When AI stands between brands and buyers
The brand you built
is not always
the brand AI sees.
Digital Twin of Brand Equity investigates how AI systems represent brands, bring them into consideration, compare alternatives, and participate in choice.
The framework
How AI mediates the path from brand meaning to action
A tested AI answer reflects the system, the task, learned associations, retrieved or supplied evidence, available tools and product rules. The resulting account can shape the options a person encounters before a direct interaction with the brand.
Information
What evidence can the machine encounter?
Representation
What does the system conclude about the brand?
Access
Which information, products, services and environments can it reach?
Action
What is it able and authorized to do?
Information, Representation, Access and Action describe four system layers. DTBE separately tracks four commercial outcomes: Discovery, Representation, Evaluation and Choice. Representation appears in both lists because the account formed by a system can become an outcome for a buyer. These layers and outcomes do not establish a single linear causal pathway. A brand can be represented accurately but remain inaccessible to an agent. It can be accessible but evaluated poorly. An agent can identify the right brand and still lack permission to act.
Core concepts
The Concept
Library
A vocabulary for studying the brands AI systems construct.
Explore all conceptsSemantic Entity
The picture of a brand that appears in an AI answer to a particular question.
Semantic Coherence
Whether a brand’s core facts fit together across AI answers, even when the details differ.
Statistical Availability
How often a brand appears when we repeat a defined AI test. Praise, position and choice are measured separately.
Agentic Narrowing
An observed reduction in the options an AI system presents. Human choice requires separate evidence.
Representation Inertia
When an older account of a brand keeps returning after the relevant evidence changes.
Brand Wake
The public trail left by a company’s actions and the responses directly connected to them.
Public Signal Field
The public conversation and wider public information around a brand at a given time.
Vector Safety
Whether AI connects accurate facts and associations to the right brand.
Essays
Read beyond
the headlines.
Arguments, observations, and field notes on brand strategy in an AI-mediated world.
Browse the essaysAI Agents Are Only as Powerful as What They Can Access
Amazon, Meta Muse, and the coming war over who controls the data AI agents can reach
What is the Semantic Entity?
Why the same brand can be represented differently across tasks and systems.
The Brand You Built Is Not the Brand AI Sees
A company’s intended identity competes with the evidence AI systems encounter about it.
Research in focus · Response variance
One answer is only
one observation.
Ask the same question again. The wording can change, the brands can change, and the first name can change. Repeated observations reveal which parts of a brand’s AI representation hold steady.
Observed brand mentions · 1,500 responses
More detail. Shorter lists. More disagreement.
In this CRM study, adding a business profile and operating constraints increased variation in which brands appeared across repeated answers.
Disagreement between brand sets
Mean pairwise Jaccard distance: 0 = identical sets, 1 = no shared brands. Each prompt was repeated 500 times.
95% bootstrap interval: 0.187–0.216
Secondary literal-mention analysis. The interval resamples whole randomized triplets.
Which brand appears first?
First named is a position measure, not necessarily an endorsement. HubSpot appeared somewhere in all 500 answers in every condition.
Inspect the exact prompt and mention counts
What is the best CRM for a small to medium-sized business?
| Brand | Broad question | Business profile | Operating constraints |
|---|---|---|---|
| HubSpot | 500/500 | 500/500 | 500/500 |
| Pipedrive | 500/500 | 500/500 | 500/500 |
| Zoho | 500/500 | 500/500 | 495/500 |
| Salesforce | 500/500 | 500/500 | 340/500 |
| Microsoft Dynamics | 423/500 | 414/500 | 40/500 |
| Freshsales | 491/500 | 65/500 | 245/500 |
| Copper | 1/500 | 100/500 | 256/500 |
| Monday | 61/500 | 42/500 | 5/500 |
Study design, evidence and interpretation limits
This frozen publication snapshot includes all three completed October 6–9 variance-related experiments: the 500-response CRM baseline, the 1,500-response randomized CRM specificity study (500 per condition), and the 20-response live-search boots pilot (10 per condition). Collection counts were checked against captured observations on October 9, 2026. No additional model calls were made for this exhibit.
CRM studies used GPT-5.6 Luna with medium reasoning and search disabled. The specificity study randomized the three prompt conditions within each of 500 triplets. Its interval uses 10,000 whole-triplet bootstrap resamples; response pairs are not independent observations. The baseline is historical replication context, not a randomized control for the later study.
Charts show literal mentions, first-named position, exact text uniqueness and recorded source counts. Mentions can include criticism. Semantic recommendation coding remains provisional and is not presented as a confirmed finding. Automated extraction and AI-assisted checks have not received independent human adjudication; intervals omit coding uncertainty.
The tested constraints combine budget, software environment and staffing needs. These results cannot establish a universal effect of prompt specificity, generalize to every model or consumer chat interface, or demonstrate business outcomes. More specific prompts changed both the question and its response distribution.
Studies: CRM baseline, collected October 6, 2026; randomized CRM specificity study, collected October 7, 2026; live-search boots pilot, collected October 9, 2026. Analysis snapshots: October 6 and 8, 2026.
