How we compute fair online prices

We read public-facing product listings from a small list of vetted Ethiopian sources. For every source we:

• Review the site's terms of service and robots.txt before enabling it.
• Identify ourselves with a User-Agent and contact email.
• Throttle requests (default 2.5 s between fetches per host).
• Never bypass paywalls, logins, captchas, or anti-bot systems.
• Never copy product photos or full descriptions — only title, price, city, condition, source URL, and collection date.

Each listing is stored once. Duplicate detection uses a SHA-256 hash of source URL + title + price, so a listing that's re-fetched tomorrow doesn't double-count.

Raw titles are messy ("Iphone 14 promax 256 gb used", "14PM 256 slightly used"). We parse them into structured fields — brand, model, variant, storage, RAM, condition, city, price-in-ETB — using a regex-driven rule set. When an OPENAI_API_KEY is configured, the rules' best guess is passed to an LLM that produces a strict JSON object, which is validated against a zod schema. If the LLM call fails or returns invalid data, we fall back to the rule-based output.

Normalized listings are then matched to a canonical product via: exact brand+model+variant → exact brand+model → alias substring → fuzzy Jaccard over the canonical name. Listings that fail all four steps appear in the admin "Unmatched" queue for review.

Ethiopian listings include occasional bogus prices (a missing zero, a placeholder "10 ETB"). Before computing fair prices we remove outliers using two filters in combination:

• IQR fence — drop anything outside [Q1 − 1.5·IQR, Q3 + 1.5·IQR].
• Hard ratio caps — drop anything below 25% of the median or above 4× the median.

We keep the cleaned distribution for everything that follows. The same filter is applied to the price-distribution chart on each product page, so what you see is what we compute against.

A used phone and a new phone of the same model have very different fair prices. Mixing them into one median tells you nothing useful. We compute fair-range stats once per condition:

• all — every listing, regardless of condition.
• new, used, refurbished, unknown — only listings tagged with that condition.

Per-condition stats are only written when the bucket has at least 3 listings — below that, the number would be noise. Each product page shows the available strata as chips; pick the one you're shopping for.

For each stratum we publish:

• Median — the price most listings ask. We display this as the headline number because the average can be skewed by extreme listings.
• Fair range — the inter-quartile range (P25 – P75). The middle 50% of honest listings sit here.
• Suspicious-low threshold — 75% of P10. A listing below this may indicate fraud, incorrect specs, damaged condition, a locked device, unpaid loans, or another issue. We flag, never accuse.
• Overpriced threshold — 125% of P90. Listings above this are well above the recent online range; the seller may be over-asking.

Confidence is derived from listing volume and source diversity:

• High: 25+ listings from 3+ sources.
• Medium: 10+ listings from 2+ sources.
• Low: fewer than 10 listings or only one source.

Treat Low confidence as a hint: the fair range is plausible but the sample is too thin to bet on. Confirm against at least one other source before paying.

• We are not a marketplace. We don't take payments and we don't broker deals.
• We cannot guarantee any listed device's authenticity, condition, or legal status. Always inspect in person, check IMEI, and verify warranty.
• Our fair ranges are statistical, not prescriptive. A skilled negotiator regularly beats them; a careless buyer can pay above them.
• Stale snapshots are clearly labeled. If a product shows "Updated 9 days ago", treat the number as a starting point, not the current market.