thinkrank / includes / seo / class-builder-content.php

class-builder-content.php in ThinkRank AI SEO – AI SEO Plugin for WordPress: Schema, XML Sitemaps, Meta Tags, Search Console & Local SEO 2.12.0, at includes/seo/class-builder-content.php

1,849 lines 72.1 KB

Raw Download Zip

1	<?php
2	/**
3	* Page-builder content extraction.
4	*
5	* SEO analysis reads `post_content`, which is only the real text on a classic
6	* post. Page builders keep the words somewhere else, and every server-side
7	* scoring path — bulk analysis, the post-list SEO Overview column, the MCP
8	* abilities, cron reports — saw an empty page as a result:
9	*
10	* - Oxygen / Breakdance leave `post_content` completely EMPTY and store the
11	* node tree in postmeta. Nothing to render, nothing to strip: the analyzer
12	* reported "No content" on pages with well over a thousand visible words.
13	* - Elementor does the same via `_elementor_data`.
14	* - Divi 5 and Gutenberg do store block markup in `post_content`, but Divi
15	* keeps module text inside the block's JSON attributes — inside an HTML
16	* comment, which tag stripping removes wholesale.
17	* - Divi 4 and other shortcode builders keep text in shortcode attributes.
18	*
19	* Extraction reads the builder's own stored data rather than invoking its
20	* render engine. Rendering an Oxygen page outside a front-end request is slow,
21	* stateful and can fatal in an admin context, whereas the stored tree is just
22	* JSON — cheap, side-effect free and safe to touch during a bulk run.
23	*
24	* @package ThinkRank\SEO
25	* @since 1.23.0
26	*/
27
28	declare(strict_types=1);
29
30	namespace ThinkRank\SEO;
31
32	if (!defined('ABSPATH')) {
33	exit;
34	}
35
36	/**
37	* Resolves the analyzable content of a post, whatever built it.
38	*/
39	class Builder_Content {
40
41	/**
42	* Post meta keys that hold builder data, in priority order.
43	*
44	* Several generations of the same builder are listed on purpose: Oxygen 6
45	* is Breakdance under the hood (`_breakdance_data`), while earlier Oxygen
46	* releases used `_oxygen_data` or the shortcode-based
47	* `ct_builder_shortcodes`. A site can only have one of them.
48	*
49	* @var string[]
50	*/
51	private const BUILDER_META_KEYS = [
52	'_breakdance_data', // Oxygen 6+ / Breakdance
53	'_oxygen_data', // Oxygen (earlier releases)
54	// Oxygen classic. 4.x writes the tree as JSON to `ct_builder_json`
55	// while still keeping `ct_builder_shortcodes`. A post carrying only
56	// the JSON key used to match no key at all and fall through to an
57	// empty `post_content`, which reads as a one-word page (#776).
58	//
59	// Oxygen 4.8.3 then renamed every `ct_` post meta key to `_ct_`
60	// (`oxygen_vsb_update_4_8_3()` runs `oxy_prefix_meta_keys()` on
61	// upgrade, and `oxy_get_post_meta()` only reads the prefixed name
62	// from then on). A current Oxygen classic site therefore has only the
63	// underscored keys, which nothing here listed, so every one of its
64	// pages resolved as empty. The prefixed keys come first because they
65	// are what Oxygen itself reads; the bare ones cover a site that has
66	// not run the migration (or reverted it with `?unprefix_meta`).
67	//
68	// These four are not read in this loop: from_oxygen_classic() pairs
69	// each JSON key with its shortcode sibling so the two forms can be
70	// compared. They are listed here because this list is also what the
71	// word-count index watches and what FAQ detection scans.
72	'_ct_builder_json', // Oxygen classic 4.8.3+ (JSON tree)
73	'ct_builder_json', // Oxygen classic 4.0-4.8.2 (JSON tree)
74	'_ct_builder_shortcodes', // Oxygen classic 4.8.3+ (shortcode tree)
75	'ct_builder_shortcodes', // Oxygen classic < 4.8.3 (shortcode tree)
76	'_elementor_data', // Elementor
77	// Beaver Builder. Published layout first: `_fl_builder_draft` holds
78	// unsaved changes and would score content the visitor cannot see.
79	// Both are arrays of stdClass nodes, which is why the walker below
80	// has to treat objects like arrays (#449).
81	'_fl_builder_data', // Beaver Builder (published)
82	'_fl_builder_draft', // Beaver Builder (unsaved changes)
83	];
84
85	/**
86	* Bricks' content-area meta key, used when Bricks itself isn't loaded.
87	*
88	* Bricks exposes `BRICKS_DB_PAGE_CONTENT` and renames the underlying key
89	* between generations (it gained the `_2` suffix in 1.7.3), so the
90	* constant is authoritative and this literal is only the fallback for the
91	* contexts where it is undefined — Bricks is a theme, so on an admin or
92	* CLI request against a site that has since switched themes the constant
93	* is simply not there while the post meta still is.
94	*
95	* Bricks stores three areas — header, content and footer. Only the content
96	* area belongs to the post being scored; the header and footer areas live
97	* on Bricks' own template posts and would double-count site chrome into
98	* every page's word count, so they are deliberately not read here.
99	*
100	* @since 2.2.1
101	* @var string
102	*/
103	private const BRICKS_CONTENT_META_KEY = '_bricks_page_content_2';
104
105	/**
106	* Bricks' per-post editor-mode meta key, used when Bricks isn't loaded.
107	*
108	* @since 2.2.1
109	* @var string
110	*/
111	private const BRICKS_EDITOR_MODE_META_KEY = '_bricks_editor_mode';
112
113	/**
114	* Bricks' components option, used when Bricks itself isn't loaded.
115	*
116	* @since 2.2.1
117	* @var string
118	*/
119	private const BRICKS_COMPONENTS_OPTION = 'bricks_components';
120
121	/**
122	* Bricks' element that renders the post's own `post_content`.
123	*
124	* A Bricks page normally discards `post_content` entirely, which is why
125	* anything left there is invisible. Dropping this element onto the canvas
126	* is the one way an author puts it back on the page, so its presence flips
127	* `post_content` from stale leftovers to content the visitor reads.
128	*
129	* @since 2.3.1
130	* @var string
131	*/
132	private const BRICKS_POST_CONTENT_ELEMENT = 'post-content';
133
134	/**
135	* Resolved Bricks trees for this request, keyed by post ID.
136	*
137	* Rendering one page asks for the tree about twenty times — every
138	* description, every schema node, the FAQ guard — and resolving it is not
139	* free. `bricks_content_source()` clears `Bricks\Database::$active_templates`
140	* before asking Bricks which content template applies, which defeats
141	* Bricks' own early-return and re-runs its whole template-condition engine;
142	* `expand_bricks_components()` then walks the tree again. Measured on a
143	* Bricks page with no content of its own, that was ten full runs of the
144	* rules engine per request.
