Guide: How I’d Build a Delta Air Lines Campaign for Possible in Miami

April 4, 2025 · 3 min read · Originally on LinkedIn

AI, Attention & EARO Build Logs New here? Start with the most read →

The gist

Privacy-first personalization doesn't have to mean dumber targeting. Using himself as the test persona — a NYC executive flying to the POSSIBLE conference — Popov pairs anonymized, hashed PII with a five-vector psychographic profile, then lets an LLM turn that schema into scored creative variants and activation rules. Reach someone by how they think and what they value, not just where they are.

In English, please

A walkthrough of how an airline could build ads for one very specific kind of high-value flyer -- a New York-based executive in his 40s or 50s, flying business class to a major marketing conference in Miami -- without ever touching that traveler's real name or contact details. The test-case persona is status-conscious, schedule-driven, and loyalty-card-carrying: the kind of frequent flyer airlines most want to reach.

The first layer of data is ordinary information -- age range, general neighborhood, device, loyalty-card number -- but scrambled into codes instead of stored as real identifiers. It works like a claim-check ticket instead of a passport: the airline's systems can recognize a valuable repeat customer and their habits without any record of who that person actually is by name.

A second, less common layer scores how the traveler thinks and behaves, across five categories: how quickly and clearly they process information; whether they reason things out or lean on gut-level shortcuts; how loyal they stay to a brand once they trust it; how open-minded, organized, outgoing, agreeable, and emotionally steady they are; and how predictable their routines are -- a planner versus someone who decides at the last minute.

Those scores sort the traveler into a labeled personality type -- here, "Rational Explorer," with a calculated engagement score of 82 -- and that label gets handed to AI writing tools (several are named, including Claude and GPT-4) to produce several different versions of the same ad. A logic-minded traveler gets a data-heavy, time-saving pitch; someone who values comfort gets a calmer, personal-feeling video ad; someone motivated by status gets a rewards-focused social post -- each version, down to its imagery and how often it's shown, matched to a specific trait.

The pitch is that this threads a needle: no real name or contact info is ever needed, yet the messaging can still be tuned to how a specific kind of person thinks, not just their age bracket and location. The same approach is proposed for other traveler segments -- frequent-flying executives, conference-goers, high-loyalty customers -- and eventually for more advanced signals: biometric cues that hint at intent, real-time behavior scoring, and Apple Wallet-based identity syncing -- all described throughout as privacy-preserving.

On this page

If I were leading personalization strategy at Delta Business Air Lines, targeting high-value business travelers flying from NYC to the POSSIBLE conference in Miami, I’d start with a very specific persona.

Mine :)

I’m a New York-based executive, aged 40–50, attending POSSIBLE a premier global marketing event. I care about efficiency, comfort, predictability, and status. And if Delta were running a campaign targeting someone like me, here’s exactly how I’d design it using Anonymized PII and a Psychographic framework, with LLMs orchestrating creative outputs at scale.

Campaign Setup: Me as the Persona

Advertiser: Delta Business (Delta Air Lines)
Objective: Drive bookings for premium/business-class flights from NYC to MIA (aligned to POSSIBLE conference timing)
Persona:
Male, 40–50
NYC-based
Executive, business traveler
High openness and conscientiousness emotionally stable
Conference-goer: status-driven, schedule-conscious

The Data Stack: How I’d Build the Profile

To make this work in a privacy-first environment, I’d use a two-layered approach: anonymized PII + psychographics.

Layer 1: Anonymized PII Metadata

Here’s how I’d map the foundational identity layer inside a CDP - no raw identifiers, just hashed or tokenized fields:

Demographics: Age range, gender, job category, employment status
Geolocation: Geo-bucketed location (NYC), country code
Device & Network: Pseudonymous device ID, fingerprint hash
Identity: Hashed email, user token
Financial: Loyalty card hash, payment type
Behavioral: Preference cluster, purchase intent

Layer 2: Psychographic Profile (My Cognitive Fingerprint)

This is where it gets interesting—and personal. These five vectors describe how I think, decide, relate, and engage.

