Share of Model Is the New Market Share

You rank third for "merino base layer for hiking." Someone opens an AI assistant, asks roughly the same thing in a full sentence, and gets five brands back. You are not one of them.

Nothing broke. Your position did not move. Traffic looks like last week. There is no report anywhere in your stack called "did not get recommended."

That blind spot now has a name, and it has started turning up on the same slide as market share.

Where the term came from

The term is credited to Jack Smyth at Jellyfish, and the agency launched a platform under that name on December 5, 2024, running betas with Danone and Chivas Brothers. His colleague Tom Roach made the broader case in Marketing Week: if buying decisions increasingly route through a model, your standing inside that model is an asset worth measuring.

Strip the agency framing and it is arithmetic. Take a fixed set of questions a shopper in your category would actually ask an assistant. Run them. Count how many of the answers include you. Divide.

Everything after that is implementation detail.

Rankings do not carry over the way you would hope

The obvious objection: if I rank well, surely the model picks me up anyway.

Less than it used to. Ahrefs re-ran its study in March 2026 across 863,000 keywords and 4 million AI Overview URLs and found 38% of cited pages ranked in Google's top 10 for the same query. In July 2025 the same measurement came out at 76%. The rest split nearly evenly: 31.2% ranked somewhere between 11 and 100, and 31% ranked nowhere at all.

Ahrefs is upfront that some of the drop is better detection on their end, plus query fan-out, where Google splits the original question into sub-queries and cites whatever wins those. So do not read 38% as a clean collapse.

Read it as this. The top ten is now one input among several, and about a third of what gets cited never ranked for the query in the first place. Your position report cannot tell you whether that third includes you. This is the working difference between AEO and SEO, and it is why the two need separate scoreboards.

Three states, not one

Before counting anything, decide what counts. An AI answer can do three different things with your store, and merging them produces a number that feels good and means very little.

• Mentioned. Your brand name appears in the text. No link.
• Cited. A URL on your domain shows up as a source, usually the homepage or a collection page.
• Product cited. The answer links the specific product page for the thing it just recommended.

Track all three separately. A mention with no link is real but weak. A product page citation is the one that puts a shopper a single click from a buy button. Most dashboards flatten these into one "visibility score," which is how stores end up celebrating a number that never moved a sale.

Building the prompt set

The prompt set is the measurement. Get it wrong and everything downstream is theatre.

Write questions, not keywords. Nobody types "merino base layer womens" into an assistant. They type "what should I wear under a shell for winter hiking in Scotland, I run cold and I hate wool that itches."

Thirty to sixty prompts is enough, spread across four types:

• Open category. "Best X for Y." You lose these first and win them last.
• Constrained. Budget, size, material, use case, shipping region. This is where a small store realistically places.
• Comparison. You against two named competitors.
• Brand direct. "Is [your store] any good." Tests whether the model knows you exist at all.

Then freeze the list. The moment you edit prompts mid-quarter, the trend line is garbage and you will not notice for two months.

The noise problem is worse than people admit

Run the same prompt three times and you can get three different brand lists. That is not a flaw in your tracking. Sampling is probabilistic by design, which is also why the models can write.

Two consequences follow.

One screenshot proves nothing. If a colleague sends you an assistant recommending a competitor and calls it a trend, that is one draw from a distribution.

And the unit of analysis is the run, not the answer. Same prompts, same assistants, same day of the week, then compare run over run and smooth across four weeks. Weekly is enough. Daily buys you more noise at higher cost.

Log the date on everything. Models ship updates without announcing them, and a step change in your numbers is at least as likely to be a model release as anything you did.

What good looks like

Nobody can hand you a benchmark, and anyone who does is selling one. Category size, competitor count and query mix move the number so much that a cross-industry average tells you nothing about your store.

Your own baseline is the benchmark. Measure for four weeks before you change anything and call that zero.

Expect the four prompt types to move in a particular order, because each one depends on something different. Brand-direct prompts move first, sometimes within weeks, since they mostly test whether the model can find and read anything about you at all. Constrained prompts move next, over a couple of months, because they turn on specifics that are either present in your product data and in third-party content or are not. Open category prompts move last, if ever, since they are dominated by roundups and forum threads you do not control.

Give it three to six months before you judge the programme. Not because that is a comfortable number, but because model release cycles, indexing and third-party publication all run on roughly that clock. A four-week read is weather, not climate.

What we would not claim

That share of model maps cleanly onto revenue. It does not, yet. Adobe's data ties AI-referred traffic to strong conversion, but the chain from "appeared in more answers" to "made more money" has too many unmeasured links to model honestly. Treat it as a leading indicator with a plausible mechanism, the way share of search was treated a decade ago. Good for direction. Not a line item.

What it is unambiguously good for is catching absence. If you appear in 4% of answers in your category and your two nearest competitors sit at 40%, you do not need a regression to know you have a problem. You need a way in.

If you would rather not maintain that spreadsheet by hand, Seokai's AI Visibility tracker runs a fixed prompt set on a schedule and reports mentions, citations and product-page citations as separate numbers.