145	*
146	* Per-request only, and only ever read back within one page render — a
147	* request that writes Bricks content does not also render it.
148	*
149	* @since 2.3.1
150	* @var array<int,array<int,mixed>>
151	*/
152	private static array $bricks_trees = [];
153
154	/**
155	* JSON keys whose values are user-visible text.
156	*
157	* Builder trees mix content with configuration, so a blind string sweep
158	* would count CSS classes and option slugs as words. Matching on the key
159	* keeps the word count honest.
160	*
161	* @var string[]
162	*/
163	private const CONTENT_KEYS = [
164	'text', 'title', 'subtitle', 'heading', 'subheading', 'content',
165	'description', 'caption', 'excerpt', 'label', 'value', 'html',
166	'editor', 'quote', 'answer', 'question', 'body', 'button_text',
167	// Oxygen classic keeps an element's copy in `options.ct_content`
168	// (headline, text block, rich text, link and button labels). It is the
169	// field Oxygen's own serializer moves between the tags when it writes
170	// shortcodes (`parse_components_tree()`), and the one Relevanssi and
171	// Oxygen's WPML integration read. Missing from this list, the walker
172	// kept only copy that happened to contain markup: a page of plain
173	// headings and paragraphs lost almost all of its words.
174	'ct_content',
175	// Oxygen's composite elements keep their copy under `options.original`
176	// instead, one key per field. Taken from the list Oxygen itself treats
177	// as text when it serializes (`$options_to_encode`); the numeric price
178	// fields and the progress bar's right-hand percentage are left out.
179	'testimonial_text', 'testimonial_author', 'testimonial_author_info',
180	'icon_box_heading', 'icon_box_text',
181	'pricing_box_package_title', 'pricing_box_package_subtitle', 'pricing_box_content',
182	'progress_bar_left_text',
183	];
184
185	/**
186	* Oxygen classic's storage generations, as JSON key => shortcode key.
187	*
188	* @since 2.10.0
189	* @var array<string,string>
190	*/
191	private const OXYGEN_CLASSIC_KEYS = [
192	'_ct_builder_json' => '_ct_builder_shortcodes',
193	'ct_builder_json' => 'ct_builder_shortcodes',
194	];
195
196	/**
197	* JSON keys whose values hold a link destination.
198	*
199	* Builders store a link's destination in a structured field separate from
200	* its label, either as a bare URL string or as a `{ url: … }` object.
201	* Neither shape survives a text sweep — the key is not content and a bare
202	* URL contains no `<` — so no `<a>` tag reached the link counters.
203	*
204	* @var string[]
205	*/
206	private const URL_KEYS = [
207	'link', 'url', 'href', 'link_url', 'button_link', 'permalink', 'link_to',
208	];
209
210	/**
211	* JSON keys whose values hold an embedded video's source.
212	*
213	* A builder's video widget keeps its destination in a provider-specific
214	* field — Elementor picks `youtube_url`, `vimeo_url`, `dailymotion_url` or
215	* `hosted_url` according to the chosen source type — none of which is a
216	* link field or a content field, so a video on a builder page reached the
217	* analyzers as nothing at all.
218	*
219	* These are deliberately kept out of URL_KEYS. A video is an embed, not an
220	* outbound link: rendering one as `<a href>` would add a spurious external
221	* link to every page carrying a video and skew the link counts. They are
222	* reconstructed as `<iframe>`/`<video>` instead, which the video detector
223	* recognises and the link and image counters ignore.
224	*
225	* @since 2.3.1
226	* @var string[]
227	*/
228	private const VIDEO_KEYS = [
229	'youtube_url', 'vimeo_url', 'dailymotion_url', 'videopress_url',
230	'hosted_url', 'video_url', 'video_src', 'video_link',
231	];
232
233	/**
234	* File extensions that mean a video source is a file, not a provider page.
235	*
236	* @since 2.3.1
237	* @var string[]
238	*/
239	private const VIDEO_FILE_EXTENSIONS = ['mp4', 'webm', 'ogv', 'mov', 'm4v'];
240
241	/**
242	* Source keys to trust for a declared video source type.
243	*
244	* A widget keeps one field per provider and does not clear the others when
245	* the author switches source: an Elementor video moved from YouTube to Self
246	* Hosted still carries the earlier `youtube_url`. Reading whichever key
247	* turns up first then emits the video the author replaced. The widget says
248	* which one it is actually playing, so that is read first and the flat key
249	* sweep is only the fallback for a builder that declares nothing.
250	*
251	* @since 2.3.1
252	* @var array<string,string[]>
253	*/
254	private const VIDEO_KEYS_BY_TYPE = [
255	'youtube' => ['youtube_url'],
256	'vimeo' => ['vimeo_url'],
257	'dailymotion' => ['dailymotion_url'],
258	'videopress' => ['videopress_url'],
259	'hosted' => ['hosted_url', 'video_url', 'video_src', 'video_link'],
260	'media' => ['hosted_url', 'video_url', 'video_src', 'video_link'],
261	'file' => ['hosted_url', 'video_url', 'video_src', 'video_link'],
262	'self_hosted' => ['hosted_url', 'video_url', 'video_src', 'video_link'],
263	];
264
265	/**
266	* Keys a builder uses to name which video source a widget is playing.
267	*
268	* @since 2.3.1
269	* @var string[]
270	*/
271	private const VIDEO_TYPE_KEYS = ['video_type', 'videotype', 'video_source', 'source_type'];
272
273	/**
274	* JSON keys whose values hold an image, as a URL string or `{ url, alt }`.
275	*
276	* @var string[]
277	*/
278	private const IMAGE_KEYS = [
279	'image', 'src', 'image_url', 'background_image', 'bg_image', 'photo',
280	];
281
282	/**
283	* JSON keys that carry a heading level for the node's text.
284	*
285	* A builder heading's text is collected (its key is in CONTENT_KEYS) and so
286	* counts toward the word count, but it arrives as bare text with no `<h2>`
287	* wrapper — which is why heading-structure checks saw none.
288	*
289	* @var string[]
290	*/
291	private const HEADING_TAG_KEYS = [
292	// Lower-cased on both sides of the comparison, so `headingtag` is the
293	// camelCase `headingTag` Bricks uses throughout its own controls and
294	// which ThinkRank's Bricks elements declare. Without it their section
295	// headings counted as body copy and never reached heading structure.
