Antavo’s 2026 AI Loyalty: 25% CLTV Boost

Using large language models (LLMs) for customer retention isn’t a theoretical exercise anymore. The results are measurable. Brands are spinning up AI-powered loyalty programs for hyper-personalized experiences, trying to get past generic points systems to actually anticipate what customers need. But how well do these systems really work, especially for a startup trying to gain a foothold in a crowded market? This campaign teardown digs into a recent initiative that used an LLM to get serious engagement and show a tangible ROI on AI marketing.

Campaign Overview: “Next-Gen Rewards” by Antavo

In Q3 2025, loyalty tech provider Antavo launched its “Next-Gen Rewards” campaign to prove its LLM-enhanced platform wasn’t just smoke and mirrors. The goal was to land new B2B clients by showing real, concrete improvements in retention and engagement for a fictional D2C beauty brand they created called “GlowUp Cosmetics.” This was a live-fire exercise. Antavo used a controlled group from GlowUp’s real customer base and ran a targeted acquisition strategy for new members. The whole thing ran for 12 weeks, from September 1st to November 24th, 2025.

Strategy: AI-Powered Personalization at Scale

The strategy was all about getting past basic demographic segmentation to deliver true 1-to-1 personalization. Antavo’s platform, hooked up to a proprietary LLM, chewed on customer behavior from all over: purchase history, website browsing data, email clicks, and even sentiment from customer service chats. This deep read on customer activity allowed it to dynamically create personalized offers, content, and conversational prompts right inside the loyalty program’s interface.

• Target Audience: They used existing GlowUp Cosmetics customers to demonstrate retention and then targeted lookalike audiences on Meta platforms for new acquisition. Those lookalikes were modeled off GlowUp’s top 15% of spenders and most active loyalty members.
• Key Differentiator: The LLM could predict future buying behavior and recommend entire routines or solutions for specific skin concerns, not just single products. This was designed to build a much stronger connection to the brand.
• Retention Focus: Existing members got AI-curated “mystery boxes” with samples based on their predicted tastes, plus early access to new products.
• Acquisition Focus: New customers were drawn in with a tiered welcome offer. The discount and free gift were dynamically changed by the LLM based on what a user clicked on right after landing on the GlowUp site.

Creative Approach: Human Touch, AI-Driven Content

All the creative assets kept GlowUp Cosmetics’ consistent brand voice, focused on natural beauty and self-care. But the actual copy in the ads, emails, and on-site prompts was heavily influenced by the LLM, in some cases, the AI wrote it directly. For example, email subject lines were A/B tested, pitting AI-generated options against human-written ones. The AI also wrote personalized product descriptions that zeroed in on the benefits most relevant to a specific user’s known or inferred problems.

• Ad Copy: They ran short, benefit-focused headlines like “Unlock Your Best Skin: AI-Powered Routine Inside!” The call-to-action was almost always “Join GlowUp Rewards & Get Your Personalized Welcome.”
• Email Content: The email bodies did more than just use a different subject line. They dynamically pulled in product images and descriptions specifically for that recipient. One big win was a cart abandonment sequence where the LLM generated unique reasons to complete the purchase, often referencing products the person had looked at before.
• On-Site Experience: The loyalty member dashboard had dynamic content blocks. If you were a customer interested in anti-aging, you might see a blog post about collagen next to product recommendations for serums, all selected by the LLM.

Targeting and Channels: Precision at Scale

This was a multi-channel campaign, but it leaned heavily on Meta Ads (Facebook and Instagram) for getting new customers and on email marketing for keeping the ones they had.

• Meta Ads:
  • Audiences: They used custom audiences of existing customers (but excluded current loyalty members), 1% and 3% lookalike audiences built from high-value purchasers, and standard interest-based targeting (skincare, beauty, wellness).
  • Placements: Instagram Feed, Facebook Feed, and Instagram Stories were the main spots.
  • Ad Formats: They ran carousel ads showing off personalized product bundles, single-image ads with strong testimonials, and quick video ads that showed how easy the loyalty program was to use.
• Email Marketing:
  • Segments: They segmented by existing loyalty members, recent buyers who hadn’t joined the program yet, and people who gave their email but never bought anything.
  • Automation: Workflows included a welcome series for new sign-ups, post-purchase follow-ups with personalized tips on how to use the products, and win-back campaigns for members who had gone cold.
• On-Site Pop-ups: These were triggered by exit intent or when someone viewed a specific page, offering instant enrollment in the loyalty program with a personalized reward right off the bat.

