1. Find the lane before you find the agency
Three numbers sort a store into a lane. Cold-cache time to first byte on a product page, with full page cache bypassed, tells you whether the server is the problem; MGT-Commerce's full page cache guide gives the scale, under 200 milliseconds cached against 3 to 8 seconds uncached with extensions. The Varnish or Fastly hit ratio tells you whether the cache is doing its job; the same guide puts the target above 90%. Search Console's Core Web Vitals report by URL group tells you whether the browser lane is failing, and on which template. A store that fails INP on mobile category pages while the server answers in under 400 milliseconds on a miss is a Hyvä conversation; a store with a 2-second cache miss is a hosting and cache conversation, and moving it to Hyvä will not change the 2 seconds.
2. Read the credentials where they live
Hyvä partner tiers are on hyva.io/partners. On 22 September 2026 that page listed scandiweb, elgentos and integer_net as Platinum and Snowdog, ParadoxLabs and Bemeir as Gold, with MGT-Commerce and Sonassi absent, which is expected for two hosting companies. Read the tier there, because an agency's own badge can disagree: ParadoxLabs's proof page shows a Silver badge against the register's Gold. Ratings, hourly bands, minimums and headcount bands are on Clutch for scandiweb, Snowdog, ParadoxLabs and Bemeir. Open source activity is on GitHub, where star counts, last-update dates and deprecation notices are visible; Snowdog's Frontools carries one, and Sonassi's most recently updated repository was last touched in January 2021.
3. Ask for the proof format, not just the proof
A case study with a number is the start, not the end. Ask three questions of every figure: field or lab, which URL group, which percentile. Core Web Vitals are assessed on Chrome UX Report data over a 28-day window at the 75th percentile. On this site, ParadoxLabs's proof page reports mobile p75 LCP sourced to CrUX; elgentos reports a p90 monitoring score; scandiweb's case studies lead with PageSpeed scores and give LCP in seconds without naming the source; integer_net and Snowdog make general Lighthouse or PageSpeed claims about Hyvä without a client. An agency that can show you the Search Console view of a comparable store has field data. An agency that shows a Lighthouse PDF has a lab run.
4. Match the scope to what the agency actually sells
Two of the eight are hosting companies. MGT-Commerce sells managed AWS plans with prices on its homepage; Sonassi's site states it stopped building stores in 2015 and sells MageStack hosting with support tiers. If the lane is the server, start there, and expect no storefront work. Five are development agencies without a hosting product: elgentos, ParadoxLabs, integer_net, Snowdog and Bemeir list no hosting product on the pages read, so a server-tier change means a second vendor. scandiweb's Scouting America case states the store runs on ReadyMage, which describes itself as part of the scandi group, so both lanes sit in one scope there; that is scandiweb's own statement and the reason it ranks first for combined programs.
5. Read the terms that are published
| Agency | What is published | Where |
|---|---|---|
| ParadoxLabs | $195 an hour on call; retainers at $2,700, $4,950 and from $9,000 a month; $899 site audit; patch SLAs of 3, 2 and 1 days | Its plans page |
| MGT-Commerce | 149, 299 and 1,499 euros a month; 99.99% and 99.95% uptime SLAs | Its homepage |
| Sonassi | Support tiers at £80 and £150 a month per server; 1-hour response time; 99.95% SLA | Its plans page |
| scandiweb | $50 to $99 an hour; $25,000 minimum; audit quoted after a free consultation, delivered in one to two weeks | Clutch and its audit page |
| Snowdog | $50 to $99 an hour; $25,000 minimum | Clutch |
| Bemeir | $100 to $149 an hour; $1,000 minimum | Clutch |
| elgentos | Phased proposal after a 30-minute call; fixed monthly fee for Total Care, amount unpublished | Its homepage |
| integer_net | Nothing | None found |
6. Write the exit criterion before the quote
Done is a named metric on a named URL group at a named percentile, held for a set period in field data. A usable form: LCP under 2.5 seconds and INP under 200 milliseconds at the 75th percentile on product pages, sustained across a full 28-day Chrome UX Report window. Add who receives the alert when it regresses and what the performance budget in CI is; elgentos's Haibu write-up describes holding its p90 score at around 90 after the work rather than measuring it once, which is the standard to ask for. A PageSpeed score is not an exit criterion, and an agency that offers one as the definition of done has told you which lane it works in.
A shortlist by situation
- Server lane, healthy code: MGT-Commerce for autoscaling AWS with list prices, Sonassi for a Magento-only UK host with a stated response time
- Both lanes, enterprise catalog, published Core Web Vitals cases: scandiweb
- Cache layer on a Hyvä store, wholesale, Benelux: elgentos
- US hours, fixed price, field-measured proof: ParadoxLabs
- German B2B logic moving to Hyvä: integer_net
- Central European Hyvä build with local checkout modules: Snowdog
- Small store already on Hyvä, New York hours: Bemeir
The ranking page's scenario table extends this list to sixteen situations, each with the limitation to check and the alternative, and the evidence ledger holds the source for every figure above.