An agent that clusters return comments across thousands of orders to find the two or three fixable causes — a sizing chart, a packaging change — behind most of your returns.
Reason codes and free-text comments across channels and SKUs, with order values.
2
Cluster by real cause
Comments are grouped by meaning — "too small" and "ordered my usual size, didn't fit" are one cause, not two.
3
Price each cause
Returns per cause × average return cost (refund + shipping + damaged resale) turns complaints into dollars.
4
Match cause to fix
Sizing cluster → new chart and a fit note. Damage cluster → packaging. Confusion cluster → clearer listing and instructions.
5
Track the fix
After a change ships, the agent watches that cluster's rate — you find out if the fix worked, in numbers.
What it needs
Inputs & tools
ReturnsCodes + comments
CostsAvg return cost
CadenceMonthly + on demand
OutputPriced cause list
FixesYour team ships them
Your time≈ 1 hr/month
An apparel brand's biggest cluster was "fabric feels cheap" — at one price point. The fix wasn't the fabric; it was the listing photos setting a premium expectation the mid-tier product couldn't meet. New photos, returns down, same product.
Run this agent yourself
⚡
1 · Get a FreshCrawl subscription
The agent runs on live marketplace data. FreshCrawl plans start free (1,000 credits), paid plans from $29/mo — every plan includes MCP access and an API key. freshcrawl.io →
🔗
2 · Link the MCP to your LLM
In Claude or ChatGPT, add a custom FreshCrawl MCP connector — server crawl.onewally.com/mcp — and paste your FreshCrawl API key. Your LLM can now pull live offers, Buy Box history, and sales estimates itself.
📋
3 · Paste the prompt
Copy the prompt below into your LLM and name your store. The agent runs against the live data it can now reach — no code, no exports.
Copy the prompt
Starter prompt
Step 3 · Paste this into your LLM
# Returns Reason Miner
Use the FreshCrawl MCP connected to this chat for all live marketplace data.
MY STORE (fill in once):
- Store name: [your store name]
- Seller ID or storefront URL: [paste it — the agent uses it to find your catalog]
- Marketplace: [Walmart US / Amazon US]
First, use the FreshCrawl MCP to look up my store, pull my catalog,
and work only on my listings. If you can't identify my store from
what I gave you, ask me for my storefront URL before anything else.
RUN SCHEDULE (set this where you run the prompt):
- Run this prompt: monthly on the 1st, plus on demand.
- If your LLM app supports scheduled/recurring tasks, set it there; otherwise re-paste the prompt on that cadence.
Pull my returns from the last 90 days with reason codes and
comments. Cluster the comments by root cause (meaning, not
keywords). For each cluster: number of returns, % of total, and
estimated cost at my average return cost. Suggest the specific fix
(listing, packaging, product) for the top 3 clusters.
EXECUTION — READ CAREFULLY:
Analysis only. Fixes are shipped by my team.
The prompt gets you the shape. The reliable version runs on live data through the MCP — that's steps 1 and 2 above.
Clusters your 1–3 star reviews into a ranked fix list for the listing and the product.
Questions sellers ask
Reason codes are useless at my marketplace — still works?
Yes, that's the point: it reads the free-text comments buyers actually write, not just the dropdown code they picked at random.
How many returns do I need?
A few hundred returns gives stable clusters; smaller volumes still surface the top causes, labeled with their sample size.
Does it track fixes?
Yes — post-fix, it reports the cluster's return rate against the pre-fix baseline.
Run this on your store.
Every agent here runs on live FreshCrawl data through the MCP. Start free, connect your LLM, paste the prompt — or get the free Buy Box Loss Report and see the data first.