Campaign Performance Metrics and Analysis

The “Next-Gen Rewards” campaign put up strong numbers, especially when you isolate the uplift that came directly from the LLM integration. Here’s how it broke down:

Acquisition Performance (Meta Ads)

The ad creative and targeting optimized by the AI, which used LLM insights to tweak messaging for specific audience segments, clearly beat the standard control group. The 3.9x ROAS from the AI-optimized ads was a big deal, showing that the personalized messages hit home and drove immediate sales from new members. This lines up with a recent eMarketer report that found companies using generative AI for ad copy are seeing around a 15% lift in CTRs which is right in the ballpark of what we saw here.

Retention Performance (Existing Loyalty Members)

• Average Order Value (AOV) Increase: A 15% jump for members who got LLM-personalized recommendations compared to those getting standard ones.
• Repeat Purchase Rate: 28% higher for the LLM-engaged segment versus the control group over the 12-week campaign.
• Customer Lifetime Value (CLTV) Projection: The first three months of data projects a 25% higher CLTV for customers brought in and engaged through the LLM program.
• Support Ticket Reduction: Common customer questions dropped by 18%, thanks to the LLM-powered FAQ and self-service options in the loyalty portal.

The AOV and repeat purchase rate increases show how the LLM drives actual incremental revenue from existing customers. When people feel like you get them and send them relevant suggestions, they spend more and come back faster. This is about predicting their next need with a pretty high degree of accuracy. The drop in support tickets also shows a real operational win, freeing up human agents to handle more difficult problems.

What Worked: The Power of Contextual Personalization

The biggest win was the LLM’s ability to deliver contextual personalization. It went way beyond simple rules-based engines to understand the “why” behind what customers were doing. For instance, if a customer looked at several anti-aging serums but didn’t buy one, the LLM might guess they’re price-sensitive and offer a small discount on a similar, more affordable product, or maybe suggest a sample kit. That kind of nuanced response is where these models really deliver.

Another victory was dynamic content generation. The sheer number of personalized email variations and ad copy versions we could generate and test without a human having to write every single one was huge. This let us iterate and optimize fast, quickly figuring out what worked for different micro-segments.

Finally, the predictive analytics were key. By anticipating churn risk or a customer’s next best action, the LLM allowed for proactive engagement. For example, customers who were showing signs of drifting away (like lower email open rates or fewer site visits) got a targeted re-engagement offer that was much more effective than a generic “we miss you” discount.

What Didn’t Work: Over-Reliance on Pure Automation

While automation was a plus, there were times when the pure LLM-generated content felt generic or just slightly off-brand. This happened mostly in the early tests where we didn’t have enough human oversight. Some of the first AI-generated blog post snippets, for instance, just didn’t have the empathetic tone GlowUp Cosmetics was known for. It was obvious that a “human-in-the-loop” process was non-negotiable, especially for top-of-funnel content. This meant having human editors review and polish the AI’s suggestions.

The other headache was data cleanliness and integration. An LLM is only as good as the data it’s fed. In the beginning, we had inconsistencies in customer data coming from different systems (the CRM, the e-commerce platform, support logs), which caused some bad personalization attempts. This just proved how important it is to have your data governance and integration strategy locked down *before* you roll out an advanced AI tool.

Optimization Steps Taken

Based on what we saw early on, we made a few important changes:

1. Enhanced Human Oversight: We assigned a dedicated content reviewer to approve all LLM-generated email campaigns and ad copy before they went out the door. This kept the brand voice tight and caught any weird phrasing.
2. Refined LLM Training Data: We fed the LLM more proprietary brand guidelines, tone-of-voice docs, and examples of past marketing copy that had performed well. This helped the AI learn and copy GlowUp’s specific personality.
3. Improved Data Hygiene Protocols: We started doing weekly data audits to find and fix mismatches in customer profiles. This meant cross-checking purchase history against loyalty activity and website behavior logs to make sure everything lined up.
4. Dynamic Offer Testing Framework: Instead of doing static A/B tests, we switched to a multi-armed bandit approach for personalized offers. This let the LLM continuously learn and send traffic to the best-performing offer variations in real-time, which sped up optimization.
5. Feedback Loop Integration: Any time a customer service interaction showed a personalization fail (e.g., a customer complained about a recommendation), we fed that data back into the LLM’s learning model to refine future suggestions for similar customers.

These adjustments mattered. The initial deployment was promising, but the process of constant refinement, guided by both the numbers and real human feedback, is what really got the most out of the LLM. It’s a good reminder that even the smartest AI needs to be managed and improved over time.

Conclusion

The “Next-Gen Rewards” campaign proved that putting LLMs in loyalty programs is a powerful strategy for getting real, measurable wins in both customer retention and acquisition. Brands that are willing to invest in good data infrastructure and keep a human in the loop will get a serious competitive advantage, turning generic interactions into personal, revenue-driving relationships. A good place to start is to find one key customer journey where your personalization is weak and run a pilot with an LLM-driven solution to prove the concept’s value.