296	'header_size', 'heading_tag', 'headingtag', 'html_tag', 'title_tag', 'tag', 'level', 'size',
297	];
298
299	/**
300	* Keys whose value is alternative text for a sibling image.
301	*
302	* @var string[]
303	*/
304	private const ALT_KEYS = ['alt', 'alt_text', 'image_alt', 'title'];
305
306	/**
307	* The global post and every global `setup_postdata()` writes.
308	*
309	* Rendering points them at the post being analyzed, then puts each one
310	* back exactly as it was, unset included (#860).
311	*
312	* @since 2.12.0
313	* @var string[]
314	*/
315	private const POSTDATA_GLOBALS = [
316	'post', 'id', 'authordata', 'currentday', 'currentmonth',
317	'page', 'pages', 'multipage', 'more', 'numpages',
318	];
319
320	/**
321	* Resolve the content worth analyzing for a post.
322	*
323	* @param \WP_Post $post Post being analyzed.
324	* @return string HTML/text to analyze.
325	*/
326	public static function resolve(\WP_Post $post): string {
327	$raw = (string) $post->post_content;
328
329	// A page built in Gutenberg and then switched to Bricks keeps its old
330	// blocks in `post_content` forever — Bricks never clears them, and
331	// never renders them either. Resolving that first meant the stale draft
332	// beat the tree the visitor actually reads, and it did not stop at the
333	// score: the same string becomes the meta description, og:description,
334	// twitter:description and the schema description. Starting from nothing
335	// sends the resolution straight to Bricks' storage, which is where this
336	// page's words are (#651).
337	//
338	// Only for the stored path. `resolve_markup()` is also called with live
339	// editor content, and the Bricks panel's resolver reads the canvas —
340	// discarding that would replace what the author is typing with the last
341	// save.
342	if (self::bricks_supersedes_post_content((int) $post->ID)) {
343	$raw = '';
344	}
345
346	return self::resolve_markup($raw, $post);
347	}
348
349	/**
350	* Whether Bricks renders this post and throws its `post_content` away.
351	*
352	* True means anything still stored in `post_content` is invisible: it is
353	* not on the page, so it must not be scored, described or published as
354	* structured data. False covers both a post Bricks does not own and a
355	* Bricks page that puts `post_content` back with a Post Content element.
356	*
357	* @since 2.3.1
358	*
359	* @param int $post_id Post being resolved.
360	* @return bool
361	*/
362	public static function bricks_supersedes_post_content(int $post_id): bool {
363	$tree = self::bricks_tree($post_id);
364
365	if (empty($tree)) {
366	return false;
367	}
368
369	foreach ($tree as $element) {
370	if (is_array($element)
371	&& self::BRICKS_POST_CONTENT_ELEMENT === ($element['name'] ?? null)
372	) {
373	return false;
374	}
375	}
376
377	return !self::bricks_tree_prints_post_content($tree);
378	}
379
380	/**
381	* Whether a Bricks tree prints the body through a dynamic-data tag.
382	*
383	* The Post Content element is not the only way back onto the page: Bricks'
384	* `{post_content}` tag renders the same thing from inside an ordinary text
385	* element, and a single-post template written that way is a common shape.
386	* Missing it would mean the post's real body is discarded everywhere —
387	* scoring, the meta/og/twitter descriptions, the schema description — for a
388	* page that is displaying it.
389	*
390	* Matched over the encoded tree rather than per setting, because the tag can
391	* sit in any string field of any element and Bricks allows modifiers after
392	* the name (`{post_content:...}`).
393	*
394	* @since 2.3.1
395	*
396	* @param array $tree Bricks element tree.
397	* @return bool
398	*/
399	private static function bricks_tree_prints_post_content(array $tree): bool {
400	$encoded = wp_json_encode($tree);
401
402	return is_string($encoded) && false !== stripos($encoded, '{post_content');
403	}
404
405	/**
406	* The post's content as the visitor actually receives it.
407	*
408	* `post_content` for everything except a Bricks page that discards it, and
409	* there the Bricks tree's text. Descriptions are derived from a post's body
410	* in half a dozen places; every one of them wants this rather than the raw
411	* column (#651).
412	*
413	* @since 2.3.1
414	*
415	* @param \WP_Post $post Post being described.
416	* @return string
417	*/
418	public static function visible_content(\WP_Post $post): string {
419	$superseding = self::superseding_content($post);
420
421	return '' !== $superseding ? $superseding : (string) $post->post_content;
422	}
423
424	/**
425	* Replacement body text for a post whose `post_content` does not render.
426	*
427	* Empty for every ordinary post, which is what makes this safe to call from
428	* paths that already handle excerpts their own way: they keep that handling
429	* and only a Bricks page is diverted.
430	*
431	* @since 2.3.1
432	*
433	* @param \WP_Post $post Post being described.
434	* @return string Visible body text, or '' when `post_content` is fine.
435	*/
436	public static function superseding_content(\WP_Post $post): string {
437	if (!self::bricks_supersedes_post_content((int) $post->ID)) {
438	return '';
439	}
440
441	$bricks = self::from_bricks((int) $post->ID);
442
443	return self::is_blank($bricks) ? '' : $bricks;
444	}
445
446	/**
447	* Body text to derive a description from, when the usual source is wrong.
448	*
449	* A hand-written excerpt is the author's own summary and is correct however
450	* the page is built, so it yields '' here and the caller's normal
451	* `get_the_excerpt()` path keeps it. Only a Bricks page with no excerpt —
452	* where core would derive one from discarded `post_content` — gets diverted.
453	*
454	* @since 2.3.1
455	*
456	* @param \WP_Post $post Post being described.
457	* @return string Text to summarize, or '' to leave the caller's path alone.
458	*/
459	public static function superseding_excerpt_source(\WP_Post $post): string {
460	if ('' !== trim((string) $post->post_excerpt)) {
461	return '';
462	}
463
464	return self::superseding_content($post);
465	}
466
467	/**
468	* The Bricks element tree that renders for a post.
469	*
470	* Public because what Bricks puts on the page is not only a scoring
471	* question: the schema graph has to know whether a Bricks element already
472	* publishes the page's FAQ before adding one of its own (#649, #650).
473	*
474	* Flat, in Bricks' own storage shape — `expand_bricks_components()`
475	* appends component definitions to the same list rather than nesting them,
476	* so one `foreach` reaches every element.
477	*
478	* @since 2.3.1
479	*
480	* @param int $post_id Post being resolved.
481	* @return array<int,mixed> Elements, or [] when Bricks renders nothing here.
482	*/
483	/**
484	* The builder meta keys, for callers that need to inspect the raw storage
485	* rather than the text extracted from it.
