By 2026, Eleanor Vance, Aura Retail’s Head of Digital Marketing, was facing a familiar problem that had become a full-blown crisis. The online fashion brand had poured money into large language model (LLM) shopping experiences and even saw a promising 15% lift in average order value from early tests, but their analytics were a mess. Repeat purchases weren’t being attributed correctly, and tracking a single customer journey across a phone and a laptop was nearly impossible. Eleanor knew their fancy LLM was basically running blind without solid identity resolution, meaning its ability to personalize and justify ad spend was just a fraction of what it could be. With third-party cookies completely dead across all major browsers, they had to build a new way to see their customers’ digital footprints, particularly for LLM purchases.
Key Takeaways
- Build a first-party data strategy by getting explicit consent for identifiers like emails and phone numbers through things like loyalty programs.
- Use probabilistic matching, which analyzes signals like device fingerprints and IP addresses, to connect user sessions when you don’t have a direct login.
- Feed LLM outputs into a unified customer profile so that your personalized offers are consistent on every channel a customer uses.
- Put privacy tech first. Look into differential privacy and federated learning to handle customer data in a way that builds trust and keeps you compliant.
- Constantly audit your identity resolution models against real purchase data to make them more accurate and stop wrongly connecting different users.
The Disappearing Cookie: A Catalyst for Change
We all know the story: for years, third-party cookies were how digital marketers tracked people, pieced together customer journeys, and did personalization. When they were finally phased out for good by all the big browsers in early 2024, companies like Aura Retail were left with a huge hole in their data. “It was like trying to navigate a city without street signs,” Eleanor said in a strategy meeting. “Our LLMs were great at writing product descriptions and suggesting things, but that intelligence was fragmented because they couldn’t tell if the person on the phone was the same one who was on their desktop earlier.”
This was especially painful for Aura’s personalized product discovery, which was powered by their LLM. Imagine the common scenario: a customer browses on their work laptop at lunch, puts some stuff in a cart on their tablet that night, and then buys on their phone the next day. Without a single ID connecting those actions, the LLM saw three different people. The result? Wasted opportunities and annoyed customers getting recommendations for items they already had in a cart on another device. The problem was the LLM’s data foundation, not its intelligence.
So Eleanor’s team went back to the drawing board, looking for a way to get a full picture of the customer that wouldn’t run afoul of new privacy rules. They knew any new identity resolution system had to be scalable and totally compliant with regulations like GDPR and the recently-strengthened CCPA. “We couldn’t just replace one tracking mechanism with another equally opaque one,” Eleanor insisted. “Building trust with our customers was the whole point.”
Building a First-Party Data Fortress
Aura’s first move was to get serious about their first-party data collection, which meant collecting the right data with explicit consent, not just vacuuming up everything they could. They reworked their newsletter sign-up pop-ups to make the value-add obvious for handing over an email. They also launched a new loyalty program that gave people a real reason (exclusive discounts and early sale access) to share their phone number or birthday. “We shifted our mindset from ‘collect everything’ to ‘collect what’s essential and provide clear value for it’,” Eleanor explained. It took more UX work upfront, but the payoff was higher-quality data that people had actually agreed to share.
The new loyalty program, “Aura VIP,” became the anchor for this new strategy. Members got LLM-generated recommendations sent right to their inbox or via SMS. Because this was tied to a unique customer ID in Aura’s database, they could finally track what people were doing across all their devices with real accuracy. If a logged-in VIP member used the LLM on their phone, that conversation immediately updated their profile, giving the model a better read on their style and what they were likely to buy. This direct connection was a breakthrough for increasing LLM purchases.
On the back end, this meant connecting their customer data platform (they went with Segment, which was pretty standard for this kind of work in 2026) directly to their LLM inference engine. This setup fed real-time updates to customer profiles from any interaction, browsing, adding to a cart, or just chatting with the LLM. So when a logged-in VIP asked the LLM for “sustainable evening wear” on their desktop, and then later looked at a specific dress on their phone, the phone’s LLM knew about the earlier search. It created the connected cross-device conversation that Eleanor had been aiming for.
Probabilistic Matching: Inferring the Unseen Connections
Of course, not every user logs in or joins a loyalty program on every visit which left a lot of anonymous traffic. To connect those dots, Aura turned to probabilistic identity resolution. They worked with a data clean room provider, InfoSum, which let them match anonymized identifiers across touchpoints without ever sharing raw customer data. The process used encrypted, hashed IDs that could be compared to find likely matches between sessions.
This probabilistic approach is all about finding patterns. It analyzes signals like device type and IP address, along with behavioral clues such as browsing speed and how a person scrolls. “It’s like forensic science for digital marketing,” Eleanor mused. “You’re looking for enough circumstantial evidence to confidently say two seemingly disparate data points belong to the same person.” The LLM itself became a key source of this evidence. By analyzing the words and product preferences from anonymous chats, the LLM could flag similarities. For example, if two sessions from different devices but a similar IP range both kept asking for “boho chic” styles, the matching algorithm would raise its confidence score that it was the same person.
To avoid getting it wrong, Aura’s data science team built their own algorithm to score every potential match. They set a high confidence threshold, starting at 85% and tweaking it from there, and only matches that cleared this bar were allowed to update a unified customer profile. This was all about preventing false positives, because as Eleanor warned her team, “The worst thing you can do is incorrectly assume you know a customer. That’s how you get creepy, not personalized.”
