Share of model is how often an AI engine names you among the options for a question, measured across a fixed set of questions run repeatedly. If you ask forty category questions and you appear in ten of the answers that named anybody, your share of model is twenty five percent.
It is a genuinely useful idea and the term is being ruined at speed, mostly by tools reporting a number with no published method behind it. The idea is worth keeping. The number is worth distrusting until you know how it was made.
Where the idea comes from
It is share of voice with a different substrate. Marketing measurement has used a family of share metrics for decades on a consistent finding, which is that a brand’s share of category attention tends to track its share of category sales, and that gaps between the two predict which direction share is about to move.
The most recent useful version of this is share of search, which the IPA and the practitioners around it popularised as a cheap, fast-moving predictor of market share. The appeal was that it used a public behavioural signal rather than a survey, and it moved early enough to be worth acting on.
The underlying theory sits in the Ehrenberg-Bass tradition, which holds that buyers choose from a small set of options that come to mind rather than evaluating the whole market. Being in the set is most of the battle. An AI engine does mechanically what a buyer does psychologically, which is why the metric transfers as cleanly as it does.
Why it is worth measuring at all. Almost everything else people report about AI visibility is either a count of mentions, which rewards being talked about badly, or a proprietary score, which rewards whoever built the scoring. Share of a named option set is at least a ratio with a meaningful denominator.
How to compute it without kidding yourself
Four decisions determine the number, and every one of them is an opportunity to flatter yourself.
The question set. Fixed, written before you look at any results, and representative of what buyers ask rather than what you would like them to ask. Twenty to forty questions. This is where the most inflation happens, because a question set weighted toward your strengths produces a wonderful number that predicts nothing.
The denominator. Count only answers where at least one vendor was named. Including answers that named nobody drags everyone’s share down and makes the metric move when engine behaviour changes rather than when your position does.
The counting rule. Decide in advance whether being named third counts the same as being named first, and whether being named as a caveat counts at all. We count presence rather than position, and record position separately, because presence is what is achievable and position is largely not.
The engines. Report per engine, never blended. A blended figure hides the single most most useful fact available, which is that you are strong in one engine and absent from another, and those have different causes. The reasons they differ are in how Perplexity chooses its sources.
share of model (per engine)
answers naming you
= ----------------------------------------
answers naming at least one vendor
fixed at: 40 questions, 2 runs each, logged out
counted as: presence, not position
reported: per engine, with the question set publishedThe four ways the number gets inflated
If you are shown a share of model figure by a tool or an agency, these are the things to ask about, in order of how much they move the result.
| Method choice | Effect on the number | What to ask |
|---|---|---|
| Questions chosen after seeing results | Large. Easily doubles it | Was the set fixed before the first run? |
| Branded questions included in the set | Large. You always appear | How many questions name us? |
| Answers naming nobody left in the denominator | Moderate, and unstable over time | What is the denominator exactly? |
| One run per question | Unpredictable in both directions | How many runs, and how is variance handled? |
| Engines blended into one figure | Hides the detail you could act on | Can I see it per engine? |
The fix for all five is the same and it is not technical. Publish the method alongside the number, fix it in advance, and treat changes to the method as a reset of the series rather than an improvement in the score. This is ordinary measurement hygiene of the kind the open science community has spent years arguing for, and it applies here for exactly the same reason. A flexible method will always find a result somebody wanted.
A number you cannot reproduce is not a measurement.
If a dashboard gives you a share of model figure and cannot tell you the question set, the run count and the denominator, you are being shown a rendering of somebody’s business model rather than a fact about your company. That is a harsh way to put it and we have not found a softer one that is still true.
Our own method is published in full, including the parts that make our numbers look worse. Book an audit call if you want a baseline built that way, or just take the method and run it yourself.
A worked example, with the awkward part left in
Take a company selling incident management software. Forty questions, four engines, two runs each, which is three hundred and twenty answers. Of those, two hundred and sixty named at least one vendor and sixty named nobody, usually because the question was too vague or the engine declined.
The company appeared in seventy eight of the two hundred and sixty. That is a share of model of thirty percent, which sounds respectable until you split it. In Perplexity it was forty four percent. In ChatGPT without search it was eleven percent. Those two numbers describe completely different problems and the blended thirty describes neither.
The awkward part is what happened next. Of the seventy eight appearances, fifty one came from the twelve questions that carried a constraint, meaning the company was winning narrow questions and losing broad ones. That is the correct position for a challenger and it looked like underperformance on the headline number. Reporting the blended figure to a board would have made a healthy programme look like a failing one, which is the practical case for never blending.
What it is good for, and what it is not
Good for direction over quarters, on a fixed set, per engine. It is the closest thing available to a competitive position measure in AI answers, and tracked honestly it will tell you whether the work is compounding.
Not good for month-to-month reporting, because variance between runs will swamp real change at that timescale. Not good as a single company-wide number, because it blends questions you can win with questions you cannot. And not good as a target, because it is trivially gameable by adjusting the question set, which means the moment somebody’s bonus depends on it the number stops measuring anything.
The deeper limitation is that share of model measures presence among options, which is only one of three things an engine can do with your company. Being mentioned, being cited and being recommended are different outcomes with different causes, separated in mention, citation, recommendation. A rising share of model with no change in recommendation rate means you have become an option rather than a choice.
What moves it
In rough order of speed. Crawler access, which is binary and fast. Passage extractability, which is an editing pass. Then discrimination, meaning pages that answer constrained questions nobody else has answered. Then corroboration, which is slow and mostly outside your control.
What does not move it is publishing volume on your own domain. Forty more posts saying what your existing posts say adds nothing to the denominator you are competing in, because the engine is choosing among companies rather than among pages. The stages this runs through are in indexed, retrieved, cited, and most share of model problems are actually stuck at the first two.
One honest caveat about benchmarks. There is no credible cross-industry benchmark for share of model and you should be suspicious of anyone offering one, because the figure depends entirely on the question set and every category has a different number of plausible vendors. In a category with four real players, twenty five percent is par. In one with forty, it would be extraordinary. Your only meaningful comparison is your own previous measurement under an identical method.
Related reading
- Our full testing methodology. The protocol this metric is computed from, published so you can argue with it.
- The twenty prompt set. A smaller fixed set you can run yourself this week.
- What AEO is. The wider picture this metric sits inside.
- Training data versus live retrieval. Why share of model moves at different speeds for different question types.
Before you adopt the metric, write the question set down and date it.
That single act does more for the honesty of this number than any tooling. A set fixed in advance, run the same way each quarter, will tell you something true even if your method is otherwise crude. A set assembled after seeing results will flatter you indefinitely.
If you want a baseline built under a published method you can inspect and keep, book a visibility audit. The output is yours regardless of whether we work together.
Shaban runs answer engine optimisation for B2B SaaS companies at Uncited Brands. He works the unglamorous end of the problem, mostly crawler access, retrieval diagnostics and measurement baselines, and publishes the tests that do not go his way alongside the ones that do.