Egypt · Furniture · Living catalog

Every desk we could find, in one place.

A source-linked public catalog of 1,173 audited desk listings from 26 Egyptian sellers. Price, availability, and first-image references are rechecked across a rolling 48-hour window.

Source-linked

Every public row keeps its seller, original product URL, observation time, and audit trail. The catalog always leads back to the source.

Bilingual gate

English and Arabic evidence pass through strict and full-page checks. Ambiguous table words stay reviewable instead of being quietly guessed.

Last known, never invented

If a live check fails, the previous value stays visible with a stale state. Missing prices remain missing; AI never supplies commercial facts.

01 / Overview

A census, not a shortlist.

The broad crawl sought every plausible desk listed online by Egyptian furniture sellers. The public catalog is narrower: each of its 1,173 rows survived the strict classifier and a full listing audit.

The order matters. Broad collection protects recall; strict evidence gates protect the public claim. Combining those jobs would either lose desks too early or publish weak matches too confidently.

Non-desks remain in rejection logs, ambiguous products remain in review queues, and 37 records that failed the final full-page audit are retained outside the public catalog rather than silently disappearing.

The result is a browsable marketplace index, not a merchant. It narrows discovery and links out; sellers still own checkout, delivery, returns, and warranty.

What a record carries

  • seller_id
  • product_url
  • title_raw
  • brand
  • category_path
  • price_raw
  • price_value
  • currency
  • availability_raw
  • description_raw
  • observed_at
  • snapshot_ref
  • inclusion_reason
  • canonical_id
  • canonical_confidence

Original evidence remains in the audit layer. The public payload contains only the consumer-facing fields required to browse and verify a listing.

02 / Numbers

What survived each gate.

These counts come from generated reports, not a manual tally. The broad census and public catalog are separate layers; both remain visible here so the narrower claim is auditable.

What happened to 28,670 unique records

Kept as EGP desks 10.5%
Sent to review 8.6%
Non-EGP excluded 5.6%
Non-desk rejected 75.4%

Percentages are of 28,670 unique records. Two duplicate URLs are not shown.

MetricValue
Unique records reviewed28,670
EGP-priced desks2,997
Desks in stock2,234
Desks with no usable price963
Review queue2,462
Non-desk rejected21,620
Non-EGP excluded1,591
Canonical groups2,992
Sellers with a desk29
Classifier tests28
Strict EGP-eligible input27,079
Strict verified desks1,210
Final audit rejects37
Public catalog1,173
03 / Method

Six steps, most of them unglamorous.

Fetching is the easy part. The work is making different shops produce the same record shape, requiring evidence for public labels, and keeping disagreement and failure visible.

  1. Seller inventory

    The census became 139 sellers in five buckets: 57 retail, 38 B2B, 21 probe, 10 channel, and 13 dismissed. The bucket chooses the path, not the desk decision.

  2. Fetch

    Shopify /products.json and WooCommerce Store API first. JSON-LD for product pages, Magento where needed, and IKEA’s own category walk. Firecrawl only for blocked or rendered pages.

  3. Normalize

    One source shape across every seller: URL, title, brand, category, price, currency, availability, description, image references, timestamp, and snapshot.

  4. Desk gate

    Title first, then category, then description. Strong desk signals keep a record; strong non-desk signals reject it; anything genuinely ambiguous goes to review.

  5. Market gate

    Non-EGP prices go to a separate audit file. LE, L.E, and ج.م count as EGP. Unknown currency is kept, because unknown is not the same as foreign.

  6. Evidence audit

    Text, full-page content, and the first image are cross-checked. Unsupported labels become unknown; hard non-desks are rejected with evidence.

ExtractionDesks
JSON-LD direct1,072
Shopify JSON1,064
WooCommerce Store API857
Firecrawl product format4

Most desks came from structured data

Not from an LLM reading pages. That is deliberate: structured data is cheaper, faster, and easier to audit. Firecrawl was the fallback, not the default.

04 / Sources

Where the broad census came from.

These historical seller counts describe the 2,997-record broad collection, not the 1,173-row public catalog. Shopify and WooCommerce were the quick wins; IKEA and a long tail of shops needed targeted extractors.

SellerDesksSellerDesks
Chichomz706IKEA81
Nablco441Elansary81
Raneen394Stels42
Lamsa225Elhelow42
Mige204Woodway41
MecaWorks186Flaketech35
El Berry175Capital Furniture33
Al Zahra12714 other sellers184

The broad collection represented 29 sellers. The final audited public catalog represents 26; the remaining seller records are B2B, probe, channel, dismissed, uncrawled, or absent after stricter gates.

05 / Desk gate

What counts as a desk.

The classifier asks one question: is this product a desk? It does not ask whether it is wide enough, electric, or good value. Those are later questions.

