Byte attribution of a served product snapshot: inlined CSS is 13%, attributes are 75%

首先阅读 postProcess 路径以及现有的 pruneUnmatchedCss 和 minifyInlineCss 行为，然后使用所述的 browser v1.21.0 选项复现 /prerender_admin/page-content 的字节分解。将 SVG、类 token 和重复 Style-URL 这些杠杆与外部样式表以及 paint 或几何一致性进行比较；不要移除 review widget CSS。完成的要求是：选定一个经过测量的优化方案，在渲染内容不变的情况下，为达到 1 MB 上限提供清晰的路径。

Summary

A served product snapshot was reported as having a low text-to-HTML ratio, attributed to the CSS flattenShadowDom inlines for the review widget. Measured against a real served page, that attribution is wrong: inlined CSS is 13.4% of the document and attributes are 74.7%. Deleting every byte of injected styling moves the ratio 2.66% → 3.07%.

This issue records the byte attribution so the numbers aren't re-derived, and separates the levers that are provably safe from the one that isn't.

Method

One review-heavy product page (desktop), read from the serving cache via /prerender_admin/page-content, so these are the bytes a crawler receives. Rendered by browser v1.21.0 with flattenShadowDom, pruneUnmatchedCss and minifyInlineCss all on — i.e. the CSS below is already the residue after ~74% of style rules were pruned and the rest CSSOM-minified.

Class-token reference counts were computed against the inline <style> blocks and the four external stylesheets the snapshot still links. Measuring against inline CSS alone overstates the unreferenced total by ~6× (359 KB vs the true 61 KB) — the page keeps a 588 KB external utility stylesheet, so most tokens are referenced from a sheet that is not in the document.

Where 1,468,123 bytes go

	bytes	% of doc
class attributes	451,713	30.8%
inlined <style> CSS	197,160	13.4%
style= attributes	140,957	9.6%
href / src / srcset	214,751	14.6%
SVG d path data	81,344	5.5%
aria-label	52,604	3.6%
tag names + brackets	55,700	3.8%
ld+json	17,486	1.2%
visible text	39,078	2.66%

10,137 elements. 74.7% of the document is HTML attributes.

The review widget

The flattened review container is 385,482 B — 26.3% of the page — carrying 7,707 B of review text (2.0% of itself). Eight review cards; one card is 22,446 B / 235 elements for 956 B of text.

inside the review container	bytes	% of block
inlined <style> (host sheet + nested shadow hosts)	152,624	39.6%
SVG markup (d alone = 62,473)	108,954	28.3%
class	60,930	15.8%
review text	7,707	2.0%

Two candidates measured and rejected

• Deduplicating identical shadow stylesheets across hosts. 9 distinct sheets across 10 hosts; total bytes sitting in duplicate copies: 196 B. Not a lever.
• minifyInlineCss on the flattened blocks. Already CSSOM output, so re-emission is a no-op — consistent with the existing docstring.

Levers that remain, with sizes

1. Repeated SVG path data — ~64,093 B (4.4% of the document), concentrated in the review block. 193 d occurrences, only 61 distinct; a single 1,033-byte path appears 45 times (46,485 B). Hoisting exact-duplicate path data into <defs><symbol> + <use> renders identically. No such pass exists in postProcess today.
2. Class tokens no surviving rule references — 61,277 B page-wide, 34,467 B inside the review block. Same soundness argument that already licenses pruneUnmatchedCss: with stripScripts on, nothing can add a class after serialization, so a token no sheet mentions is inert. Must be evaluated against external sheets, not just inline ones.
3. A repeated background-image URL in style= attributes — ~93,000 B. One sprite URL repeated 489 times across style= attributes. Origin-authored, not introduced by post-processing.

Cloaking: what actually carries risk here

Worth stating explicitly, because "fidelity" has been used to decline byte-cutting levers that do not actually implicate it.

Cloaking is presenting different content to users and crawlers in order to manipulate rankings. The unit is content — text, links, images, prices, availability, structured data — not markup. Removing an attribute no rule reads, deduplicating identical SVG path data, or pruning a rule that matches nothing changes zero content: a user and a crawler still perceive the same page. Levers 1–3 are outside the policy, and rendering-fidelity checks (paintParity, geometry) are the right bar for them for a different reason — Google indexes the rendered DOM and devalues content it determines is not visible, so a layout regression costs indexing quality on its own merits.

The risk is asymmetric, and it runs the other way: revealing content to a crawler that a user does not see is the cloaking-shaped move. That is exactly what dropping the review widget's CSS would do — the widget's own rules hold collapsed panels, modals and secondary UI at display:none, and stripping them injects text into the indexable page that no user is shown. It would also reintroduce the inbound-encapsulation defects fixed across v1.19–v1.21. Do not pursue "just drop the review CSS" for byte savings.

Note for answer/generative engines

Structured data carries aggregateRating (ratingValue, ratingCount) but no individual review objects. So the 385 KB review DOM is the only place individual review content exists in the snapshot — it is load-bearing for extraction, not redundant with the JSON-LD, and should not be treated as trimmable content. Emitting review items into the JSON-LD would serve extraction far better per byte than the DOM does, but that is an origin change, not a post-processing one.

Separately: text-to-HTML ratio is not a documented ranking signal — it is a third-party audit-tool metric. The documented hard constraint is the 1 MB soft cap a crawler applies to a document, and this page is at 1.47 MB. Sizing work against "get under 1 MB" is better defined than "raise a ratio", and levers 1–3 together are ~218 KB of the ~470 KB needed.