Methodology · November 2026

From a broad census to a living catalog.

How 28,670 collected records became 1,173 public desk listings—and how those exact URLs are rechecked without hiding uncertainty or overwriting the last trustworthy value.

1,173 public listings 26 represented sellers Rolling 48-hour checks Source-linked and reproducible

Four counts, four different claims.

The broad collection, strict desk set, audited catalog, and live overlay answer different questions. They should not be presented as interchangeable totals.

28,670unique product records collected
27,079EGP-eligible records sent through the strict pass
1,210desks verified before the full listing audit
1,173public listings after 37 final rejects

The public products page is generated from desks_audited.jsonl, not from the 2,997-record broad candidate set. The larger numbers remain useful as an audit trail and as evidence of coverage, but only the 1,173-row set is presented as the current catalog.

The operating rule

Collection can be broad. Public claims must be narrow enough to defend, and live updates may change price, stock, and imagery—not the membership of the audited catalog.

Collection and normalization.

The census began with 139 seller records across retail, B2B, probe, channel, and dismissed buckets. Structured sources were preferred because they are cheaper to collect, easier to compare, and more auditable than rendered-page guesses.

  1. Inventory sellers

    Record the seller, route, platform, and coverage state before collecting any products.

  2. Fetch structured sources first

    Use Shopify catalogs, WooCommerce Store API, JSON-LD, sitemaps, and targeted category walks. Browser rendering is a fallback.

  3. Normalize one record shape

    Keep seller, URL, original title, category, price, currency, availability, description, image references, observation time, and snapshot provenance.

  4. Apply the bilingual desk gate

    Read English and Arabic evidence with title-first precedence. Ambiguous “table” and “workstation” records remain reviewable instead of being forced into a clean total.

  5. Run text and image audits

    Normalize dimensions, form, materials, colours, and visible features while retaining source and cross-check status.

  6. Publish through a separate schema

    Reduce the audit model to familiar consumer labels. Raw model output and conflict fields stay outside the public payload.

What qualifies for the catalog.

A product must be evidenced as a desk and remain relevant to the Egyptian market. Office, computer, study, writing, gaming, standing, executive, reception, drafting, corner, and L-shaped desks qualify. Chairs, lamps, pads, shelves, drawer units, cabinets, dining tables, vanities, and desk converters sold as accessories do not.

Market gate

Known non-EGP prices are excluded from the Egypt catalog. EGP, LE, L.E., and ج.م are normalized as Egyptian pounds. Unknown currency remains unknown rather than being silently converted into EGP.

Ambiguity policy

Generic Arabic table terms such as ترابيزة and طاولة require context. A plausible desk signal can move a product to review; only strong non-desk evidence produces a hard rejection.

The full listing audit.

The strict 1,210-row set was checked again against title, full product-page text, price, seller, and first image. This final pass rejected 37 products that did not survive the combined evidence gate, leaving 1,173 public rows from 26 sellers.

Public attributePrimary evidenceConservative behavior
Desk type and shapeTitle and product text; image as supporting evidenceUnclear records use “Other” or “Not specified”
PowerExplicit motor, electric-height, sit-stand, manual, or gas-lift evidenceA power outlet, LED strip, or image alone cannot create an electric claim
DimensionsStructured fields and stated product textConflicting or absent measures remain unknown
Material and colourText plus image cross-checkPublic value is normalized; audit conflicts remain retained internally
Price and stockSeller-published structured or page dataNever generated by AI; missing values are not estimated

AI-assisted fields are constrained by strict output schemas and evidence requirements. Core commercial facts—URL, seller, price, currency, and availability—come from seller data, not a language model.

What the public labels mean.

Fixed height
No supported standing or height-adjustment mechanism was evidenced. It is a catalog decision, not a laboratory claim that the product can never be configured differently.
Not specified
The available evidence did not support a public value. It does not mean “no,” and it is always filterable as an unknown.
Price on request
The seller invites a quote or withholds a numeric public price. No estimate is substituted.
In stock
The seller page currently publishes an available state. It is not a warehouse or telephone verification.
Last known value
The most recent confidently parsed value, retained when a later page request is blocked, times out, or cannot be parsed safely.

How listings stay current.

The deployed site keeps the audited catalog static and stores only its live commercial overlay in Cloudflare D1. One hourly cron selects up to 25 due rows. Across 48 hours that provides 1,200 available checks—slightly more than the 1,173 URLs in the catalog—without approaching the free Worker limit of 50 outbound subrequests per invocation.

  1. Select due URLs

    Only active URLs already in the audited catalog are eligible. The refresher does not discover or add products.

  2. Fetch the seller page

    Requests are distributed across the window and capped by timeout, response size, and concurrency.

  3. Parse conservative sources

    Product JSON-LD and commerce meta tags come first, followed by narrow Shopify, WooCommerce, and semantic HTML fallbacks.

  4. Update fields independently

    A confident stock result can update even when price is absent. Unobserved fields retain their own earlier values and timestamps.

  5. Expose freshness

    The API returns check status and per-field observation times. The page shows checked, awaiting, or stale states instead of implying every value is equally fresh.

Images are referenced, not copied

The catalog uses the first valid HTTPS image URL published by the seller. Browsers load it lazily from the seller’s host with no referrer; the project does not download, proxy, or store image bytes.

Failure is a state, not a deletion.

Seller pages can block data-center traffic, return a temporary error, change markup, or disappear. A failed fetch or parser miss updates check status and retry metadata only. It does not replace a known price with null, mark a product in stock without evidence, or erase its image.

ResultCatalog behaviorVisible state
Confident extractionObserved fields and timestamps update; changes enter historyChecked recently
Partial extractionOnly the supported fields updateChecked, with older field timestamps retained
403, 429, timeout, or parser missAll last known commercial values remainLatest check failed; last verification shown
404 or 410Availability becomes out of stock; last price and image remain for contextUnavailable at source
Live API unavailableThe browser keeps the generated static catalogAudit snapshot notice

Limits and buying caveats.

  • This is not every desk in Egypt. It is the audited public subset of a seller census; marketplaces, classifieds, offline inventory, and blocked sources remain incomplete.
  • Seller truth can lag reality. Published stock does not guarantee a warehouse unit, delivery date, assembly service, return right, or warranty.
  • Attribute coverage varies. Dimensions, load limits, motors, materials, and features are absent or conflicting on many product pages.
  • Images are external dependencies. Sellers can replace or remove them; a broken thumbnail does not remove the listing.
  • Automated parsing is conservative, not infallible. The original seller link remains the final place to verify a purchase.

Before buying

Reconfirm price, availability, dimensions, delivery, assembly, returns, and warranty directly with the seller. The catalog reduces search cost; it does not become the merchant of record.

Reproduction and public boundary.

The public build is an explicit allowlist. Raw pages, audit caches, credentials, source JSONL, scripts, reports, and local paths are excluded from deployment. The Worker serves static assets as a fallback even if D1 cannot answer.

python scripts/export_products_web.py
python scripts/generate_d1_seed.py
python scripts/build_site.py
python -m unittest discover -s tests -q
node --test worker/test/*.test.js

The generated D1 seed is idempotent: a new seed can update catalog membership and static metadata without overwriting values already observed by the live refresher.