486	*
487	* The SEO Analyzer reads these to answer "is there a ThinkRank FAQ element
488	* on this post?", which is a question about the stored tree, not about the
489	* words in it (#686).
490	*
491	* @since 2.7.0
492	* @return string[]
493	*/
494	public static function builder_meta_keys(): array {
495	return self::BUILDER_META_KEYS;
496	}
497
498	public static function bricks_tree(int $post_id): array {
499	if (array_key_exists($post_id, self::$bricks_trees)) {
500	return self::$bricks_trees[$post_id];
501	}
502
503	self::$bricks_trees[$post_id] = self::resolve_bricks_tree($post_id);
504
505	return self::$bricks_trees[$post_id];
506	}
507
508	/**
509	* Discard the resolved-tree memo. Test seam.
510	*
511	* @since 2.3.1
512	* @return void
513	*/
514	public static function flush_bricks_cache(): void {
515	self::$bricks_trees = [];
516	}
517
518	/**
519	* Read and resolve a post's Bricks tree, ignoring the memo.
520	*
521	* @since 2.3.1
522	*
523	* @param int $post_id Post being resolved.
524	* @return array<int,mixed>
525	*/
526	private static function resolve_bricks_tree(int $post_id): array {
527	if (!self::bricks_owns_post($post_id)) {
528	return [];
529	}
530
531	$source = self::bricks_content_source($post_id);
532	if (!$source) {
533	return [];
534	}
535
536	$stored = get_post_meta($source, self::bricks_meta_key(), true);
537
538	if (is_string($stored)) {
539	$stored = '' === trim($stored) ? null : json_decode($stored, true);
540	}
541
542	if (!is_array($stored) \|\| empty($stored)) {
543	return [];
544	}
545
546	return self::expand_bricks_components($stored);
547	}
548
549	/**
550	* Resolve an arbitrary chunk of editor markup for the given post.
551	*
552	* The editor sends its live content to the scorer so an author sees their
553	* unsaved edits reflected. On a builder page that live string is the raw
554	* builder markup — the block editor hands over Divi's
555	* `<!-- wp:divi/... -->` comments verbatim, because it cannot render
556	* blocks it has no client-side registration for. Analyzed as-is it reads
557	* as zero words, which is how a Divi page could show a correct saved score
558	* while the live Content Analysis panel next to it still said
559	* "No content".
560	*
561	* Running the live string through the same chain as stored content keeps
562	* both paths honest, and falling through to the post's builder storage
563	* covers builders (Oxygen) whose editor content is empty to begin with.
564	*
565	* @since 1.23.0
566	*
567	* @param string $raw Markup to analyze.
568	* @param \WP_Post $post Post the markup belongs to.
569	* @return string Content to analyze.
570	*/
571	public static function resolve_markup(string $raw, \WP_Post $post): string {
572	$content = self::render_post_content($raw, $post);
573
574	// Block markup that renders to nothing usually means the builder that
575	// owns those blocks did not register them in this context — Divi 5
576	// loads its module library lazily per-request, so in CLI, REST, admin
577	// and block-editor requests do_blocks() yields an empty string while
578	// the words sit right there in the block attributes. Read them
579	// directly.
580	if (self::is_blank($content)) {
581	$from_blocks = self::from_block_attributes($raw);
582	if (!self::is_blank($from_blocks)) {
583	$content = $from_blocks;
584	}
585	}
586
587	// Only reach for builder storage when the markup yielded nothing — a
588	// classic post must never pay for this.
589	if (self::is_blank($content)) {
590	$builder = self::from_builder_meta((int) $post->ID);
591	if (!self::is_blank($builder)) {
592	$content = $builder;
593	}
594	}
595
596	// A resolution that collapsed to nothing is worse than the raw markup.
597	if (self::is_blank($content) && !self::is_blank($raw)) {
598	$content = $raw;
599	}
600
601	/**
602	* Filter the content ThinkRank analyzes for a post.
603	*
604	* Use this to teach ThinkRank about a builder it does not know, or to
605	* override extraction for one it does.
606	*
607	* @since 1.23.0
608	*
609	* @param string $content Resolved content.
610	* @param \WP_Post $post Post being analyzed.
611	* @param string $raw Markup this resolution started from.
612	*/
613	return (string) apply_filters('thinkrank_analyzable_content', $content, $post, $raw);
614	}
615
616	/**
617	* Render blocks and shortcodes found in post_content.
618	*
619	* Best-effort: a third-party block that fatals must not take the whole
620	* score down with it.
621	*
622	* Runs as the post's own context, as it would on the front end. Admin and
623	* REST requests have no current post, so a shortcode reading
624	* `get_the_ID()` got nothing, and one looping a related-posts query left
625	* the global post on the last of them: its `wp_reset_postdata()` goes back
626	* to the main query's post, and there is none. On the Classic Editor this
627	* runs after the form prints its hidden `post_ID` and before the title and
628	* editor, which then showed the related post, and Update saved it over the
629	* original (#860).
630	*
631	* @param string $raw Raw post content.
632	* @param \WP_Post $post Post the content belongs to.
633	* @return string Rendered content.
634	*/
635	private static function render_post_content(string $raw, \WP_Post $post): string {
636	if ('' === trim($raw)) {
637	return '';
638	}
639
640	$content = $raw;
641	$previous = self::snapshot_post_globals();
642
643	try {
644	// phpcs:ignore WordPress.WP.GlobalVariablesOverride.Prohibited -- Made current for the render, restored in finally.
645	$GLOBALS['post'] = $post;
646
647	// Fires `the_post`, which these paths never fired before: admin,
648	// REST and cron analysis had no current post at all. That is the
649	// same signal the front-end loop sends and it is what makes
650	// `get_the_ID()` work inside a shortcode, but it is a new call on a
651	// path that runs in bulk — the word-count index resolves every post
652	// it visits — so a theme that counts views on `the_post` will count
653	// them during indexing. Accepted deliberately: without it a
654	// shortcode cannot resolve its own post, which is the bug (#860).
655	if (function_exists('setup_postdata')) {
656	setup_postdata($post);
657	}
658
659	if (function_exists('has_blocks') && function_exists('do_blocks') && has_blocks($raw)) {
660	$content = do_blocks($raw);
661	}
662
663	// Block output can itself contain shortcodes, so this runs either way.
664	if (function_exists('do_shortcode') && strpos($content, '[') !== false) {
665	$content = do_shortcode($content);
666	}
667	} catch (\Throwable $e) {
668	return $raw;
669	} finally {
670	self::restore_post_globals($previous);
671	}
672
673	return self::is_blank($content) ? $raw : $content;
674	}
675
676	/**
677	* The post globals as they are now; a global that is unset has no key.
