eBay Seller Feedback API: Read 1-, 6- and 12-Month Ratings Correctly
| Feedback row | m1: recent 1 month | m6: recent 6 months | m12: recent 12 months |
|---|---|---|---|
positive | count or unknown | count or unknown | count or unknown |
neutral | count or unknown | count or unknown | count or unknown |
negative | count or unknown | count or unknown | count or unknown |
This is the review table to build before choosing a chart. The cells are separate measurements, not nine numbers to collapse into one reputation score. A blank cell must remain distinguishable from a measured zero. The eBay seller feedback API exposes a nested seller profile that can support this display, but it does not turn a public summary into a seller-safety decision.
Put m1, m6 and m12 side by side—never add them
A recent month sits inside the recent six-month window, which sits inside the recent twelve-month window. Adding the three counts double-counts overlapping observations. Show one row per sentiment and one column per window, with the API field names available in an export or tooltip.
Do not silently rename these fields to exact 30-, 180- and 365-day periods. The documented labels here are m1, m6 and m12. Other interfaces may define their reporting periods differently. Likewise, subtracting two snapshots is not automatically a count of newly received feedback: an older observation can leave a rolling window while a new one enters it.
The useful question is “what summary was available when we checked?” rather than “how many new orders happened since yesterday?” Feedback counts and transaction counts are different measurements.
Request the profile with a seller username and marketplace domain
Use the public seller username from the profile or listing seller link, not an assumed match to a store display name. Keep the marketplace domain with the username. The verified request used musicmagpie and co.uk; that is a reproducible input example, not an endorsement or a promise about current values.
The request is a read using POST and a JSON body. Replace the platform-key placeholder locally before deliberately sending it. Copying the following command or opening the project page does not run it.
curl --request POST \
'https://api.opendata-api.com/api/project/ebay/seller-profile?key=YOUR_API_KEY' \
--header 'Content-Type: application/json' \
--data '{"domain":"co.uk","seller":"musicmagpie"}'
Keep a real key out of source control, screenshots and shared request logs. If you use an environment variable to construct the URL, remember that URL query strings can still enter logs. The example needs no seller-account session or private cookie.
Extract data.seller_profile without flattening every rating
There are two useful boundaries to check. First inspect HTTP status and the platform response's code, message and data. Then inspect the operation result: its documented successful envelope has ok, meta and data. The seller profile is data.seller_profile inside that operation result, not a sibling of the platform code.
A successful recorded request established an object profile with nested feedback_ratings and detailed_seller_ratings. It did not establish that every documented scalar will be present in every response. Diagnose a missing container rather than replacing it with an empty profile that looks like a seller with no feedback.
def extract_profile(platform_reply):
if not isinstance(platform_reply, dict):
raise ValueError("Expected platform response object")
if str(platform_reply.get("code")) != "200":
raise RuntimeError("Platform request did not succeed")
result = platform_reply.get("data")
if not isinstance(result, dict) or result.get("ok") is not True:
raise RuntimeError("Profile operation did not report success")
data = result.get("data")
profile = data.get("seller_profile") if isinstance(data, dict) else None
if not isinstance(profile, dict):
raise ValueError("Missing seller_profile object")
return profile
This deliberately refuses an error envelope instead of converting it into nine unknown cells and silently calling the request successful. Avoid logging the full request URL when reporting failures.
Keep feedback_score, feedback_positive_pct and positive_pct_12mo separate
The documented profile has separate names for a feedback score, a positive percentage and a twelve-month positive percentage. Preserve those names in your review record. A score is not a percentage; two percentages with different labels are not automatically interchangeable.
Do not fill one missing percentage from another or derive a replacement from the nine-cell matrix without a documented denominator. You would be introducing a new metric with potentially different treatment of repeat raters, reporting windows or excluded records. An interface can show “unavailable” more honestly than an impressive-looking computed value.
Display fields such as items_sold_display and followers_display are also not guaranteed integer counters. Retain their representation rather than parsing every decorated display string into an exact quantity. None of these fields establishes accepted-offer prices or transaction history.
Preserve missing counts and inspect detailed_seller_ratings on their own
Build the matrix only from nested objects actually returned. Missing, explicit null and a wrong type all deserve attention, but none should become zero by default. The following extractor accepts numeric counts and preserves every other representation as unknown; retain the original profile separately if you need to inspect schema variation.
SENTIMENTS = ("positive", "neutral", "negative")
WINDOWS = ("m1", "m6", "m12")
def feedback_matrix(profile):
ratings = profile.get("feedback_ratings")
ratings = ratings if isinstance(ratings, dict) else {}
matrix = {}
for sentiment in SENTIMENTS:
row = ratings.get(sentiment)
row = row if isinstance(row, dict) else {}
matrix[sentiment] = {}
for window in WINDOWS:
value = row.get(window)
is_count = type(value) in (int, float) and value >= 0
matrix[sentiment][window] = value if is_count else None
return matrix
A string count is not proof of zero; this code leaves it unknown for deliberate normalization later. The boolean type is excluded even though Python treats booleans as integers.
Keep detailed_seller_ratings in its own section. The verified evidence established a nested object, not a universal fixed list of dimensions. Inspect the returned keys at runtime and do not hardcode a dimension as mandatory merely because another seller or another interface exposes it.
Save dated observations without inventing a seller risk score
A useful internal record contains the caller's domain and seller, your own UTC observed_at, the three separate reputation metrics, the matrix and the returned detailed-rating object. Your timestamp says when your application observed the response, not when eBay last updated every underlying measurement.
If you also need the seller's active products, request seller-items with the same explicit username and domain. Inspect its pagination rather than treating one page as a complete storefront. That separate listing scan should not alter or “correct” the feedback summary.
For repeated observations, compare like-named fields and preserve unknowns. A sudden change is a prompt to inspect the record and request context, not a reason to manufacture a proprietary risk score or announce a new sale. Current-data responses do not create a historical archive unless your application retains dated observations itself.
What this profile cannot establish about a seller
The summary cannot provide individual feedback text, prove transaction volume, verify identity or guarantee that a seller is safe. Positive execution status establishes that the operation completed with a usable structure, not that every business value is complete, fresh or relevant to a particular purchase.
Use the matrix as one input to human review. Keep overlapping periods separate, label unknowns, and preserve the boundary between public feedback measurements and decisions that require additional evidence. That is a more defensible result than adding the windows and attaching an unexplained “trust” label.