Cognitive: I process info fast and prefer clarity (high reasoning, verbal comprehension)
Rationality: I weigh outcomes logically, but still respond to persuasive framing
Attachment: I’m brand-loyal if the UX is seamless and consistent
Personality: I lean toward openness and structure (but ignore hype)
Temperament: I work in rhythms, act on reminders, and avoid last-minute chaos

What I’d Feed Into the CDP

No raw identifiers — just hashed or tokenized fields, flowing from anonymized PII into the psychographic layer, a segmentation engine, and finally activation:

[PII_Metadata_Anonymized]
 ├── Demographics
 ├── Geolocation
 ├── Device & Network
 ├── Identity & Contact
 ├── Financial & Transactional
 └── Behavioral
        ↓
[Psychographic Input Layer]
 ├── Cognitive
 ├── Rationality
 ├── Attachment
 ├── Personality
 └── Temperament
        ↓
[Segmentation Engine / ML]
 ├── Archetype: Rational Explorer
 └── Engagement Score: 82
        ↓
[Activation Layer]
 ├── Attention Modeling
 ├── Creative Selection
 └── Media Planning

And the psychographic layer in detail — the five vectors, scored:

[Psychographic Input Layer]
 ├── Cognitive (G)
 │    ├── Fluid Reasoning: High
 │    ├── Working Memory: Moderate
 │    ├── Verbal Comprehension: High
 │    └── Processing Speed: Low
 ├── Rationality
 │    ├── Instrumental Rationality: High
 │    ├── Epistemic Rationality: Moderate
 │    ├── Reflective Thinking: High
 │    └── Heuristic Reliance: Low
 ├── Attachment
 │    └── Style: Secure
 ├── Personality (Big Five)
 │    ├── Openness: High
 │    ├── Conscientiousness: Moderate
 │    ├── Extraversion: Low
 │    ├── Agreeableness: High
 │    └── Neuroticism: Low
 └── Temperament
        ├── Activity Level: High
        ├── Emotional Reactivity: Moderate
        ├── Sociability: Low
        ├── Self-Regulation: High
        └── Rhythmicity: Predictable

Powering the Messaging: LLMs at Work

With this schema in place, I’d pipe the structured psychographic profile into an LLM (Gemini, DeepSeek, Claude, Grok, GPT-4) to generate tailored creative messaging variants. Those creatives would feed into DCO platforms, CRM, or native ad buys on LinkedIn, The New York Times, or newsletter partners.

Psychographic-Aligned Creatives for the Delta Campaign

Looking at format, message style, and creative, the LLM would normalize its recommendations into something like this (easily output as JSON):

Creative variant	Format	Messaging style	Justification
“Make Time Work for You”	Display / Native	Data-backed, time-saving	High instrumental rationality and conscientiousness
“Your Schedule, Our Priority — Delta One to Miami”	Carousel / Email	Structured itinerary builder	Matches fluid reasoning and self-regulation for planners
“Business Class That Feels Personal”	Video pre-roll	Emotionally subtle; comfort + autonomy	Tailored to secure attachment and low neuroticism
“Elevate Possible: Arrive Ready”	Native article	Executive experience story tied to POSSIBLE	Combines openness, reflective thinking, and relevance to context
“Unlock Your Travel Ritual”	Push / In-app	Dayparted reminder to book	Leverages rhythmicity and predictable self-regulation
“Delta × POSSIBLE: Get Rewarded”	Sponsored LinkedIn post	Social, reward-oriented	Taps moderate extraversion and status-signaling

Insights Behind Activation

Psychographic signal	Activation approach
High Openness + Rationality	Educate and inspire with logical creative, premium positioning
Secure Attachment + Low Neuroticism	Confident, low-friction booking journey
High Conscientiousness + Rhythmicity	Align with work cadence and structured planning
Moderate Sociability + Conference Attendee	Light peer-oriented social proof, but not influencer-heavy

Why It Works

This approach blends anonymity with intelligence:

PII is protected; signals are still actionable
Psychographics enhance precision where behavioral data alone can’t
LLMs translate data into empathetic personalization, not just targeting

Delta would be able to reach someone like me not just because of where I am—but because of how I think and what I care about.

Result: One Creative, Built for One Mind

MAKE TIME WORK FOR YOU

Your Schedule, Our Priority — Delta One to Miami

[ Book Now ]

If I Were Running This at Delta…

I’d scale this approach to micro-segments:

CXOs traveling monthly
AI/marketing conference goers
High-loyalty/low-discount travelers

And I’d explore predictive upgrades using biometric intent, session-level psychographic classifiers, and even Apple Wallet-based identity sync (privacy-preserving, of course).