678	*
679	* @since 2.12.0
680	*
681	* @return array<string,mixed>
682	*/
683	private static function snapshot_post_globals(): array {
684	$snapshot = [];
685
686	foreach (self::POSTDATA_GLOBALS as $name) {
687	if (array_key_exists($name, $GLOBALS)) {
688	$snapshot[$name] = $GLOBALS[$name];
689	}
690	}
691
692	return $snapshot;
693	}
694
695	/**
696	* Put the post globals back as snapshot_post_globals() found them.
697	*
698	* Assigned directly rather than through `setup_postdata()`: there may have
699	* been no post to set up, and re-running it would fire `the_post` again.
700	*
701	* @since 2.12.0
702	*
703	* @param array<string,mixed> $snapshot From snapshot_post_globals().
704	* @return void
705	*/
706	private static function restore_post_globals(array $snapshot): void {
707	foreach (self::POSTDATA_GLOBALS as $name) {
708	if (array_key_exists($name, $snapshot)) {
709	// phpcs:ignore WordPress.NamingConventions.PrefixAllGlobals.NonPrefixedVariableFound -- Core's own globals, put back as they were.
710	$GLOBALS[$name] = $snapshot[$name];
711	} else {
712	unset($GLOBALS[$name]);
713	}
714	}
715	}
716
717	/**
718	* Extract text from the attributes of parsed blocks.
719	*
720	* @param string $raw Raw post content containing block markup.
721	* @return string Collected text, or '' when nothing was found.
722	*/
723	private static function from_block_attributes(string $raw): string {
724	if (!function_exists('parse_blocks') \|\| !function_exists('has_blocks') \|\| !has_blocks($raw)) {
725	return '';
726	}
727
728	try {
729	$blocks = parse_blocks($raw);
730	} catch (\Throwable $e) {
731	return '';
732	}
733
734	$attrs = [];
735	$collect = static function (array $items) use (&$collect, &$attrs): void {
736	foreach ($items as $block) {
737	if (!empty($block['attrs']) && is_array($block['attrs'])) {
738	$attrs[] = $block['attrs'];
739	}
740	if (!empty($block['innerBlocks']) && is_array($block['innerBlocks'])) {
741	$collect($block['innerBlocks']);
742	}
743	}
744	};
745	$collect($blocks);
746
747	return empty($attrs) ? '' : self::text_from_tree($attrs);
748	}
749
750	/**
751	* Everything Bricks contributes to this post's analyzable content.
752	*
753	* Bricks is the only builder here that needs more than a meta key, on
754	* three counts:
755	*
756	* - It leaves its stored tree behind when a post is switched back to the
757	* block editor, so an editor-mode gate has to run first or ThinkRank
758	* scores markup the visitor never sees — the same failure
759	* `_fl_builder_draft` was ordered against in #449.
760	* - A post's content can live on ANOTHER post. Bricks' Templates feature
761	* assigns a content template by condition, and a page using one stores
762	* nothing of its own; reading only the page's meta scores it blank
763	* while the visitor reads a full page.
764	* - Its stored text carries dynamic-data tags and internal element names
765	* that never reach the rendered page.
766	*
767	* @since 2.2.1
768	*
769	* @param int $post_id Post being resolved.
770	* @return string Extracted text, or '' when Bricks has nothing for it.
771	*/
772	private static function from_bricks(int $post_id): string {
773	$tree = self::bricks_tree($post_id);
774
775	if (empty($tree)) {
776	return '';
777	}
778
779	return self::strip_bricks_dynamic_tags(
780	self::text_from_tree(self::without_bricks_element_labels($tree))
781	);
782	}
783
784	/**
785	* Whether Bricks — not the block editor — renders this post.
786	*
787	* Bricks writes `bricks` or `wordpress` into its editor-mode meta as the
788	* author toggles between the two, and never clears the content it stored
789	* for the other mode. Only the `wordpress` value is disqualifying: an
790	* absent value is the normal state for a post Bricks built and never
791	* toggled. This follows Bricks' own `Helpers::render_with_bricks()`, which
792	* bails on exactly that one value.
793	*
794	* It deliberately does not match it exactly: the comparison here is
795	* case-insensitive, where Bricks' is strict. Bricks 2.3.12 only ever writes
796	* the value lowercase, so the two agree on everything Bricks itself
797	* stores; they part company only on a value some other integration wrote.
798	* The two shipping today disagree about the casing — SureRank compares
799	* against `'WordPress'`, AIOSEO against `'bricks'` — and of the two ways to
800	* be wrong about `'WordPress'`, blocking costs a score on a page that has
801	* one, while allowing scores stale content the visitor never sees, which is
802	* the failure this gate exists to prevent.
803	*
804	* @since 2.2.1
805	*
806	* @param int $post_id Post being resolved.
807	* @return bool
808	*/
809	private static function bricks_owns_post(int $post_id): bool {
810	$mode = get_post_meta($post_id, self::bricks_editor_mode_key(), true);
811
812	// phpcs:ignore WordPress.WP.CapitalPDangit.MisspelledInText -- Bricks' own stored meta value, lower-cased for the comparison.
813	return !(is_string($mode) && 'wordpress' === strtolower(trim($mode)));
814	}
815
816	/**
817	* The post whose Bricks tree actually renders for this post.
818	*
819	* Usually the post itself. When it stores nothing of its own, Bricks falls
820	* back to whichever content template's conditions match, and that template
821	* is a separate post carrying the words the visitor reads.
822	*
823	* Resolution is delegated to Bricks rather than reimplemented: template
824	* conditions are a whole rules engine (post IDs, types, taxonomies,
825	* archives), and a second implementation would drift from it. Bricks
826	* answers through statics, so they are saved and restored around the call —
827	* `set_active_templates()` returns early once populated, and on a
828	* front-end request Bricks has already populated it for the page being
829	* served. Clobbering that would corrupt the render in progress.
830	*
831	* Best-effort by design: any failure returns the post's own data, which is
832	* exactly today's behaviour.
833	*
834	* @since 2.2.1
835	*
836	* @param int $post_id Post being resolved.
837	* @return int Post ID holding the Bricks tree, or 0 when there is none.
838	*/
839	private static function bricks_content_source(int $post_id): int {
840	$own = get_post_meta($post_id, self::bricks_meta_key(), true);
841	if ((is_array($own) && !empty($own)) \|\| (is_string($own) && '' !== trim($own))) {
842	return $post_id;
843	}
844
845	if (!class_exists('\\Bricks\\Database')
846	\|\| !method_exists('\\Bricks\\Database', 'set_active_templates')
847	) {
848	return 0;
849	}
850
851	// `set_active_templates()` writes TWO statics — `$active_templates` and,
852	// when a header template resolves, `$header_position`. Both are saved,
853	// and both are restored in `finally` rather than on the happy path: a
854	// throw part-way through (a third-party hook on
855	// `bricks/database/content_type`, `bricks/builder/data_post_id` or
856	// `bricks/active_templates` is enough) must not leave Bricks' render
857	// state holding this lookup's values. Restoring only after a clean
858	// return is what the `catch` below would otherwise skip.
