Seller intelligence starts with identity and observation dates
Build seller-level research from stable store identities, observed catalogue coverage and clearly labelled inferences.
In this article
THE KEY IDEA
A seller's visible catalogue is an observation, not a complete inventory or a verified revenue statement.
01Resolve the seller before aggregating listings
Store names can change, and similar names can belong to different businesses. Preserve the marketplace's seller identifier where available, the public store URL and a separate internal entity ID. Do not merge sellers across platforms solely because their display names match.
When a cross-platform relationship matters, require supporting evidence and keep the match status. An unresolved identity should remain unresolved. A mistaken merge can contaminate every later measure of assortment, pricing and activity.
02Describe observed coverage precisely
Record how seller listings were discovered: store pages, category search, keyword results or supplied URLs. Each method exposes a different slice of the catalogue. Keep pagination and filtering boundaries with the collection run.
Count observed unique listings after deduplication, not raw rows. If search results are capped or some pages fail, report that limitation. Avoid labelling the result total inventory unless the scope and source behaviour support that claim.
| Metric | Honest interpretation |
|---|---|
| Observed listings | Unique listings found within this collection scope |
| First seen | First observation in your dataset |
| Last seen | Most recent successful observation |
| Price range | Range of comparable observed offers |
| Rating count | Published count at a stated observation time |
03Keep published facts apart from estimates
Visible prices and rating counts do not directly establish sales or revenue. If a model estimates an outcome, label it as an estimate and disclose the assumptions. Do not let an inferred field appear next to observed fields without a clear distinction.
Similarly, a newly observed listing is not necessarily newly created. Your crawler may have discovered it late or gained access to another category. Use source-published dates when available and preserve the difference between those dates and your own observation history.
04Compare consistent seller cohorts
Changes in discovery coverage can make a seller appear to expand or contract. Compare the same collection scope over time and report when that scope changes. Keep failed runs from being interpreted as sudden catalogue collapse.
Normalise categories only with a documented mapping. A seller's source category can have a different meaning from your internal taxonomy. Preserve the original value so analysts can revise the mapping without recollecting every page.
05Use the output for reviewable decisions
A seller dataset can support assortment research, partner investigation or anomaly review. It should not imply misconduct, financial performance or business health without supporting evidence. Keep source references and uncertainty with material conclusions.
Start with a small set of well-defined measures and validate them manually for representative sellers. A modest dataset with known coverage is more defensible than an elaborate score built from uncertain identities and unexplained assumptions.
References & further reading
Put it into practice.
Try a public-page workflow, or scope a maintained dataset with our team.