The newest AI marketing metric is designed to make executives feel informed. That should make everyone nervous.
Share of model measures how often an AI system mentions, recommends, or ranks a brand when people ask questions in a category. It sounds like the natural successor to search share of voice. Agencies are building dashboards around it. Vendors are promising optimization. Marketers are adding it to the weekly report.
The problem isn't that the metric is useless. The problem is that it is easy to improve without improving the business.
The number looks cleaner than reality
AI visibility is messy by design. Ask the same question twice and you may get a different answer. Change one word and the model may pull from a different set of sources. Add a location, a budget, a customer type, or a product constraint, and the brand list can change again.
That makes a single share-of-model percentage a strange thing to treat as a stable business signal. It is not a shelf position. It is a snapshot of a probabilistic system responding to a prompt under specific conditions.
The IAB's new AI visibility measurement guidance is a useful step because it recognizes the fragmentation. Marketing Dive reported that the trade group found more than 20 vendors offering AI visibility measurement tools, with little consistency between them. That isn't a minor reporting inconvenience. It means two teams can measure the same brand and produce different answers without either team making a math error.
The first question for any share-of-model report should be simple: what exactly was measured?
Which models? Which prompts? Which regions? Which dates? Which answer types? Were citations counted? Were paid placements separated from organic recommendations? If the vendor can't show the sample, the score is closer to a mood ring than a measurement system.
Visibility isn't preference
A brand can appear in an AI answer for reasons that have nothing to do with preference. It might be included as a warning. It might be named because the model is comparing it with a competitor. It might be cited in a source list that nobody opens. It might be recommended for one narrow use case while being ignored for the purchase that actually matters.
Mention is not consideration. Consideration is not trust. Trust is not conversion.
That chain is where many AI visibility programs get lazy. The report climbs from 12 percent to 19 percent, and everyone calls it progress. Nobody checks whether the new mentions are accurate, positive, relevant, or connected to qualified demand.
This is the same measurement drift I wrote about in AI visibility measurement's biggest blind spot. The industry keeps inventing sharper labels for signals that still sit several steps away from revenue.
The answer isn't to ignore visibility. It is to split it into useful layers.
A serious scorecard should separate at least four questions:
- Presence: Does the brand appear for the prompts that matter?
- Position: Is it listed early, or buried in a long comparison?
- Accuracy: Does the answer describe the brand correctly?
- Action: Does the exposure lead to a click, conversation, visit, trial, or sale?
A fifth layer matters for brand teams: context. The same mention can help or hurt depending on the sentence around it.
The optimization trap
Once a metric becomes visible, teams start optimizing for it. That is normal. It is also where the trouble starts.
A brand can chase share of model by publishing more pages, repeating the same claims, adding formulaic comparison content, or flooding the web with low-value material that models may encounter during retrieval. The metric can rise while the brand becomes less distinctive and less trustworthy.
That is not a hypothetical risk. Google's own people-first content guidance makes the point in a different way: content created mainly to manipulate discovery tends to be a poor experience for people. AI systems are not magic filters. If the open web gets noisier, the inputs get noisier too.
The better move is to improve the evidence a model can find, not to manufacture more surface area.
That means making product facts clear. Keeping pricing, availability, policies, and qualifications current. Giving experts a real byline. Publishing original data that others can cite. Correcting false claims in the places where they appear. Building a brand that customers describe consistently because the experience actually deserves the description.
Some of that work is unglamorous. It also survives algorithm changes better than prompt hacks.
A useful operating rule is this: don't celebrate a share-of-model gain until you know what caused it and what changed downstream. If the answer is only “we published 40 new pages,” you haven't found growth. You've found an input.
What the report should include
The first version of an AI visibility dashboard shouldn't try to recreate an entire media mix model. It should make uncertainty visible.
Start with a fixed prompt library tied to real customer decisions. Include category questions, comparison questions, problem questions, local questions, and branded questions. Store the exact wording, date, model, region, and answer. Do not let a vendor quietly change the test set while preserving the historical chart.
Then add a human review layer. Automated scoring can identify mentions, sentiment, citations, and position. People still need to judge whether the answer is useful, current, and commercially meaningful.
The report should also show confidence ranges. A move from 18 percent to 21 percent may sound important, but if the sample is small and the answers vary widely, the honest interpretation may be “no clear change.” Marketing dashboards rarely reward that sentence. They should.
Finally, connect visibility to behavior wherever the path exists. Add tagged links to cited resources. Watch branded search patterns. Compare qualified traffic, assisted conversions, lead quality, and sales conversations across the same period. In a long-cycle business, the signal may arrive as a better sales call before it arrives as a clean last-click conversion.
This is where the work touches the broader failure pattern behind AI agents that hallucinate on live data. The system can produce a confident answer, and the dashboard can produce a confident number. Neither confidence proves the underlying claim.
The metric needs a job
Jellyfish's current expansion of its “Share of Model” strategy shows where the category is heading. The company is tying AI visibility insights to paid media recommendations across Google Performance Max, ChatGPT, TikTok Search, DV360, YouTube, and Reddit Ads. Its pitch is practical: use model insights to decide where to act, then connect those actions to campaign performance.
That is more useful than admiring a visibility score in isolation. Still, even the reported claim of a 30 percent improvement in ROAS needs the usual questions. For which brands? Against what baseline? Over what period? With what spend and attribution model?
The metric doesn't have to prove everything. It just needs a job.
Use share of model to find where the brand is absent, misunderstood, or poorly represented. Use human review to decide why. Use content, product, customer experience, and media changes to address the cause. Use commercial outcomes to decide whether the work mattered.
That sequence keeps the score in its place. It becomes a diagnostic, not a trophy.
The uncomfortable part
Marketing has a long history of turning useful signals into status symbols. Follower counts became influence. Impressions became impact. Traffic became demand. Now AI mentions are lining up for the same promotion.
The temptation will be strongest because the market is still new and executives want a number they can repeat. A share-of-model score fits neatly into the slide. It suggests control at exactly the moment when the systems feel least controllable.
But visibility is only the beginning of the customer journey. A model can mention you. A person still has to believe you.
The teams that win here won't be the ones with the prettiest AI dashboard. They'll be the ones that can explain what changed in the answer, why it changed, and whether anybody made a better decision because of it.
For now, treat share of model as a lead, not a verdict. The number may tell you where to look. It cannot tell you what you mean.