859	$has_header_position = property_exists('\\Bricks\\Database', 'header_position');
860	$saved_templates = \Bricks\Database::$active_templates;
861	$saved_header_position = $has_header_position ? \Bricks\Database::$header_position : null;
862
863	try {
864	\Bricks\Database::$active_templates = [];
865	\Bricks\Database::set_active_templates($post_id);
866	$template = (int) (\Bricks\Database::$active_templates['content'] ?? 0);
867	} catch (\Throwable $e) {
868	return 0;
869	} finally {
870	\Bricks\Database::$active_templates = $saved_templates;
871	if ($has_header_position) {
872	\Bricks\Database::$header_position = $saved_header_position;
873	}
874	}
875
876	// A template that is the post itself adds nothing over the empty read
877	// above, and would otherwise recurse conceptually.
878	return $template === $post_id ? 0 : $template;
879	}
880
881	/**
882	* Splice component definitions into the tree.
883	*
884	* A Bricks component keeps its markup in the `bricks_components` option,
885	* not on the page. The page stores only an instance: an element carrying
886	* `cid` and, usually, empty `settings`. Walking the page alone therefore
887	* found no words at all, and a page built entirely from components scored
888	* blank — the same failure as a page built from a content template.
889	*
890	* Confirmed on Bricks 2.3.12: `Bricks\Frontend::render_data()` renders the
891	* component's copy from an instance this walker extracted '' from.
892	*
893	* The definition is read straight from the option rather than through
894	* `Bricks\Helpers::get_component_instance()`. That helper resolves an
895	* instance's property overrides, which would be better, but it reads
896	* `Bricks\Database::$global_data['components']` — populated once per
897	* request, and empty in the admin and CLI contexts where bulk scoring
898	* runs. Refreshing it would mean writing to Bricks' live render state, the
899	* same hazard the template resolver is careful to avoid, and gating on it
900	* would make a page score differently in wp-admin than on the front end.
901	* Reading the stored definition is consistent everywhere.
902	*
903	* The trade-off: an instance that overrides a component property is scored
904	* with the component's authored copy rather than the override. That is the
905	* text the component renders by default, and it is much closer than the
906	* nothing this returned before.
907	*
908	* @since 2.2.1
909	*
910	* @param array $tree Bricks content area.
911	* @return array Tree with component elements spliced in after each instance.
912	*/
913	private static function expand_bricks_components(array $tree): array {
914	$expanded = [];
915	$open = [];
916
917	$walk = static function (array $elements, int $depth) use (&$walk, &$expanded, &$open): void {
918	foreach ($elements as $element) {
919	$expanded[] = $element;
920
921	if (!is_array($element) \|\| empty($element['cid']) \|\| !is_string($element['cid'])) {
922	continue;
923	}
924
925	$cid = $element['cid'];
926
927	// A component nested inside its own definition would recurse
928	// forever; the depth cap covers deep but legitimate nesting.
929	if (isset($open[$cid]) \|\| $depth > 4) {
930	continue;
931	}
932
933	$children = self::bricks_component_elements($cid);
934	if (empty($children)) {
935	continue;
936	}
937
938	// Re-entrant per branch, not per page: the guard is released
939	// after the walk so a second instance further along the page
940	// still expands, rather than being mistaken for recursion.
941	//
942	// That does NOT double the word count — `text_from_tree()`
943	// ends in `array_unique()`, which collapses a repeated
944	// component's copy the same way it collapses a value repeated
945	// across responsive breakpoints. Expanding both instances is
946	// about not silently dropping the second one's structure.
947	$open[$cid] = true;
948	$walk($children, $depth + 1);
949	unset($open[$cid]);
950	}
951	};
952
953	$walk($tree, 0);
954
955	return $expanded;
956	}
957
958	/**
959	* The stored elements of one Bricks component.
960	*
961	* @since 2.2.1
962	*
963	* @param string $cid Component id held by an instance element.
964	* @return array Component elements, or [] when it cannot be resolved.
965	*/
966	private static function bricks_component_elements(string $cid): array {
967	$components = get_option(self::bricks_constant('BRICKS_DB_COMPONENTS', self::BRICKS_COMPONENTS_OPTION), []);
968
969	if (!is_array($components)) {
970	return [];
971	}
972
973	foreach ($components as $component) {
974	$component = self::as_children($component);
975	if (null === $component) {
976	continue;
977	}
978
979	if (isset($component['id']) && $component['id'] === $cid && !empty($component['elements'])) {
980	return is_array($component['elements']) ? $component['elements'] : [];
981	}
982	}
983
984	return [];
985	}
986
987	/**
988	* Drop each Bricks element's internal name before the tree is walked.
989	*
990	* A Bricks element carries an optional top-level `label` — the nickname an
991	* author types in the Structure panel to find it again ("Hero headline",
992	* "CTA row"). It is builder chrome and is never rendered, but `label` is in
993	* CONTENT_KEYS because it is real content for other builders' form fields,
994	* so it was being counted as page copy.
995	*
996	* Only the element's own `label` is removed. A `label` inside `settings`
997	* is a rendered field label and stays.
998	*
999	* @since 2.2.1
1000	*
1001	* @param array $tree Bricks content area.
1002	* @return array Tree with element nicknames removed.
1003	*/
1004	private static function without_bricks_element_labels(array $tree): array {
1005	foreach ($tree as $index => $element) {
1006	if (is_array($element) && isset($element['id'], $element['label'])) {
1007	unset($tree[$index]['label']);
1008	}
1009	}
1010
1011	return $tree;
1012	}
1013
1014	/**
1015	* Remove Bricks dynamic-data tags from extracted text.
1016	*
1017	* Bricks stores `{post_title}`, `{post_meta:price}`, `{echo:my_fn}` and the
1018	* like verbatim and resolves them when it renders. Extraction reads the
1019	* stored tree, so without this the placeholders were counted as words, and
1020	* a heading whose text is `{post_title}` reported the literal token as its
1021	* heading text.
1022	*
1023	* The pattern is deliberately narrower than Bricks' own
1024	* (`/{([\wÀ-ÖØ-öø-ÿ\-\s\.\/:...]+)}/u`), which also matches braces
1025	* containing spaces. Bricks only substitutes tags that resolve to a
1026	* registered provider and leaves anything else on the page as literal text,
1027	* so the broad pattern would delete prose the visitor can actually read.
