MIT-Haiti Initiative · Technical Report · November 2026

mit-ayiti.net Migration Technical Report

The full measurement record behind the migration from WordPress to a Next.js front end on Pantheon, including method, per page figures, the effect of a CDN change made during testing, and the limits of what was measured.

Lighthouse 12.8.2 40 runs, 4 page pairs Lantern simulate, 7 Mbps, 100 ms, 4x CPU 7 November 2026

Method

  1. Identical URLs on both sites. The migration preserved the existing URL structure, so each pair compares the same content rather than an approximation of it. Four pairs were tested, covering the home page, the resource library index, a single resource page and the blog index.
  2. Three throttling profiles. A Haitian mobile profile at 7 Mbps and 100 ms round trip time with a 4x CPU slowdown, Lighthouse's own Slow 4G default for an industry comparable figure, and an unthrottled desktop profile. Medians of three runs on the Haitian profile, single runs on the others.
  3. Lantern simulated throttling. The page loads at full speed and the slow connection is modeled afterwards. This is the method PageSpeed Insights uses and is far more repeatable than applying real packet delay, but it is a model rather than a measurement of a real connection.
  4. Simulated from the United States. No measurement was taken in Haiti. Real geographic latency is not modeled by this method.
  5. Third party content is included, because readers download it. The old site loads Google Analytics. Both sites embed the Google Drive document viewer on resource pages.

Why 7 Mbps was chosen

Two independent real world samples converge near that figure. SpeedOf.Me in browser tests report a median of 6.9 Mbps and 101 ms latency for Haiti. Speedgeo puts Digicel's mobile average at 7.1 Mbps. Ookla's Speedtest median is considerably more optimistic at 16.08 Mbps for the same carrier, drawn from self selected tests. The lower and more conservative figure was used throughout. Internet penetration in Haiti was 39.3% as of October 2025 according to DataReportal.

Response timing

Medians of three runs on the Haitian mobile profile. Time to First Byte improved by between 80% and 93%, and First Contentful Paint by between 48% and 55%.

Haitian mobile profile, 7 Mbps, 100 ms, 4x CPU
Page TTFB before TTFB after FCP before FCP after Speed Index before Speed Index after
Home 0.95s 0.07s 1.59s 0.73s 3.13s 1.20s
Resource library 0.52s 0.10s 1.57s 0.71s 2.27s 1.55s
A resource page 0.49s 0.07s 1.73s 0.90s 2.35s 1.80s
Blog index 0.46s 0.06s 1.38s 0.67s 1.74s 1.38s

The old site sent Cache-Control: no-store, no-cache, must-revalidate on every response, so each visit regenerated the page in PHP. The new site serves a pre-rendered page from a CDN edge cache. Variance reflects that difference. Home page TTFB samples on the old site ranged from 0.89s to 2.03s across the full test set, against 0.04s to 0.14s on the new site.

Largest Contentful Paint

LCP is the one timing measure that did not improve uniformly. It is reported here in full because a reader running their own test against the live site will see these numbers.

Page LCP before LCP after Change
Home 1.66s 1.96s 18% slower
Resource library 2.47s 2.96s 20% slower
A resource page 1.73s 1.45s 16% faster
Blog index 1.76s 1.57s 11% faster

The cause is JavaScript volume rather than network time. The home page ships roughly 1.4 MB of uncompressed JavaScript, which gzips to around 440 KB but still has to be parsed and executed on a device running at a quarter speed under the test profile. Reducing that payload is ongoing work and is not a property of the hosting change.

Transfer size

Transfer is what crossed the wire. Content is what those bytes represent once decompressed, which is a better measure of how much the pages were actually reduced.

Page Requests Transfer before Transfer after Content before Content after
Home 138 to 60 2,520 KB 1,007 KB 6,850 KB 2,107 KB
Resource library 92 to 77 1,419 KB 1,193 KB 2,826 KB 2,397 KB
A resource page 133 to 89 4,579 KB 3,978 KB 14,277 KB 13,785 KB
Blog index 77 to 48 6,062 KB 854 KB 7,106 KB 1,915 KB

The resource page figure is dominated by the embedded Google Drive document viewer, which contributes roughly 9 MB of Google's own JavaScript and CSS. That embed is present on both sites, so the comparison holds it constant rather than claiming it as an improvement.

A CDN change during testing

Testing began while the site was served by one CDN and finished after a migration to Pantheon's Advanced Global CDN. The two configurations were briefly serving the same build at the same time, which produced an unusually direct comparison.

Home page, same build, two CDN paths, minutes apart
Path Content-Encoding Bytes
Advanced Global CDN gzip 13,383
Previous CDN none 65,223

On the previous configuration no response reached the browser compressed, although the origin compressed correctly when asked. The same build served by a plain Node process returned 12,465 bytes with Content-Encoding: gzip. The effect was a home page transfer of 2,074 KB where roughly 1,010 KB was achievable. The CDN migration resolved this, and every figure in this report comes from the current configuration.

Conditional requests also improved. A request carrying a matching If-None-Match header now returns 304 with an empty body, where it previously returned 200 with the full resource.

correct etag  ->  304, 0 bytes
no etag       ->  200, 229,287 bytes
bogus etag    ->  200, 229,287 bytes

Two items remain open. Static assets under /_next/static/ are served without a Cache-Control header, so they revalidate rather than being treated as immutable. Brotli is not negotiated, and a request advertising only Accept-Encoding: br receives an uncompressed response rather than a gzip fallback. Brotli would reduce these assets by a further 16% against gzip at maximum quality.

SEO and Accessibility

Lighthouse SEO reached 100 on all four pages. The old site failed the same two audits sitewide, for a missing meta description and for links that could not be crawled.

Page SEO before SEO after Accessibility before Accessibility after
Home 85 100 82 92
Resource library 83 100 91 96
A resource page 85 100 88 97
Blog index 83 100 91 96

Why SEO was measured against a local production build

Pre-launch environments carry a hosting platform header instructing search engines not to index them, and Lighthouse reads that header alongside robots.txt. The score is therefore suppressed on any pre-launch host for reasons unrelated to the site. SEO and Best Practices were measured against a build configured identically to production, with the production origin set, which reproduces the indexable configuration exactly. Timing and transfer figures come from the deployed site.

Accessibility figures are Lighthouse automated checks. They cover a subset of WCAG success criteria and are a floor rather than a verdict. Manual testing is a separate exercise.

Limits of this measurement

  • Conditions were simulated from the United States and not measured in Haiti.
  • Lantern models a slow connection rather than applying real packet delay.
  • Every figure for the new site is a warm cache hit. The site scales to zero after a period without traffic, and the first request after that costs considerably more than these figures show.
  • Medians of three runs on the Haitian profile, single runs elsewhere. Lighthouse results vary between runs even under simulated throttling.
  • Best Practices scores on the deployed site are depressed by the CDN's bot detection challenging the headless test browser. The figures reported here come from the local production build, where no challenge is issued.
Designed and built by

Design and engineering of great websites, including the rebuild measured in this report.

fastforward.sh
Hosted on

mit-ayiti.net runs on the Pantheon platform which provides the build pipeline and the CDN behind the figures above as well as Pantheon P1, a next generation content editing platform.

pantheon.io
Measurements taken 7 November 2026 with Google Lighthouse 12.8.2. A shorter summary is available in the overview report.