[Case Study] Using DataFeedWatch AI to create a scalable title optimization architecture

Some may use AI to automatically rewrite titles once and hope for the best. But we designed a repeatable, two-phase framework using DataFeedWatch AI.

When our team at Adalystic Marketing stepped in to optimize two distinct B2B healthcare ecommerce catalogs we ran into a classic enterprise bottleneck. Scaling manual optimization is practically impossible, yet generic automation risks ruining feed quality.

We’re sharing the framework we created so that you can try the strategy out for yourself, across any vertical.

Our strategy: creating an optimization framework in two phases 

Before we get into the specifics, I’ll walk you through the overarching strategy we used for both of our clients.

A lot of marketers treat feed optimization as a one-and-done project. But we turned it into a scientific loop using a structured, two-phase approach.

Here’s how it goes:

1. Phase 1: Mass A/B title testing (learning)

We use DataFeedWatch’s AI title generation to run 50/50 split tests across thousands of SKUs. Our goal here isn’t to find a permanent fix, but to gather fast data on what buyers actually respond to.

2. Phase 2: Create optimized titles based on test results (scaling)

We take the exact insights proven by the AI in Phase 1 and hardcode them into repeatable templates inside DataFeedWatch.

Here is how this framework played out across two of our real-world B2B healthcare catalogs. 

Case study 1: Title optimization at scale (+20,000 SKUs)

The challenge

Our first client had a large catalog of over 20,000 SKUs. Manual title rewrites were completely out of the question.

We needed a time-efficient way to:

• Test title variations at scale
• Gather learnings on what drives CTRs and conversions
• Apply those new strategies systematically across all products

Phase 1: Using AI-powered title optimization 

We used DataFeedWatch’s AI title generation to instantly deploy a 50/50 split test across the entire catalog. Roughly 10,000 SKUs received AI-rewritten titles, while the matched control group kept their original titles.

How to run AI A/B title testing with DataFeedWatch

You can use DataFeedWatch to run your A/B title tests and various Google products to track their performance. 

We ran the test for 30 days and followed the impact across paid and organic Google Shopping listings.

To track performance, we look across Google Analytics, Google Merchant Center, and Google Ads for a 360 view. GMC and Google Ads are really where the paid and organic Shopping picture lives, with GA4 layered on for the post-click view.

Example results from Google Analytics

For a step-by-step guide on analyzing A/B test results, visit the DataFeedWatch Help Center.

The breakthrough insights

Because of the scale of the test, we surfaced statistically meaningful data in 3 main areas.

Some of them were contrary to common assumptions:

1. Can brand placement be neutral?
   There is a common ecommerce assumption that putting the brand name first always improves performance. Our AI test proved otherwise. Moving the brand to the front, middle, or end of the title showed minimal difference in CTR.
   
   The importance of brand placement could be vertical-specific. For example, if a shopper is looking for a certain brand of shoe, the brand may hold more weight.
   
   It’s worth testing.
   
   
2. The 70-Character cliff is real
   Google Shopping truncates titles at 70 characters. Our data validated this hard limit. Performance dropped off meaningfully the moment a title pushed past the 70-character threshold.
   
   
3. Specs over descriptions
   B2B buyers tend not to respond to generic marketing descriptors.
   
   The biggest lifts in CTR and conversions came from surfacing hard data:

• Pack quantities
• Technical specifications
• Weights
• Concentrations
• Compatibility callouts

After performing your A/B title tests, pull similar observations from the data you have.

These will be foundational for in the next phase of this strategy.

Phase 2: Using our insights to scale results

Instead of leaving the AI running indefinitely, we took these insights and built a structured 5-slot template within DataFeedWatch using feed-attribute substitution.

The template enforces a strict 70-character cap and features a "fallback chain" that automatically drops less critical attributes first, ensuring the most search-relevant specs are never truncated.

Case study 2: legacy catalog cleanup (+7,000 SKUs)

The challenge

Our second client possessed a catalog of roughly 7,000 SKUs, but it was plagued by years of legacy data inconsistencies.

These are some recurring data configurations that we needed to clean up:

• Attributes formatted inside parentheses
• Erratic ordering of brand vs. product type
• Hidden technical specifications that buyers were actively searching for

Phase 1: Our AI approach

We deployed a similar 50/50 split test, giving roughly 3,500 SKUs the AI-rewritten treatment. Because organic free listings heavily rely on title clarity, we ran this test for a longer 60-day window to ensure we had fully closed attribution data on both sides for a clean read.

The breakthrough insights

1. AI excels at cleaning legacy messes
   The SKUs that saw the highest CTR gains were the ones with the most awkward original formatting. DataFeedWatch’s AI stripped out messy syntax and restructured them into natural, search-friendly phrasing.
   
   
2. Safety and compliance drive conversions
   In B2B healthcare, compliance is a non-negotiable. The AI surfaced critical purchase signals that buyers actively type into search bars, like: powder-free, latex-free, or BPA-free.
   
   
3. Fewer clicks, better customers?
   Interestingly, the AI-rewritten titles showed a unique pattern. They generated fewer but significantly higher-quality clicks.

Post-click engagement and conversion efficiency increased, proving that clearer titles successfully filter out irrelevant traffic.

Phase 2: Using our insights to scale results

Just like in the first scenario, the Phase 1 experiment served its purpose. We codified the winning attribute hierarchy into a repeatable DataFeedWatch template. Our client can now apply this optimized structure to all future product launches and seasonal merchandising seamlessly.

The takeaway: Don't guess, use AI for product feed management

What we’ve demonstrated across these engagements is an a repeatable approach to modern feed management. If you use AI blindly, you risk losing control of your data consistency. But if you treat AI as a rapid testing sandbox, you can extract invaluable data-driven insights from your own customers.

By pairing an experimental use of DataFeedWatch’s AI with the strict discipline of feed mapping rules, we've built an optimization loop that is continuously successful.