1028	* Matching only tag-shaped tokens keeps every real sentence and still
1029	* removes every placeholder — the same trade-off SureRank makes.
1030	*
1031	* @since 2.2.1
1032	*
1033	* @param string $text Extracted text.
1034	* @return string Text with placeholders removed.
1035	*/
1036	private static function strip_bricks_dynamic_tags(string $text): string {
1037	$stripped = preg_replace('/\{[a-z0-9_][a-z0-9_:\-\.]*\}/i', '', $text);
1038
1039	if (null === $stripped) {
1040	return $text;
1041	}
1042
1043	// Collapse the runs of spaces a removed tag leaves mid-sentence,
1044	// without touching the newlines that separate collected nodes.
1045	$tidied = preg_replace('/[ \t]{2,}/', ' ', $stripped);
1046
1047	return null === $tidied ? $stripped : $tidied;
1048	}
1049
1050	/**
1051	* Bricks' content-area meta key, preferring Bricks' own constant.
1052	*
1053	* @since 2.2.1
1054	*
1055	* @return string
1056	*/
1057	private static function bricks_meta_key(): string {
1058	return self::bricks_constant('BRICKS_DB_PAGE_CONTENT', self::BRICKS_CONTENT_META_KEY);
1059	}
1060
1061	/**
1062	* Bricks' editor-mode meta key, preferring Bricks' own constant.
1063	*
1064	* @since 2.2.1
1065	*
1066	* @return string
1067	*/
1068	private static function bricks_editor_mode_key(): string {
1069	return self::bricks_constant('BRICKS_DB_EDITOR_MODE', self::BRICKS_EDITOR_MODE_META_KEY);
1070	}
1071
1072	/**
1073	* Read one of Bricks' key-name constants, falling back to the literal.
1074	*
1075	* @since 2.2.1
1076	*
1077	* @param string $name Constant name.
1078	* @param string $fallback Key to use when the constant is unavailable.
1079	* @return string
1080	*/
1081	private static function bricks_constant(string $name, string $fallback): string {
1082	if (defined($name)) {
1083	$value = constant($name);
1084	if (is_string($value) && '' !== trim($value)) {
1085	return $value;
1086	}
1087	}
1088
1089	return $fallback;
1090	}
1091
1092	/**
1093	* Pull text out of whichever builder stored this post.
1094	*
1095	* @param int $post_id Post ID.
1096	* @return string Extracted text, or '' when no builder data was found.
1097	*/
1098	private static function from_builder_meta(int $post_id): string {
1099	// Bricks first: it is the only builder whose content can live on
1100	// another post, and the only one gated on an editor mode.
1101	$bricks = self::from_bricks($post_id);
1102	if (!self::is_blank($bricks)) {
1103	return $bricks;
1104	}
1105
1106	foreach (self::BUILDER_META_KEYS as $key) {
1107	if (self::is_oxygen_classic_key($key)) {
1108	// Resolved as a pair, once, at the first of its keys.
1109	if ('_ct_builder_json' !== $key) {
1110	continue;
1111	}
1112
1113	$oxygen = self::from_oxygen_classic($post_id);
1114	if (!self::is_blank($oxygen)) {
1115	return $oxygen;
1116	}
1117	continue;
1118	}
1119
1120	$stored = get_post_meta($post_id, $key, true);
1121
1122	if (is_string($stored) && '' !== trim($stored)) {
1123	$decoded = json_decode($stored, true);
1124
1125	// JSON node tree (Breakdance/Oxygen 6, Elementor).
1126	if (is_array($decoded)) {
1127	$text = self::text_from_tree($decoded);
1128	if (!self::is_blank($text)) {
1129	return $text;
1130	}
1131	continue;
1132	}
1133
1134	continue;
1135	}
1136
1137	// Some builders store an already-decoded tree — an array for most,
1138	// an array of objects for Beaver Builder (#449).
1139	$tree = self::as_children($stored);
1140	if (null !== $tree) {
1141	$text = self::text_from_tree($tree);
1142	if (!self::is_blank($text)) {
1143	return $text;
1144	}
1145	}
1146	}
1147
1148	return '';
1149	}
1150
1151	/**
1152	* Whether a meta key is one of Oxygen classic's storage keys.
1153	*
1154	* @since 2.10.0
1155	*
1156	* @param string $key Meta key.
1157	* @return bool
1158	*/
1159	private static function is_oxygen_classic_key(string $key): bool {
1160	return isset(self::OXYGEN_CLASSIC_KEYS[$key]) \|\| in_array($key, self::OXYGEN_CLASSIC_KEYS, true);
1161	}
1162
1163	/**
1164	* Text of an Oxygen classic page, from whichever stored form holds more.
1165	*
1166	* Oxygen 4.x keeps the same tree twice: as JSON, and as the shortcodes it
1167	* used before 4.0. The JSON is preferred because it carries copy the
1168	* shortcode form hides (a composite element's text is base64-encoded
1169	* inside `ct_options`, which is configuration and stripped). It is not
1170	* trusted blindly, though. Reading `ct_builder_json` first once meant a
1171	* key missing from CONTENT_KEYS silently threw the page away while the
1172	* shortcode copy sat unread next to it, because a non-empty JSON result
1173	* stopped the search. Comparing the two means the next such gap costs
1174	* nothing: the richer form wins.
1175	*
1176	* A generation is only read as a pair. The prefixed keys are what Oxygen
1177	* 4.8.3+ reads, so an unprefixed leftover next to them is stale.
1178	*
1179	* @since 2.10.0
1180	*
1181	* @param int $post_id Post ID.
1182	* @return string Extracted text, or '' when Oxygen classic stored nothing.
1183	*/
1184	private static function from_oxygen_classic(int $post_id): string {
1185	foreach (self::OXYGEN_CLASSIC_KEYS as $json_key => $shortcode_key) {
1186	$json = get_post_meta($post_id, $json_key, true);
1187	$shortcodes = get_post_meta($post_id, $shortcode_key, true);
1188
1189	$from_json = '';
1190	if (is_string($json) && '' !== trim($json)) {
1191	$decoded = json_decode($json, true);
1192	if (is_array($decoded)) {
1193	// `[oxygen data="..."]` is a dynamic-data placeholder
1194	// Oxygen fills at render time. The shortcode path drops it
1195	// with every other tag, so it goes here too or the two
1196	// forms would disagree on the same page.
