Do soft navigations repay the slow first load?

The bargain, and the measurement gap that hid it

For years the SPA-versus-MPA argument was fought with one eye closed. The initial document load was visible to every tool; route transitions inside a long-lived JavaScript app were not measured the same way. That asymmetry let both sides be right about different halves of the same session — an advocate could point to snappy later navigations, a critic to a heavy first load with hydration and interaction cost, and neither could settle it. The WICG soft-navigation spec states the gap plainly: existing performance APIs had no way to separate SPA route changes from the initial navigation, and LCP, FCP and CLS were tied to the page lifecycle rather than to later in-app navigations.

Chrome's final soft-navigation origin trial — running Chrome 147 through 149 — closes it. It emits a SoftNavigationEntry when a route change satisfies the soft-navigation criteria (a user interaction, a URL change and a visible paint), and adds a navigationId to largest-contentful-paint, interaction-contentful-paint, event-timing and layout-shift entries so paints, shifts and interactions can be attributed to the specific route that produced them. That is exactly the instrumentation an honest amortization test needs.

The real question, then, is not “SPA or MPA?” in the abstract. It is an amortization question: does the cumulative saving from soft navigations exceed the first-load penalty, within the real distribution of session depths?

The break-even model

Let P be the SPA first-load penalty versus the matched MPA; Δ the average saving of a soft navigation versus the equivalent MPA navigation; and K the number of follow-on navigations in the session. The bargain pays off only when the savings repay the penalty:

P  ≤  Σ Δ_i                 the penalty is repaid by the follow-on savings
K  ≥  P / Δ                 follow-on navigations needed (constant Δ)
D* =  1 + P / Δ             break-even session depth (load + follow-ons)

Δ_required = P / r          required per-nav saving, using r ≈ E[follow-on navs]

The last line is the whole argument. If a session contains on average r follow-on navigations, then to repay a penalty P the average soft navigation must save at least P / r. Because r is slightly below one on the public web, the required saving is slightly above the entire penalty — every navigation has to give back more than the first load cost. This turns a vague promise (“routes feel snappier”) into a falsifiable test any team can apply before choosing a framework.

Reading the ledger

Think of a session as a ledger that opens in the red by the first-load penalty and is paid down by each follow-on saving. Whether it ever reaches break-even depends on two things the team controls only partly: how big each saving is, and how many navigations the user actually makes.

That picture is the segment story in one image. Where K is small, the bargain needs an implausibly large per-navigation win. Where K is large — a tool people live in for an hour — even modest per-route savings compound into a decisive lead. So the decision is not architectural taste; it is whether your segment supplies enough follow-on navigations to clear your own penalty:

The MPA baseline moved — the test has to move with it

Even a real soft-navigation saving is measured against a counterfactual, and the honest counterfactual is no longer “an old-fashioned full reload.” Modern multi-page apps get platform help that makes document navigations fast or instant:

• Back/forward cache (bfcache) serves instant back/forward navigations from a frozen page snapshot.
• Speculation rules prefetch or prerender likely next documents, so a click can paint immediately.
• Cross-document view transitions give MPAs the animated, app-like route change that used to be an SPA-only flourish.
• The Navigation API (now Baseline) gives MPAs and SPAs alike a clean, interceptable navigation model.

A fair benchmark therefore can't compare an SPA only to an unoptimised MPA; it has to include a modern, enhanced MPA arm. Against that baseline the soft-navigation saving Δ shrinks — and the break-even gets harder still, because a prerendered MPA navigation can itself be near-instant.

Threats to validity

• Benchmark unfairness. SPA and MPA aren't single technologies — they're bundles of caching, rendering, bundling and third-party choices. The two-arm design (real matched pairs + controlled paired builds) is the defence.
• Metric mismatch. A fast route data fetch can hide a late paint. Route-level paint and interaction metrics are primary; soft-nav TTFB is secondary.
• Session-depth aggregation. The RUM Archive ratios are aggregated, anonymised beacon shares — an event-weighted proxy, not a user-level session-depth histogram. A production study should re-estimate K from first-party RUM by segment. We use the ratio only as a published proxy, and say so.
• An unfair MPA baseline. Because bfcache, speculation rules, prerendering and view transitions exist, the benchmark must include both baseline and enhanced-MPA conditions or it flatters the SPA.

Limitations — read before quoting

1. The speed saving is not yet measured

   This article derives the break-even from published session ratios; it does not claim a measured Δ. The benchmark above exists precisely to supply it. Treat the result as “how big a saving the bargain requires,” not “how big the saving is.”

2. Aggregated ratios, not your sessions

   The RUM Archive shares are device-aggregated beacon proportions. Your segment may be deeper (a SaaS editor) or shallower (a landing page). The model is a calculator; K is the input you must supply.

3. Chart values are image-derived

   The exact 27.6 / 27.2 / 26.4 / 23.1 shares are read from Russell's published charts; the headline ~1:1 ratio is stated in his text. We carry the precise figures as source-derived proxies, not as our own measurement.

4. A constant-Δ simplification

   The closed form assumes a stable average saving per navigation. Real savings vary by route; the full sensitivity grid in the pack relaxes this across a range of P and Δ.