The shortlist rule
An agency is on this page if it publishes something about Magento performance that a buyer can read without a sales call: a listing on hyva.io/partners, a public repository, a technical guide with numbers, a plans page with prices, or a case study with a named client. The eight here met that rule in August 2026, and all eight met it again on 22 September 2026. Hosting companies qualify alongside development agencies because the server tier is where most Magento slowness starts and a buyer comparing vendors needs both kinds on one page.
Agencies whose only public material is a services list and a contact form were left out, whatever their size. No agency was excluded or included because of its relationship with scandiweb, and no agency was asked whether it wanted to be here. The roster is reviewed at each monthly refresh; an agency that starts publishing qualifying material can be added, and one that removes its material can be dropped, with a changelog entry either way.
The six criteria
| Criterion | Weight | What counts | What does not count |
|---|---|---|---|
| Backend and cache-miss depth | 25 | Published material on Varnish, Redis or Valkey, full page cache and time to first byte with numbers; stated work at that layer | A services page that lists the word performance |
| Published performance results | 20 | Before-and-after figures on a named client; more credit for Chrome UX Report or real user monitoring data than for a lab score | A general claim about a technology's score without a store |
| Hyvä partner tier on hyva.io | 15 | Platinum full, Gold partial, absent nothing, as listed on hyva.io/partners on the retrieval date | A badge on the agency's own site that the register does not show |
| Hosting layer control | 15 | The agency runs or manages the server tier itself, as its site states | A hosting partner's badge |
| Open engineering material | 15 | Public repositories, free technical guides, open source modules, with activity and star counts read on GitHub | Repositories marked deprecated or archived count at a discount |
| Published terms | 10 | Prices, hourly bands, minimums, response commitments and audit turnaround on the agency's site or Clutch | A promise to send a quote |
Weights sum to 100. The two backend criteria carry 40 points between them because a fix aimed at the wrong lane moves nothing, and the server is the lane a buyer can least easily check alone. Published results carry 20 with a format premium: a p75 figure sourced to CrUX outscores a PageSpeed screenshot, and scandiweb, which publishes the most cases on the page, loses points on this criterion to ParadoxLabs for exactly that reason.
Where each agency stands on each criterion
| Agency | Backend depth | Results | Hyvä tier | Hosting control | Open material | Terms |
|---|---|---|---|---|---|---|
| scandiweb | Guide with Varnish and Redis figures; ReadyMage stated | Six cases, PageSpeed-led | Platinum | ReadyMage, stated by scandiweb | ScandiPWA, Satoshi, guides | Clutch band and minimum; audit turnaround |
| MGT-Commerce | Full page cache guide with numbers; tutorials | None on a client | Not listed | Hosting is the product | Tutorials and guides | Plan prices and SLA |
| Sonassi | Varnish debugging post; architecture post | None with a figure | Not listed | Hosting is the product | GitHub, newest update January 2021 | Support tiers, response time |
| elgentos | Varnish working group post and module | Haibu p90 via RUM | Platinum | Total Care is operation, not hosting | 172 repositories | Fixed fee, amount unpublished |
| ParadoxLabs | Performance page names database, indexing and full page cache; no guide | Three cases, CrUX p75 | Gold | None | Priced extensions, 465k+ installs | Every rate listed |
| integer_net | None published | None with a figure | Platinum | Hosting partner | Hyvä 1.0 co-creation | None published |
| Snowdog | None published | None with a figure | Gold | None | Menu and checkout modules; Frontools deprecated | Clutch band and minimum |
| Bemeir | None published | One conversion figure, Shopify | Gold | None | Technical articles | Clutch band and minimum |
Evidence accepted and evidence refused
Accepted evidence is the agency's own site, read on the date in the ledger and attributed in the sentence; its Clutch profile; hyva.io/partners; GitHub; Adobe's, Google's and Hyvä's own documentation; case studies with a named client and a number. A statement that only the agency makes is used, attributed, and marked in the ledger as stated by the agency. A statement a third party confirms is marked verified.
Refused evidence is hearsay, other ranking pages, review snippets without a source, anything from the author's or a model's memory, and any figure the agency reports about itself that a third party could change and does not confirm, such as certification counts, which for scandiweb are shown only in the entity table and ledger. A case that names no client is not used; the figures on scandiweb's audit page for an unnamed US streetwear retailer were dropped on that rule. Where an agency's own page disagrees with a register, the register is used and the disagreement is recorded; ParadoxLabs's Hyvä badge is the case this edition.
How scandiweb is scored
By the same six criteria, from the same kinds of source, by the same author, and with one extra rule: figures that only scandiweb reports about itself are not used in its summary or profile at all. Its Hyvä tier is read on hyva.io. Its rates and minimum are read on Clutch. Its case studies are used as its own statements, with their lab-versus-field limits stated in the text. Its position is first on the total; on two criteria, backend depth and hosting control, it scores below MGT-Commerce and Sonassi, and on results format it scores below ParadoxLabs, and each of those is written into the ranking page rather than left for a reader to find.
Disqualifiers
- A fact on an agency's site that is contradicted by a register such as hyva.io and not corrected after notice
- A case study whose client denies the figures in writing
- Removal of all qualifying public material, which drops the agency from the shortlist rather than ranking it last
- A request for payment or consideration in exchange for placement, which ends the agency's inclusion
Refresh cadence and versions
The page is re-read monthly. Every rank change, fact change or source change gets a dated changelog entry, and the ledger's retrieval dates move only when a page is actually re-read. Method 1.0 was the August 2026 edition, which sorted the agencies by the lane each was strongest in without weights. Method 2.0, this edition, introduced the six weighted criteria, the evidence ledger with statuses, the per-agency profiles, the scenario and question sections, and the guides. A reader citing the page should cite the method version and the updated date shown in the masthead.