This is what relevance looks like in 2025.

What would you build if this was your profile?

Let’s workshop it.

Appendix: Full Taxonomy

The complete field-level schema and a fully populated example — tucked away so they don’t interrupt the read. Expand whichever you need.

Full taxonomy — every field

[PII_Metadata_Anonymized]
   ├── Demographics
   │     ├── Age Range
   │     ├── Birth Year
   │     ├── Gender
   │     ├── Ethnicity Category
   │     ├── Marital Status
   │     ├── Household Size
   │     ├── Dependent Count
   │     ├── Education Level
   │     ├── Income Range
   │     ├── Employment Status
   │     └── Job Category
   ├── Geolocation and Address
   │     ├── Hashed Postal Code
   │     ├── Geo-Bucketed Location
   │     ├── City Code
   │     ├── Region Code
   │     ├── Country Code
   │     └── Hashed IP Address
   ├── Device and Network Identifiers
   │     ├── Pseudonymous Device ID
   │     ├── Tokenized MAC Address
   │     ├── Hashed IMEI
   │     ├── Browser Fingerprint ID
   │     ├── Anonymized User Agent Signature
   │     └── Network ID Token
   ├── Identity and Contact Information
   │     ├── Hashed Email Address
   │     ├── Phone Number Hash
   │     ├── User ID Token
   │     └── Login ID Hash
   ├── Financial and Transactional
   │     ├── Transaction Token ID
   │     ├── Loyalty Card Hash
   │     ├── Billing Region Code
   │     ├── Purchase Event ID
   │     └── Payment Method Type
   └── Online Behavior and Preferences
         ├── Anonymized Clickstream ID
         ├── Session Event Signature
         ├── Preference Cluster ID
         ├── Ad Interaction Token
         ├── Hashed App ID
         └── Inferred Purchase Intent Score

[Psychographic Input Layer]
   ├── Cognitive (G)
   │     ├── Fluid Reasoning
   │     ├── Working Memory
   │     ├── Verbal Comprehension
   │     └── Processing Speed
   ├── Rationality
   │     ├── Instrumental Rationality
   │     ├── Epistemic Rationality
   │     ├── Reflective Thinking
   │     └── Heuristic Reliance
   ├── Attachment
   │     └── Style
   ├── Personality (Big Five)
   │     ├── Openness
   │     ├── Conscientiousness
   │     ├── Extraversion
   │     ├── Agreeableness
   │     └── Neuroticism
   └── Temperament
         ├── Activity Level
         ├── Emotional Reactivity
         ├── Sociability
         ├── Self-Regulation
         └── Rhythmicity
             ↓
[Segmentation Engine / ML]
   ├── Archetype Cluster
   ├── Predictive Engagement Score
   ├── Inferred Travel Profile
   └── Campaign Fit
             ↓
[Activation Layer]
   ├── Attention Modeling
   │     ├── Content Length
   │     ├── Visual Complexity
   │     └── Delivery Timing
   ├── Creative Selection
   │     ├── Message Style
   │     ├── Tone
   │     ├── Imagery
   │     └── Call to Action
   └── Media Planning
         ├── Format
         ├── Channels
         ├── Dayparting
         └── Frequency Cap

