Research Track
AI Model Behavior Observations
Documented observations of how large language models interpret prompts, resolve entity ambiguity, weigh source authority, and construct brand narratives. The mechanics behind why AI says what it says, and how that shapes perception.
The Question Driving The Research
Every brand strategy now has to answer a question that did not exist five years ago.
Search engines indexed pages and ranked them. Large language models interpret intent, resolve ambiguity, weigh authority across thousands of sources, and construct a narrative on the fly. The mechanics that shape your brand's appearance in AI recommendations are observable, measurable, and increasingly predictable. This research track documents what we see.
Three Mechanics, Continuously Observed
How LLMs construct what they say about you.
Three observable mechanics shape every AI recommendation. Each operates differently across models. Each can be measured. Each can be influenced once the pattern is understood.
Prompt Interpretation
How LLMs parse buyer intent, category framing, and stage-of-journey signals embedded in everyday phrasing.
- Buyer-stage signals embedded in phrasing shift the response shape. "Best X" returns ranked recommendations. "What is X" returns explanatory content. "Should I X" returns comparative reasoning.
- Location modifiers reshape the recommendation set. Adding a city or "near me" narrows results from global brand defaults to regional or local entities.
- Plural versus singular phrasing shifts category framing. "Best fire protection company" returns a single recommendation pattern. "Best fire protection companies" returns broader lists.
- Implicit commercial intent suppresses educational content. Models drop general information when buyer signals are strong.
Entity Resolution
How LLMs disambiguate brand names from common-word collisions, parent and product confusion, and ambiguous mentions.
- Common-word brand names(Evolve, Selina, Sonder) are routinely conflated with non-brand uses of those words. Resolution depends on surrounding context, not the brand identity itself.
- Parent and product confusion is widespread. "Marriott" can resolve to the hotel chain or to Homes & Villas by Marriott depending on prompt framing.
- Most-cited entity bias means ambiguous brand mentions often resolve toward the dominant entity in training data, even when context suggests a smaller brand was intended.
- Wikipedia presence shapes entity disambiguation more than any other single source.
Source Weighting
How LLMs rank the authority of references, prioritize recent third-party content, and decide who gets cited.
- Search-augmented models prioritize recent third-party content over older owned content, even when the owned content is more authoritative.
- Reddit citations dominate opinion-driven and lifestyle-category recommendations.
- Industry publications carry significantly more weight in B2B contexts than consumer review sites do.
- Owned brand websites are systematically under-weighted. Third-party validation drives the recommendation, not the brand's own claims.
- Citation depth varies sharply by model. ChatGPT favors Wikipedia and major publications. Perplexity surfaces niche industry sources. Claude leans on academic and reference content.
Field Observations
Patterns ARDI™ has documented across categories.
A selection of recurring patterns observed across the ARDI™ canonical 6-LLM panel. These are not theoretical positions. They are what we see, monthly, across thousands of category-defining prompts.
The same brand can hold #1 in one model and be invisible in another.
A brand can rank first in Perplexity for a category-defining query and not appear at all in Gemini for the same prompt run on the same day. Multi-model measurement is the only way to see this. Single-model audits miss it entirely.
Recommendation rank shifts month-over-month as the citation landscape changes.
Brand positions are not static. They move as third-party publishers add or remove content, as Reddit threads gain or lose visibility, and as models update their training and retrieval signals. Snapshot-in-time audits go stale within weeks.
Comparison prompts surface different secondary brands than ranking prompts.
"Best X" and "X vs Y" return overlapping but meaningfully different recommendation sets. The brands that show up in head-to-head comparison queries often differ from the brands that show up in ranked-list queries, even within the same category. Buyer-stage prompt design matters as much as the category itself.
Stage-of-journey phrasing reshapes the entire recommendation set.
Pre-purchase educational queries surface authority sources. Decision-stage queries surface vendors. Comparison queries surface niche specialists. The same category, asked three different ways, produces three different brand sets.
Hallucinations cluster where authoritative content is sparse.
Models invent brand attributes when their training and retrieval signals are thin. The brands most vulnerable to AI hallucination are not the unknown ones. They are the mid-tier brands with partial coverage. Visibility without authority creates risk, not protection.
Why This Research Exists
What this means for brand and marketing teams.
Understanding the mechanics is the difference between guessing why AI says what it says, and engineering for a different outcome. Three implications shape how we structure every client engagement.
You cannot fix what you cannot see.
Single-prompt audits and single-model checks miss the structural patterns. Measurement has to be continuous, multi-model, and stage-aware to reflect how AI actually recommends.
The brand does not control the narrative.
Third-party validation, citation density, and source authority shape AI recommendations more than owned brand content does. Owned-content strategies built for search engines do not transfer to LLMs without adjustment.
The mechanics are observable, not mysterious.
AI recommendation behavior is patterned, measurable, and over time, predictable. The brands that win in AI recommendation are the brands that treat AI as a discoverable system, not a black box.
See It For Your Brand
How does AI decide what to say about your brand?
The research above describes patterns we observe across categories. The next step is seeing what those patterns look like for your brand specifically.