MODEL ZERO STUDY
We gave six AI models the same advertising brief. They did not see the same opportunity.
AI is increasingly used to inform research, strategy and decision-making. In practice, many of us settle on one model as our default, often without visibility into what that model may leave unseen. We wanted to measure the effect of that choice.
We gave the same Australian advertising brief to six leading AI models and compared what each surfaced, what each omitted and where their recommendations differed.
We call this Model Zero: the AI model you default to.
This study examines how much that choice changes what enters the decision.
Explore the study
THE BRIEF
Launch campaign for a mid-priced electric SUV in Australia, targeting first-time EV buyers aged 30-49 in metro Sydney and Melbourne. Budget A$1.2m over eight weeks. Objective is test drive bookings, with a secondary brand consideration goal. The client wants proof of incremental reach beyond their existing search spend.
VERDICT
Single-model AI use creates material blind spots. 49% of observed ideas appeared in only one of six models responses.
The risk is not simply a different answer. It is the planning ideas and assumptions a single answer leaves invisible. The evidence below shows how much each model missed and where those omissions changed the plan.
49%
of ideas appeared in only one model response
Nearly half the ideas observed were surfaced by just one of the six models.
30%
of majority-backed ideas were missed by a typical model
The median blind spot was 30%, ranging from 8% to 54% depending on the model.
56.5%
overlap between Western and Chinese model groups
The two groups surfaced similar volumes of ideas, but shared only 56.5% of their content.
ANALYSIS OF MODEL OUTPUTS
The six models received the same brief but produced materially different plans. Average overlap between any two responses was 55.3%, while 49% of all distinct ideas appeared in only one model response.
This matters for anyone using familiar AI tools such as Gemini, ChatGPT or Claude as their default. In this study, their blind spots varied significantly. Gemini omitted 54% of the ideas supported by most models, Claude omitted 34%, while GPT omitted 9%. The point is not that one provider is inherently better. It is that choosing a single model also means inheriting that model's particular omissions.
Those differences affected real planning decisions. Gemini, for example, did not surface TikTok, influencer and creator content, or email/CRM nurture. Claude, GPT, DeepSeek, Qwen and Kimi each recommended at least two of these. A planner relying only on Gemini would therefore have considered a materially different channel mix.
The models also handled uncertainty differently. Qwen asked for clarification on whether the A$1.2 million budget was media-only or all-in, and whether dealers could absorb the projected test-drive volume. Gemini proceeded without resolving those assumptions.
The practical issue with single-model AI use is therefore difficult to detect from the answer itself. The response may appear complete, but you cannot see the ideas, questions or alternatives that your chosen model never surfaced.
Single-model exposure
Relying on one AI means accepting what that model leaves out
What each model left unseen
Share of majority-backed ideas absent from each model's response
8% missed
9% missed
26% missed
34% missed
52% missed
54% missed
No model surfaced the complete picture. Blind spots ranged from 8% for Qwen to 54% for Gemini. A user relying on any one model would have no indication from the response itself of what had been left out.
Also observed
55.3%
Average overlap between any two models
24
Different budget allocations from the same brief
1
Direct contradiction across the analysis
What this means: The study does not establish which model is “right”. It shows that the model you choose materially changes what enters the decision. When one AI becomes the default, its blind spots can become your blind spots too.
This is the Model Zero problem.
WESTERN AND CHINESE MODEL GROUP ANALYSIS
Similar volume. Different perspective.
Western and Chinese models surfaced a similar number of ideas, yet the two groups shared only 56.5% of their content. Model family therefore changes not just how much you see, but what you see.
WESTERN MODELS
Gemini, Claude and ChatGPT
Gemini 15% · GPT 49%
CHINESE MODELS
DeepSeek, Qwen and Kimi
Kimi 17% · Qwen 77%
SIX MODELS, ORCHESTRATED
Syntria™ compares perspectives across all six models rather than relying on one model as the default.
The finding: Western and Chinese models surfaced similar volumes of ideas, but overlapped on only 56.5% of their content. Similar coverage does not mean the same perspective.
Percentages represent share of the 604 distinct ideas observed across all six responses. They measure coverage, not accuracy or quality.