TIKTOK SHOP · AUG 18, 2026 · 6 MIN
Stop A/B Testing Outreach: Build a Self-Improving TikTok Shop Outreach System
Replace one-off A/B tests with a five-step loop where Claude sends three message variants through the Cruva MCP, grades them on response, sample request, and open rate, and stores the winners in a ledger it learns from.
Stop A/B testing your TikTok Shop outreach messages and build a system that improves its own response rate instead. The loop has five steps: split a Cruva creator list into three CSVs with three message variants, review every message for personalization and AI copy, send all three through the Cruva MCP, grade them on response rate, sample request rate, and open rate, then store the winners with their KPIs in a ledger that Claude reads before drafting the next round.
Key takeaways:
- An A/B test produces a winner. A ledger produces a system that keeps winning.
- The quality of the inputs decides the quality of the loop. Personalize hard and strip any AI copy before the first send.
- Grade on three KPIs, because response rate alone cannot tell you whether the hook or the offer did the work.
- Agencies compound this fastest: winning patterns across 20 to 25 brands become a data set no single brand can match.
Most brands test outreach the way they were taught to test landing pages. Two versions, pick the winner, move on. The problem is that the winner gets reused until it stops working, and the learning behind why it won never gets written down. Next quarter the team starts from zero again.
Here is exactly how the self-improving version works.
What is a self-improving outreach system?
It is an outreach process where every send is graded, every winner is stored with the reason it won, and every new message is drafted from that stored record. Claude does the drafting, sending, and grading. The Cruva MCP handles the creator data and the actual messages. The ledger is a folder of files. There is no special software beyond that.
How do you build it?
1. Build your creator lists
Pull a list of creators from Cruva and separate it into three separate CSVs. The three groups should be comparable, so the message is the variable rather than the audience.
Write three versions of the message, tweaking the offer, the copy, and any other wording you want to test. Then ensure Claude loads each outreach message against one of the three CSVs, so every creator receives exactly one variant.
Load creators-a.csv with message A, creators-b.csv with message B, and creators-c.csv with message C. Confirm the counts match before sending anything.
2. Review your messages
Before anything sends, read all three messages with two questions in mind:
- Is every message extremely personalized to the list it is going to? A message written for beauty creators should read like it was written for beauty creators, down to the product language and the way the offer is framed.
- Is there any AI copy in any of these messages? Generic phrasing, filler enthusiasm, and sentence structures that sound generated all get cut.
Having the highest quality inputs here minimizes the chance of AI slop in future steps. The ledger in step 5 learns from what you send. If you send slop, it learns slop.
3. Send the messages
Have Claude connect to Cruva's MCP and send all three messages against their lists. Wait until each message has finished sending before doing anything else. Grading a half-sent variant against a fully-sent one will skew the KPIs.
4. Grade the messages
Validate each message on the following KPIs:
- Response rate. Did creators reply at all? This is the hook and the personalization doing their job.
- Sample request rate. Did the reply turn into a request for product? This is the offer doing its job.
- Open rate, for emails. This isolates the subject line from everything else.
Three KPIs matter because a message can win on one and lose on another. A strong hook with a weak offer will show high response and low sample requests, and that tells you exactly which half to keep.
| KPI | What it measures | What to change if it is low |
|---|---|---|
| Response rate | Hook and personalization | Opening line, how the creator is addressed |
| Sample request rate | Offer strength | Commission, free product terms, the ask itself |
| Open rate (email) | Subject line | Subject line only |
5. Store the winners in a ledger
Store the winning outreach messages in a folder with the copy and the KPIs side by side. This is the step that turns a test into a system. Claude starts to learn what works and why, because it can read the record before it drafts anything new.
Future messages draw from this context and become hyper optimized over time. Eventually you do not have to spend hours optimizing messaging templates, because Claude already knows your winning patterns.
Before drafting the next outreach round, read every entry in the winners ledger. Identify the offer structures and opening lines with the highest sample request rate and draft three new variants that build on them.
Why does this matter more for agencies?
Pro tip for agencies: if you compile these winning patterns across a book of 20 to 25 brands, your AI outreach becomes very hard to beat, because it is trained on a high quality data set no individual brand can assemble. A single brand learns what its own creators respond to. An agency ledger learns what creators respond to across categories, offer types, and seasons.
Where does this sit in the affiliate stack?
The ledger approach is the outreach layer of a larger system. It works best when the creator lists feeding it are already targeted, which is where finding undiscovered TikTok Shop creators with Claude Code and the Cruva MCP comes in. And if you are still deciding whether to run outreach through an off-the-shelf product or a workflow you own, AI agents for influencer and affiliate outreach covers that decision.
The Clankers 90-Day Revamp builds the loop inside your Cruva account and hands the ledger to your team.
Frequently asked questions
Why stop A/B testing TikTok Shop outreach messages?
A single A/B test tells you which of two messages won once. It does not carry that learning into next week's outreach. A system that grades every send and stores the winners in a ledger keeps improving, because each new message is drafted from the accumulated record of what worked and why.
How many message variants should you send at once?
Three. Pull a creator list from Cruva, split it into three CSVs, and write three versions of the message that vary the offer, copy, and any other wording you want to test. Claude loads one message against each CSV.
What KPIs grade an outreach message?
Response rate, sample request rate, and open rate for emails. Response rate tells you the hook worked, sample request rate tells you the offer worked, and open rate isolates the subject line on email sends.
What goes in the winners ledger?
The winning message copy alongside its KPIs, stored in a folder Claude reads before drafting the next round. Over time Claude draws on that context and the messages become hyper optimized without anyone rewriting templates by hand.
How does this work for an agency with many brands?
Compile the winning patterns across a book of 20 to 25 brands. The ledger then reflects a high quality data set of what creators respond to across categories, and every new brand's outreach starts from that record instead of from scratch.
Written by Sohun Sanka, founder of Clankers, an operator practice that builds AI systems and automation for social-commerce brands and agencies inside the tools they already use. This post expands on a LinkedIn post Sohun published on August 4, 2026.
Build the loop once, then let it learn
If your team is still rewriting outreach templates by hand every quarter, the ledger replaces that work with a record the system reads on its own. The Clankers 90-Day Revamp diagnoses your outreach bottleneck, builds the loop inside Cruva, and transfers it to the people who run it.