““We put a tremendous amount of effort into making sure your model can remember what’s important to you over days, weeks, months, and beyond,” said research scientist Kavya Kopparapu in a video accompanying the launch.”
LLMs as Identity Enablers, Not Just Content Generators
Aura’s LLMs did more than just personalize content. They actively helped improve the identity resolution process. For instance, when a new, anonymous user started a chat with the LLM and described their style, the model’s NLU could pick out attributes like “bohemian” or “minimalist.” These tags were attached to a temporary profile. If that same person eventually logged in or bought something, the inferred style attributes would merge with their new, permanent profile, giving Aura a much deeper understanding of that customer from their very first interaction.
The LLMs also acted as a watchdog for the identity graph. If a profile that looked like a loyal customer suddenly started browsing completely different styles or asking the LLM about things totally out of character, the system flagged it as an anomaly. Was it a shared device? A hacked account? Or just a genuine change in taste? Whatever the cause, the flag would make the LLM pull back on its recommendations and be more cautious. “The LLM acts as a sort of digital detective, constantly cross-referencing and refining its understanding of who you’re and what you want,” Eleanor explained. This self-correcting loop between the LLM and the identity data gave them a real edge.
With this setup, Aura could finally get past basic demographic segments. Their LLMs could now track a customer’s changing tastes and budget, all inferred from conversations and behavior on any device. It meant that a customer who spent Tuesday afternoon researching “comfortable work-from-home attire” on their tablet would get smarter suggestions when they started looking for “weekend getaway outfits” on their phone Friday morning. The LLM knew it was the same person with different needs at different times.
Unsurprisingly, LLM conversion rates climbed as this understanding deepened. As a side benefit, Aura’s unified customer profiles, full of rich LLM insights, became a goldmine of feedback for their prompt engineer team, who used it to sharpen the LLM’s conversational skills and recommendation logic.
The Privacy Imperative: Building Trust in a Data-Driven World
None of this would have worked if they’d been creepy about it, so Aura’s whole strategy was built on a “privacy-by-design” approach. Privacy wasn’t an afterthought. It was part of the spec for the identity resolution system from day one. All collected data was pseudonymized and encrypted. They also gave users very clear, granular consent options to control what data was collected and why. For Eleanor, this level of transparency was simply a requirement for doing business in 2026.
They also put money into specific privacy-enhancing tech. They started looking into differential privacy, a method that adds statistical “noise” to data before it’s analyzed. This allowed them to spot broad behavioral trends for LLM training without being able to see what any single individual did. They also used federated learning, which let their models learn from data spread out across different servers or even devices without having to pull all that raw data into one vulnerable, central place. These were complex systems to set up, but they formed the ethical backbone of their entire identity strategy.
This intense focus on ethical data handling made it much easier for Aura Retail to stay on the right side of LLM compliance and strict global regulations. The security they built for identity resolution also fit perfectly into their larger company plan for private AI for LLMs, which aligns with the trend of 85% of enterprises prioritizing security by 2028.
Looking Ahead: The Unified Customer Experience
When Eleanor’s team presented their results to the board in January 2026, the numbers spoke for themselves. Their internal report showed that better identity resolution had directly contributed to a 22% increase in customer lifetime value and cut ad spend waste by 10% by the end of 2025. Their LLMs, finally fed with a complete picture of each customer, were creating personalized experiences that actually felt helpful, not just intrusive. What Eleanor’s work demonstrated was that identity resolution isn’t just some back-end IT problem anymore. For any company using AI to talk to customers, it’s a core part of the business strategy.
For brands trying to make LLM purchases work, the lesson is clear. The companies that figure out how to connect all these dots, while being obsessive about privacy, are the ones who will build relationships that last and get the real value out of their AI.
What is identity resolution in the context of LLM purchases?
It’s the process of figuring out that the user on a laptop, a phone, and a tablet is the same person, even if they aren’t logged in. For an LLM, this creates a single, consistent memory of that customer’s preferences and history, which it uses to make smarter, more relevant recommendations during the buying process.
Why are traditional methods like third-party cookies no longer sufficient for identity resolution?
All major web browsers have blocked them over privacy concerns and new regulations. They just don’t work anymore as a reliable way to track users across different sites, so marketers need new methods that respect user privacy to connect the dots on customer journeys.
How can first-party data improve LLM-driven personalization?
This is data you collect directly from customers who give you permission, like an email from a newsletter sign-up. It provides a reliable, persistent ID. When you feed this data to an LLM, the model can build an accurate profile, making sure its personalized recommendations are consistent across all of a customer’s devices.
What role do probabilistic matching techniques play in modern identity resolution?
It uses anonymous signals like an IP address, device information, and browsing behavior to calculate the probability that two different sessions were from the same person. It’s a way to fill in the blanks when you don’t have a login or email, giving you a fuller map of the customer journey for the LLM to use.
How do privacy-enhancing technologies like differential privacy and federated learning relate to identity resolution?
They are essential tools for handling customer data ethically and legally. Differential privacy adds statistical “noise” to data so you can analyze trends without identifying individuals. Federated learning lets an LLM learn from data on different servers without centralizing it. Both techniques protect user privacy while still allowing you to get the insights needed for good personalization.
