actively recommend you, weighted by the buying influence of each question. In one sentence: it measures how often the machine recommends you on the questions that actually precede deals.

Two refinements separate it from raw mention counting. First, it scores position on the ladder, so a recommendation is worth far more than a passing mention. Score each appearance by ladder position — Recommended or Selected counts 1.0, Compared 0.5, Mentioned 0.2, absent 0. Second, it weights each question by buying intent, on a simple high-3, low-1 scale, so being recommended on "best platform for a regulated buyer" counts for more than being named in a generic listicle. Here is the same underlying data scored both ways, on an illustrative ten-question set.

Counted raw, you appear in 9 of 10 answers, a reassuring 90 percent that flatters everyone. Counted as Qualified Share of Model, you earn 5.9 of a possible 16 weighted points, about 37 percent, and the gap between those two numbers is the exact shape of your problem: you are everywhere on low-stakes questions and absent on the high-intent one that decides regulated deals. The raw metric hid that. The qualified metric points your team straight at it.

The Share of Model Measurement Stack

A metric is only as trustworthy as the method that produces it. The Share of Model Measurement Stack organizes that method into four layers, deliberately built so the output is repeatable and defensible to a board rather than a one-off screenshot.

Layer 1, Question Universe. Build a fixed set of 20 to 40 questions drawn from real deals and sales calls, not from a keyword tool. These are the questions your buyers actually ask a model, in their words. The set is fixed so you measure the same thing each cycle.

Layer 2, Citation and Recommendation Capture. Run the set across ChatGPT, Perplexity, Gemini, and Google AI Overviews. For each answer, record your ladder position and whether the description of you is accurate. Capture per engine, because the engines disagree and the disagreement is informative.

Layer 3, Influence Weighting. Assign each question an intent weight based on how close it sits to a real buying decision. This is what converts raw presence into Qualified Share of Model and stops the metric from rewarding listicle ubiquity.

Layer 4, Revenue Attribution. Connect movement in the metric to assisted pipeline, self-reported "where did you first hear about us," and influenced deals. This is the layer that earns the metric a permanent seat in the revenue review rather than a slide in a quarterly deck.

Description Drift: Visibility Without Accuracy Is a Liability

There is one failure mode the ladder does not capture, and it is the one that quietly costs the most. A model can recommend you and describe you wrong, pricing you for the wrong segment, attributing a competitor's weakness to you, or naming a capability you retired two years ago. I call this Description Drift, and it is worth measuring as its own dimension because being visible to a model that describes you wrong is worse than being invisible. An absent vendor loses a chance. A misdescribed vendor loses a chance and plants a false belief the buyer carries into the sales conversation.

Track accuracy alongside recommendation: for every answer that names you, note whether the category, ideal customer, and core claim are correct. A rising recommendation rate with rising Description Drift is not a win. It is a louder version of the wrong story.

Share of Voice vs. Share of Model

It helps to place the new metric against the one it succeeds, because the contrast clarifies what each can and cannot tell you.

The two are not in conflict, but they are increasingly independent. You can buy share of voice. You earn share of model. As more of the buying journey is mediated by machines, the earned metric becomes the leading indicator and the bought one becomes the lagging one.

Who Owns Share of Model

The most common reason Share of Model goes unmeasured is that no one owns it. It lands in the seam between SEO, which thinks in rankings, communications, which thinks in coverage, and product marketing, which thinks in messaging. Each assumes another team has it. None of them report it. The work spans owned content, earned mentions, and technical structure, so it stalls whenever it is treated as a side task.

Give it a single accountable owner in marketing, a defined metric, and a standing place in the revenue operating rhythm, the same governance discipline the AI Buying Committee Framework applies to the demand side. Measurement without an owner is theater.

Why Most Organizations Will Struggle to Improve Share of Model

Here is the uncomfortable part. Once a leadership team can measure Share of Model, the next instinct is to fix it the way they fixed every prior visibility problem: produce more content, buy more coverage, ship more pages. That instinct will fail, and understanding why is the difference between a program that compounds and a year of motion with nothing to show for it.

Answer engines do not reward volume. They reward the things volume cannot manufacture: specific evidence, an original point of view, and corroboration from sources the model already trusts. The work that moves Share of Model is precisely the work most organizations have spent a decade avoiding because it is hard, slow, and cannot be delegated to a tool. A model has already read the median content in your category. You cannot out-publish a machine with the average of its own training data.

In practice, the companies that stall do so for a small set of recurring reasons, and each one is a self-inflicted wound rather than a technology gap.

