Share of Search measures how much of a category’s search demand belongs to your brand. Share of Model measures how often AI assistants name your brand when someone asks a buying question in that category. One reads demand. The other reads recommendation.
A founder told me last month that his brand health had never looked better. Branded searches up. Direct traffic up. Then he asked ChatGPT to shortlist vendors in his category, and someone else’s name came back three times running.
Both things were true at once. People were looking for him. The machine that increasingly does the looking for other people was not passing his name along.
That gap has a shape, and two metrics sit on either side of it.
Two questions that sound alike and are not
“How many people search for us” and “does AI mention us when people ask” feel like the same question wearing different clothes. They are not.
The search question counts intent that has already formed. Someone wanted your category, typed something, and a share of those typers wanted you specifically. It is a demand reading.
The AI question happens earlier and lower down, inside the assistant, before the user has a shortlist at all. The model is doing the narrowing. Whether your name survives that narrowing is a different measurement entirely, and it is what AI visibility is trying to describe.
What Share of Search actually measures
Share of Search is not a folk metric. Les Binet, Head of Effectiveness at adam&eveDDB, presented it at the IPA’s EffWorks Global 2020 conference. The definition is tight enough to argue with: total searches for a brand divided by total searches for all brands in that category.
The data comes from Google Trends. It is free, it reaches back to 2004, and it can be read weekly.
What made the room sit up was the lead. Binet’s testing across cars, energy and mobile phones found Share of Search running ahead of share of market, with a lag that could stretch to a year in the car category. A brand’s search share was telling you where its market share was heading before the market got there.
Twelve months of warning, from a free dataset. That is why the metric spread. The IPA still keeps a running library of Share of Search work, and Binet himself was careful to say it was no silver bullet.
What Share of Model actually measures
Share of Model asks a narrower question. Across a fixed set of buying prompts in your category, how often does an AI assistant put your brand in the answer?
You fix the prompts, you run them across the assistants your buyers actually use, and you count. Two numbers fall out. Coverage is how many of those prompts mention you at all. Depth is how strongly you land when you do appear: named first, described accurately, recommended rather than listed.
A brand can score wide and shallow, appearing everywhere as an afterthought. Another can score narrow and deep, owning three prompts completely and vanishing from the rest. Those are different problems with different fixes, which is why the single headline number is worth less than the pair underneath it.
Cousins, not twins
Both metrics are measuring a share of attention. That family resemblance is real, and it is why Share of Model reads as familiar to anyone who has used the older metric.
The difference is who is doing the filtering.
Share of Search reads a human decision that has already happened. A person chose to type your name. The metric is downstream of preference, which is exactly what makes it predictive.
Share of Model reads a machine decision that happens before the human sees anything. The assistant assembles an answer from sources it trusts, and being cited by AI is the entry ticket. If your name is not in the pool the model draws from, no amount of brand affection downstream can put it there.
One measures whether people want you. The other measures whether the thing standing between people and their shortlist knows you exist.
Why the decision layer moved
For twenty years the decision layer was the results page. You ranked, the user scanned, the user clicked, the user decided. Every SEO metric was built around that sequence.
AI Overviews and assistant answers broke the sequence. The synthesis now sits above the links, and for a growing share of questions it is the only thing read. The click that used to carry the decision often never happens.
When the decision moves, the measurement has to move with it. Ranking still matters, but ranking is now a proxy for a step the user may skip. Share of Model measures the step they do not skip. The two disciplines sit side by side rather than in sequence, which is the whole argument in SEO versus GEO.
Which one should you track
Both. They answer different questions and neither substitutes for the other.
Share of Search tells you whether demand for your brand is holding up inside its category. It is your early warning on brand health, and it still leads market share.
Share of Model tells you whether that demand can survive contact with an AI assistant. A brand with rising search share and falling model share is being wanted and not recommended, which is a specific and fixable condition.
Read together they triangulate. Demand strong and recommendation weak points at your source footprint, not your brand. Demand weak and recommendation strong means the machines like you more than the market knows you yet. Which of those two problems you have changes what you do on Monday.
If you want to see how the second number is built cluster by cluster rather than keyword by keyword, that is the argument in the beachhead approach to GEO. The map those two numbers sit on, coverage against depth, is in the two dimensions of AI visibility.
- Share of Search reads demand that has already formed. Share of Model reads the recommendation layer that forms it.
- Binet’s IPA work showed search share leading market share by up to a year in some categories.
- Share of Model splits into coverage and depth, and the split matters more than the headline.
- Rising search share with falling model share is a source problem, not a brand problem.
References
Frequently asked questions
Does Share of Model replace Share of Search?
No. Share of Search still leads market share and still reads brand health from free public data. Share of Model reads a layer that did not exist when Share of Search was designed. Dropping one for the other trades a working instrument for a newer one that measures something else.
How do you calculate Share of Model?
Fix a set of buying prompts a real customer would ask in your category. Run them across the assistants your buyers use. Count how often your brand appears, and record how it appears. The appearance rate is coverage, the quality of the appearance is depth.
Can Share of Search predict Share of Model?
Not reliably. They draw on different inputs. Search share comes from what people type, model share comes from what the assistant found while reading the web about your category. A brand can be searched often and quoted rarely, and that combination is common.
How often should Share of Model be measured?
Often enough to see movement and not so often that you are reading noise. Model answers vary between runs, so a single reading is close to meaningless. A repeated measurement on the same prompt set is what turns it into a trend.
What moves Share of Model fastest?
Being present in the sources an assistant already trusts for your category. That usually means independent mentions, third-party listings and pages that answer the exact question rather than circle it. Your own site matters, but it is one voice among the many the model reads.