Include

  • Office, computer, study, writing desks
  • Standing and sit-stand desks
  • Gaming and executive desks
  • Corner and L-shaped desks
  • Secretary, drafting, reception desks
  • Desks with a hutch
  • A workstation sold as a single desk

Exclude

  • Desk chairs and office chairs
  • Lamps, mats, pads, shelves, organizers
  • Drawer units, filing cabinets, pedestals
  • Bookcases, shelving, wardrobes
  • Dining, coffee, side, console tables
  • Dressing tables and vanities
  • Desktop computers
  • Converters sold as accessories

Send to review

  • A bare “table” that might be a work surface
  • “Workstation” that might be a cubicle system
  • “Bureau”, which can mean desk or chest
  • Arabic ترابيزه or طاوله without context
  • Bundles with a desk plus other furniture
  • A converter attached to a full desk frame
Arabic termLiteral meaningWhy it is tricky
مكتبdesk / officeFurniture or room
مكتبةbookcase / libraryNot a desk
كرسيchairOften a bundle item
ترابيزةtable (colloquial)Desk or dining table
طاولةtableSame problem, formal

2,462 records are still in review. Most are Arabic table words: 2,194 ترابيزه, 149 table, 61 طاوله, 57 workstation, one bureau.

Two IKEA TROTTEN 80×80 tables are in there on purpose. They are tables, not desks, and we did not want to force a decision just to make the queue look smaller.

06 / Quality

The pass that made the catalog defensible.

This was not cosmetic cleanup. Each decision either restored evidence, removed a false listing, normalized a consumer label, or made uncertainty visible.

1,210 → 1,173

Final listing audit

Title, full product page, price, seller, and first image were checked together. Thirty-seven records failed the final desk gate and remain in the audit rejection log.

92 fields

Audit schema

Source, heuristic, text, image, and cross-check fields stay in the audit model. The public page receives a separate 18-field consumer schema rather than raw model output.

48 hours

Rolling freshness

An hourly free-tier Worker checks up to 25 existing URLs. Confident price, availability, and image changes enter D1; failures retain the prior value and expose stale status.

422

No public price at audit

Quote-only, unexposed, and unavailable prices remain blank instead of being estimated. The products page distinguishes unknown currency and price-on-request states.

Failures remain visible

Sitemap images, URL false positives, title precedence, blocked pages, and unsupported AI decisions all changed the pipeline. Live fetch and parser failures now become explicit states instead of destructive updates.

07 / Limits

What the public catalog does not prove.

A clean interface can make uneven evidence look more certain than it is. These boundaries remain true even when a listing was checked recently.

  • Not the whole Egyptian market. Marketplaces, classifieds, offline inventory, and several large retailers remain incomplete or outside the audited set.
  • Not the merchant of record. The catalog links to sellers; it does not take payment, promise stock, arrange delivery, or administer returns and warranties.
  • Not a complete price comparison. 422 of the 1,173 public listings had no numeric public price at audit and no value is estimated for them.
  • Not laboratory verification. Dimensions, materials, mechanisms, and features reflect seller evidence and conservative normalization.
  • Not equally fresh in every field. Price, availability, and image each keep their own observation time; a partial parse updates only what it supports.
  • Not a warehouse check. Availability is recorded as published on the seller page, not confirmed by a phone call or inventory feed.
  • Not free of long-tail errors. A listing can still be misclassified, changed, or blocked. Its source link and freshness state remain visible so it can be checked.
  • Not an image archive. Thumbnails remain on seller servers and may be replaced or disappear independently of the catalog.

The honest version

Treat the 1,173 rows as a well-audited discovery set, not a guarantee that every desk in Egypt is represented or that every seller claim is correct. Verify the source page before buying.

08 / Next

What improves it next.

The public catalog is operational. The next work should improve evidence, coverage, and source reliability without weakening the admission gate.

  1. Measure live-fetch health. Observe which sellers allow Worker-origin checks, tune only evidence-safe parsers, and keep block rates visible.
  2. Resolve the broad review queue. Revisit the 2,462 ambiguous records—especially Arabic table terms—without mixing them into the catalog before audit.
  3. Add coverage deliberately. Audit marketplaces, classifieds, missing large retailers, and newly discovered products through the same gates before publishing them.
  4. Deepen sparse specifications. Improve evidence for load, motors, memory presets, warranty, delivery, and assembly while preserving “Not specified.”
  5. Strengthen change review. Use D1 history to surface unusual price or availability shifts for human inspection instead of silently trusting every parse.
  6. Build optional decision tools. Add transparent hard constraints and scoring only after the underlying fields are sufficiently complete for the user’s question.

If the goal is a desk decision, not a complete census

Start with the 1,173 audited listings, filter by the supported attributes, then verify the final few with their sellers. More crawling adds coverage; it does not automatically produce a better decision.