Mock-up example — the persona, scored end to end

[Psychographic Input Layer]
   ├── Cognitive (G)
   │     ├── Fluid Reasoning: High
   │     ├── Working Memory: Moderate
   │     ├── Verbal Comprehension: High
   │     └── Processing Speed: Low
   ├── Rationality
   │     ├── Instrumental Rationality: High
   │     ├── Epistemic Rationality: Moderate
   │     ├── Reflective Thinking: High
   │     └── Heuristic Reliance: Low
   ├── Attachment
   │     └── Style: Secure
   ├── Personality (Big Five)
   │     ├── Openness: High
   │     ├── Conscientiousness: Moderate
   │     ├── Extraversion: Low
   │     ├── Agreeableness: High
   │     └── Neuroticism: Low
   └── Temperament
         ├── Activity Level: High
         ├── Emotional Reactivity: Moderate
         ├── Sociability: Low
         ├── Self-Regulation: High
         └── Rhythmicity: Predictable
             ↓
[Segmentation Engine / ML]
   ├── Archetype Cluster: Rational Explorer
   ├── Predictive Engagement Score: 82
   ├── Inferred Travel Profile: Premium Business Traveler
   └── Campaign Fit (Delta): High Relevance
             ↓
[Activation Layer]
   ├── Attention Modeling
   │     ├── Content Length: Short
   │     ├── Visual Complexity: High
   │     └── Delivery Timing: Morning & Evening (Weekdays)
   ├── Creative Selection
   │     ├── Message Style: Data-backed with Emotional Arc
   │     ├── Tone: Confident and Reassuring
   │     ├── Imagery: Executive Lifestyle
   │     └── CTA: Goal-Oriented, Low Urgency
   └── Media Planning
         ├── Format: Native, CTV, Premium Display
         ├── Channels: LinkedIn, Email, Programmatic Display
         ├── Dayparting: Pre/Post Work Hours
         └── Frequency Cap: 3x Max Exposure

Citable summary

Privacy-first ad personalization works by layering anonymized PII — hashed, tokenized fields in a customer data platform — with a five-vector psychographic profile (cognitive, rationality, attachment, personality, temperament), then piping the structured profile into large language models that generate creative matched to how a specific person thinks. A worked Delta Air Lines campaign for business travelers flying New York to the POSSIBLE conference in Miami demonstrates the full pipeline.

The privacy-first personalization framework layers anonymized PII — hashed emails, pseudonymous device IDs, geo-bucketed location — beneath five psychographic vectors: cognitive, rationality, attachment, Big Five personality, and temperament.
Structured psychographic profiles are piped into LLMs — Gemini, DeepSeek, Claude, Grok, GPT-4 — to generate creative variants for DCO platforms, CRM, and native buys on LinkedIn.
A worked Delta Air Lines premium-class campaign targets NYC executives flying to the POSSIBLE marketing conference in Miami and yields six psychographic-matched creative variants.
In the Delta example, a segmentation engine scores the mock persona as archetype "Rational Explorer" (predictive engagement score 82), activating native, CTV, and premium display at a 3x frequency cap.

Popov, E. “Guide: How I’d Build a Delta Air Lines Campaign for Possible in Miami.” No Fluff Advisory, April 2025. https://nofluffadvisory.com/writing/guide-how-i-d-build-a-delta-air-lines-campaign-for-possible-in-miami/

Where this sits by meaning · corpus 1787917944522 · rev 44330b2

sits near Behavioral Prediction Framework for Advertising close — cosine 0.75, in the middle half of neighbour pairs in this corpus
sits near Fire the Dashboard, Hire the Agent nearby — cosine 0.71, in the lower quarter of neighbour pairs — still one of this essay’s three nearest
sits near Marketing is Dead, Long live the Marketing nearby — cosine 0.70, in the lower quarter of neighbour pairs — still one of this essay’s three nearest

Proximity, not evidence — cosine measures nearness of meaning, never that one essay supports another.

Does 0.80 mean 80% similar?

No — and that is why these cards show a phrase rather than a number. Cosine similarity measures the angle between two essays once each has been turned into a vector of meaning. It is not a share of overlapping words, and it does not run 0 to 100.

cos θ = A · B ‖A‖ ‖B‖

A score of 1.00 means the two point the same direction, 0 means unrelated, −1 means opposite. So 0.80 is an angle of about 37°, not 80% of anything.

Magnitude alone also misleads here. One author writing on adjacent subjects produces a narrow band: every pair this site draws scores between 0.55 and 0.90. Read as a percentage, 0.80 sounds like a bare pass — in this corpus it sits above roughly two thirds of all neighbour pairs. The labels below therefore report rank against every other neighbour pair, which is the reading that survives.

nearby below 0.735 — in the lower quarter of neighbour pairs — still one of this essay’s three nearest
close 0.735 and above — in the middle half of neighbour pairs in this corpus
very close 0.806 and above — closer than 75% of neighbour pairs in this corpus
unusually close 0.845 and above — closer than 90% of neighbour pairs in this corpus

Thresholds are percentiles of the 252 neighbour pairs this site renders, not universal constants; another corpus would sit elsewhere. Nearness is nearness of meaning — never a claim that one essay supports another.