Share of Model: The Formula, the Vanity Charge, and the One Fix

The CMO of an AI writing company called it a vanity metric in July 2026, in a guide written for the exact marketers who had just started counting how often assistants name their brand. Half of that charge is fair. The other half is the reason this metric is worth keeping, and almost nobody separates the two.

This is for the marketing leader who now has an AI visibility figure in a deck and a board member asking what it is for. What follows is a view, not a neutral summary: what the number is good for, what it cannot see, and the one reporting change that makes it defensible. Every fact below links to where it came from.

A metric arrived before anyone agreed what it counts

Teams run a prompt list, count the answers that name them, and call the result share of model. In Writer's guide to generative and answer engine optimisation, chief marketing officer Diego Lomanto dismisses it as "a vanity metric" and argues that presence inside the way work actually happens is the durable thing instead. Writer sells enterprise AI, so its leadership has a commercial view of where value sits. The argument still deserves taking seriously, because the alternative on offer is a counting exercise with no agreed method.

How unsettled is the method? The wiki entry at the Agile Brand Guide states plainly that no single ratified calculation exists yet, and that practitioners track frequency, prominence and favourability as separate levers. That is the real condition of this metric in 2026: a name everyone uses, with the arithmetic still being argued about.

The formula, and the four settings that decide what it says

The common form is simple. Parse's guide gives it as mentions of your brand divided by total category brand mentions, multiplied by a hundred, which is share of voice with answers in place of articles. The Agile Brand Guide reaches the same shape from a different direction, dividing your mentions by all brand mentions in the same set of responses (Agile Brand Guide).

The arithmetic is not where the number gets decided. Four setup choices are. Parse recommends a category prompt set of fifty to five hundred questions, with most practitioners settling between fifty and two hundred, at least three samples of each prompt on each platform, a temperature of zero wherever the interface allows it, and a minimum of ChatGPT, Google AI Overviews and Perplexity (Parse). Change any one of those four and your share of model changes without your brand changing at all. That is why the run count belongs inside the metric rather than in a footnote.

The half of the vanity charge that lands

Two criticisms are correct and worth conceding fully.

The first is instability. Parse reports that roughly 30% of brands hold their visibility from one monitoring run to the next (Parse), which means a reading that moves a couple of points between months has told you nothing you can act on. A figure with no sample size beside it is decoration.

The second is that a count has no opinion. Being named as the budget option and being named as the right fit for a regulated finance team are different commercial outcomes, and the metric scores them identically. Nothing in the formula distinguishes a recommendation from a mention in a list of also-rans.

The half that does not

Where the dismissal fails is the implication that the number carries no information. Ahrefs looked at 75,000 brands in May 2025 and ranked the signals that track with visibility in Google's AI Overviews. Branded web mentions came first, at a Spearman correlation of 0.664, ahead of branded anchors at 0.527 and domain rating at 0.326. The same analysis found brands in the top quarter for web mentions averaging 169 AI Overview appearances against 14 in the quarter below. The author is careful that correlation is not causation and that every coefficient in the set is moderate at best.

Read conservatively, that still says something useful: how much the web talks about you travels with how often machines name you. Share of model is the cheapest available reading of that. As a thermometer it works. As a target to optimise, it fails, because the thing it correlates with is not under your direct control.

One blended figure hides the only finding worth having

The sharpest evidence against a single number comes from a case the Agile Brand Guide records: an INSEAD analysis of the detergent brand Ariel found it holding close to 24% share of model on Meta's Llama and under 1% on Gemini. One brand, one category, two answers that are not in the same universe. Average those and you get a figure that describes neither.

Parse adds the mechanism behind that spread, citing BrightEdge at second hand: platforms disagreed on brand recommendations for 61.9% of identical queries, and a typical answer has room for very few names, with roughly 2.37 brands per query on ChatGPT and Perplexity against 6.02 in Google AI Overviews. Treat those as reported rather than verified, since the original study's date is not given. The direction is what matters: the slots are scarce and they are allocated differently by each engine.

Report it per engine. A blended average is the vanity version.

So here is the call. Share of model is not a vanity metric. A single cross-engine share of model is, and that distinction is the whole argument. The evidence is the Ariel split: a brand can be a quarter of the conversation on one model and invisible on another (Agile Brand Guide). A blended number averages a success and a failure into a figure that recommends no action, which is the working definition of vanity. The per-engine version tells you which engine to go and look at, and that is a task somebody can own by Friday.

The honest counter-argument is executive patience. A board wants one line, and four numbers in a row invites four questions nobody has time for. That is a real cost, and the answer is not to blend them. Report the worst engine and the best engine with the gap between them, which is one line and one sentence.

This call is wrong under one condition: if the engines converge, and the same prompt set starts returning similar brand sets across platforms, then a blended figure stops hiding anything and the extra rows are noise. The BrightEdge disagreement rate that Parse cites suggests convergence is not close (Parse). If that rate drops by half, revisit this.

The dashboard line that survives a challenge

If you get one row in the monthly report, build it so each part can be attacked separately. Something like: present in eleven of twenty buyer prompts, three runs each, ChatGPT and Perplexity reported apart, eighty answers in total, descriptions accurate in nine of the eleven, and your share among five named rivals stated per engine with the spread.

Then add the commercial column beside it: visits referred by assistants, and deals where the buyer named one. Both will be small and jumpy to begin with, and the referral side will understate you badly. In a March 2025 browsing panel, Pew Research Center logged a click on a link inside a Google AI summary in 1% of visits where a summary appeared, with clicks on standard results falling to 8% from 15%. Report the small numbers anyway. A measure nobody looks at never improves, and logging it by hand for a quarter will teach a team more than any single study will.

A young metric, honestly reported

Track the temperature per engine, with its sample size, and spend the effort on the thing that moves it: being talked about in places the engines already read, and being genuinely useful where your buyers work. Our pieces on connectors as the new listing slot and on brand as the thing an agent falls back on both argue the same point from different ends. And hold every claim in this field loosely, including this one. The method is two years old and the benchmarks are borrowed.