Methodology
What this test can and cannot tell you
A 14-order test is not research in the academic sense. It is a structured diary with numbers, and its value depends on being explicit about the design and the limits. Both are below.
1. Research design
- Research question
- Does a spreadsheet-first workflow make agent buying predictable enough to recommend?
- Sample
- 14 orders, 4 agents tested in parallel, 90 days (late July — mid October 2026)
- Unit of analysis
- The order, not the item. A 3-item parcel counts once.
- Spend
- About $2,184 all-in (items + agent fee + domestic + international freight)
- QC actions
- 5 rejections, 3 second-round photo requests per order on average
- Comparison set
- 211 table sites, 110 community pages, 163 domain registrations crawled October 2026
2. Scoring
The four ratings on the review page are weighted, not averaged. QC carries the most weight because it is the only stage where the buyer can still change the outcome.
| Dimension | Weight | What is measured |
|---|---|---|
| Price & agent fee | 20% | Effective cost per parcel after fee and freight, benchmarked against three other agents on the same day. |
| QC photos | 30% | Whether the default gallery catches the four common defects, and what a second request costs in time. |
| Shipping speed | 25% | Days from warehouse arrival to international dispatch, not door-to-door (freight lines vary by destination). |
| Support | 25% | Time to first useful answer, and whether the answer came from someone with access to the order. |
Weighted overall: 3.9 / 5. Unweighted average of the four: 4.0 — the difference is entirely the QC weight.
3. The spreadsheet study
The wider claims — 211 table sites, 110 community pages, the domain-age distribution — come from a separate crawl: eleven keyword queries, the top 30 results per query, the HTML of each result fetched and parsed for title, meta description, headings, technology stack and outbound links, plus WHOIS/RDAP registration data for 163 of the domains.
Rows were classified by hand into table sites, social/community pages, platform-hosted pages and official brand sites. Two researchers would disagree on a handful of edge cases; the category totals are robust to that, the per-site labels less so.
4. Five limits
- 01One author, 14 orders, four agents. This is a structured personal test, not a statistically powered study.
- 02All orders shipped to Germany. Freight lines and timing do not transfer to other destinations.
- 03Coupon codes were tested on the same three accounts; a code that failed for me may work for a first-time buyer.
- 04The 211-site crawl records what ranked in October 2026. These results move month to month.
- 05Product legality is out of scope. The site documents the buying workflow; it does not vet the goods.
5. Corrections
One correction is already live: an early draft classified one site as a countdown page when it was a Cloudflare challenge screen. The claim was removed from the review and the reasoning kept in the research notes. Corrections are logged on the page they affect rather than silently edited.
If you spot an error, the contact address is on the imprint. Corrections are made within a week and noted in the affected page’s update date.