• Generic AI content. Teams respond to the AI shift by generating more AI-written pages, which is the exact substance models discount because it corroborates nothing they did not already know. Adding it raises your mention count and lowers your distinctiveness at the same time.
• No proof. Claims without evidence do not get cited. Models lift specifics, named methods, and numbers with sources. "Leading provider" is invisible. "A four-layer model with a measured decay curve" is quotable.
• Undifferentiated positioning. If you are interchangeable with three competitors, the model does the sensible thing and summarizes you into the category rather than naming you. Interchangeable vendors get averaged; distinctive ones get cited.
• No original frameworks. With nothing ownable to attribute to you, there is no hook for the model to hang your name on. Proprietary IP is not a branding luxury here; it is the unit of citation.
• Thin authority signals. No named expert stands behind the work, and no independent source corroborates it, so the model has no reason to treat your claims as more reliable than anyone else's.
• Structural inability to become recommendable. Most marketing is built to create awareness, and awareness is not a reason. Recommendation requires a defensible reason a machine can find and verify without talking to you.

I use a short diagnostic with leadership teams to cut through this, and it is deliberately blunt. I call it the Recommendability Test, and it is three questions: Is there a specific reason an engine should recommend us over the nearest alternative? Can that reason be verified without anyone contacting us? Would the model find it in a source it already trusts? If any answer is no, you are not under-marketed. You are uncitable by design, and no amount of publishing volume will change that.

This is the same pattern I described in the Revenue Execution Gap: the constraint is rarely strategy and almost always execution distributed across teams that each assume someone else owns it. The organizations that win Share of Model will not be the ones with the largest content budgets. They will be the ones willing to do the un-scalable work of being genuinely worth recommending, and then make that work legible to the machine. That is a smaller group than the market assumes, which is exactly why the advantage is worth pursuing.

A 30-Day Share of Model Baseline

You do not need a platform or a budget to start. You need a fixed question set, a disciplined cadence, and a place to write down what you see. Here is a pragmatic first month.

Week 1: Build the question universe

• Pull 20 to 40 real questions from recent deals, sales calls, and win/loss notes. Use the buyer's language, not your category jargon.
• Assign each question an intent weight (high, medium, low) based on how close it sits to a buying decision.

Week 2: Capture the baseline

• Run every question across ChatGPT, Perplexity, Gemini, and Google AI Overviews.
• For each answer, record your ladder position and whether the description of you is accurate. This is your starting Qualified Share of Model and your starting Description Drift.

Week 3: Find the gaps that matter

• Isolate the high-intent questions where you are absent or stuck below Recommended. These are your priority targets.
• Flag every instance of Description Drift; some are fixable at the source faster than you would expect.

Week 4: Assign the owner and the cadence

• Name one accountable owner and put Share of Model on the monthly revenue review.
• Set the rhythm: measure monthly, invest in substance and corroboration quarterly, and report movement against baseline, not absolute numbers.

Key Takeaways

• The shortlist is built inside AI, before sales is ever involved, so AI recommendation is a revenue metric, not a content nicety.
• Mentions are vanity. Recommendation is pipeline. Raw mention counting rises when nothing of value happens.
• Use the Recommendation Ladder. Mentioned, Cited, Compared, Recommended, Selected. Deals live in the top two rungs.
• Measure Qualified Share of Model, recommendation weighted by buying intent, paired with description accuracy and assisted pipeline.
• Give it an owner. Unowned measurement is theater; the metric only compounds when someone is accountable for it.

Frequently Asked Questions

Research & Supporting Evidence

The Recommendation Ladder, Qualified Share of Model, the Share of Model Measurement Stack, and Description Drift are original frameworks from ERM Advisory. The market context below is drawn from primary research.

• Forrester, Buyers' Journey Survey (2025): twice as many B2B buyers named generative AI or conversational search as their most meaningful source of information than any other channel.
• Gartner (2026): a majority of B2B buyers prefer a rep-free, digital, self-service research experience.
• Gartner (2025 to 2026): 45% of B2B buyers said they used AI during a recent purchase, and most validate AI-generated insights before acting.
• McKinsey, B2B Pulse (2024): B2B buyers now use roughly ten interaction channels across the buying journey, raising the cost of being absent from any decisive one.
• Google (2025): AI Overviews and AI Mode each reached billion-plus user scale, moving generated answers to the center of search.

Conclusion: Measure the Reader You Cannot See

For most of the history of B2B marketing, every metric we built assumed a human at the other end of the screen. There is still a human buyer, and they still matter most. But there is now a second reader sitting between you and that buyer, one that reads everything, forgets nothing, and quietly decides who gets considered. We have spent two years arguing about whether that reader matters. The argument is over. The question now is whether you can measure it.

Share of Model is that measurement, done honestly. Not a count of mentions that flatters everyone, but a disciplined read of whether the machine recommends you, on the questions that lead to revenue, in language that is accurate. It is the measurement leg of the same system the AI Visibility Architecture and the Answer Engine Optimization work were always building toward, and it belongs in your revenue review alongside the rest of the ERM Revenue Execution System. When your buyer asks the machine who to trust, you will either know your answer or you will be guessing. Start measuring, and stop guessing.

— Erik R. Miller