1197	$from_json = (string) preg_replace(
1198	'/\[oxygen\b[^\]]*\]/i',
1199	' ',
1200	self::text_from_tree($decoded)
1201	);
1202	}
1203	}
1204
1205	$from_shortcodes = '';
1206	if (is_string($shortcodes) && strpos($shortcodes, '[') !== false) {
1207	$from_shortcodes = self::text_from_shortcodes($shortcodes);
1208	}
1209
1210	if (self::is_blank($from_json) && self::is_blank($from_shortcodes)) {
1211	continue;
1212	}
1213
1214	return self::visible_word_count($from_json) >= self::visible_word_count($from_shortcodes)
1215	? $from_json
1216	: $from_shortcodes;
1217	}
1218
1219	return '';
1220	}
1221
1222	/**
1223	* Rough count of the words a visitor would read in extracted text.
1224	*
1225	* Only used to compare two extractions of the same page, so it needs to
1226	* be consistent rather than locale-exact.
1227	*
1228	* @since 2.10.0
1229	*
1230	* @param string $text Extracted text or markup.
1231	* @return int
1232	*/
1233	private static function visible_word_count(string $text): int {
1234	$plain = trim((string) preg_replace('/\s+/u', ' ', wp_strip_all_tags($text)));
1235
1236	return '' === $plain ? 0 : count(explode(' ', $plain));
1237	}
1238
1239	/**
1240	* Shortcode attributes that carry copy a visitor reads.
1241	*
1242	* An allow-list, not a deny-list. Oxygen Classic tags carry far more
1243	* attributes than they do copy — `id`, `class`, `selector`, `url`,
1244	* `ct_options` and friends — and a deny-list silently admits every
1245	* attribute a future builder release invents, which is how markup ends up
1246	* being counted as prose.
1247	*
1248	* @var string[]
1249	*/
1250	private const SHORTCODE_TEXT_ATTRIBUTES = [
1251	'text',
1252	'content',
1253	'heading',
1254	'title',
1255	'subtitle',
1256	'label',
1257	'caption',
1258	'description',
1259	'alt',
1260	'button_text',
1261	'link_text',
1262	];
1263
1264	/**
1265	* Extract readable text from a shortcode tree, without rendering it.
1266	*
1267	* Oxygen Classic is the only builder whose storage is shortcodes rather
1268	* than JSON, and the previous implementation handed the string to
1269	* `do_shortcode()`. That silently depends on Oxygen having registered its
1270	* `ct_*` handlers in the current request — which it has on a front-end
1271	* view, and has not during bulk analysis, the post-list column, cron or
1272	* REST/MCP. With no handlers registered `do_shortcode()` returns its input
1273	* unchanged, so the raw shortcode source was scored as if it were the
1274	* page's prose: `[ct_section`, `id="section-1"` and the rest counted toward
1275	* the word count, while the actual copy sitting in `text="..."` attributes
1276	* was never counted at all (#776).
1277	*
1278	* `strip_shortcodes()` is no help either — it also only knows registered
1279	* shortcodes, so it leaves the same text untouched.
1280	*
1281	* Reading the stored tree directly is what every other builder here already
1282	* does, and it matches the class's stated design: no render engine, no
1283	* dependency on load order, safe during a bulk run.
1284	*
1285	* Parsing unconditionally, rather than rendering when Oxygen happens to be
1286	* loaded and parsing otherwise, is deliberate. It makes the extracted text
1287	* the same in every context, so the score in the editor matches the score
1288	* from a bulk run or from MCP. The old code produced whichever of the two
1289	* the request happened to allow, which is why the same post could report
1290	* two different word counts depending on how it was asked.
1291	*
1292	* The trade-off is that rendered output (resolved images, links, anything
1293	* Oxygen pulls in from a reusable part) is no longer reflected here. For
1294	* what this text feeds — word count, content scoring, meta-description
1295	* fallbacks and schema text — that markup was never the point, and counting
1296	* it only when the builder happened to be booted was the bug.
1297	*
1298	* @since 2.10.0
1299	*
1300	* @param string $stored Raw shortcode source.
1301	* @return string Extracted text.
1302	*/
1303	private static function text_from_shortcodes(string $stored): string {
1304	// Oxygen stores each element's settings as a JSON blob in `ct_options`.
1305	// It is configuration, never copy, and it contains braces and brackets
1306	// that would otherwise confuse the tag scan below, so it goes first.
1307	//
1308	// The blob is matched as a balanced JSON object, not as "up to the
1309	// next quote". Oxygen wraps it in single quotes but does not escape
1310	// an apostrophe inside it (`"nicename":"Bob's Plumbing"`), so the
1311	// quote-to-quote match stopped mid-value and the rest of the blob,
1312	// `s Plumbing"}'` and all, was left in the tag and leaked into the
1313	// text. Strings inside the object are skipped whole, so neither a quote
1314	// nor a brace inside a value can end the match early.
1315	$source = (string) preg_replace(
1316	'/\sct_options\s=\s\'(?<obj>\{(?:[^{}"]++\|"(?:[^"\\\\]\|\\\\.)+"\|(?&obj))+\})\'/s',
1317	'',
1318	$stored
1319	);
1320
1321	// Anything not shaped like Oxygen's JSON blob keeps the old,
1322	// quote-delimited strip.
1323	$source = (string) preg_replace(
1324	'/\sct_options\s=\s(["\']).*?\1/s',
1325	'',
1326	$source
1327	);
1328
1329	$attributes = implode('\|', array_map(
1330	static fn(string $name): string => preg_quote($name, '/'),
1331	self::SHORTCODE_TEXT_ATTRIBUTES
1332	));
1333
1334	// Replace each shortcode tag with whatever readable copy its attributes
1335	// carry. Text between tags is left exactly where it is, so the result
1336	// keeps the page's reading order rather than hoisting all the headings
1337	// to the front.
1338	// The attribute blob is matched quote-aware rather than as "anything up
1339	// to the first `]`". Oxygen copy contains brackets often enough to
1340	// matter — "Best tools [2026]", "[Updated] our policy" — and a naive
1341	// scan ends the tag inside the `text` attribute, dropping the copy
1342	// before the bracket and leaking the stray `"]` after it into the
1343	// prose. Which is this bug's own failure mode: the wrong text scored.
1344	//
1345	// A tag name must start with a letter or underscore. `[2026]` is not a
1346	// shortcode anyone can register, and scanning it as one dropped the
1347	// year out of "Best tools [2026]".
1348	$text = (string) preg_replace_callback(
1349	'/\[\/?[a-zA-Z_][a-zA-Z0-9_-]((?:[^\]"\']\|"[^"]"\|\'[^\']\'))\]/',
1350	static function (array $matches) use ($attributes): string {
1351	if ('' === trim($matches[1])) {
1352	return ' ';
1353	}
1354
1355	if (!preg_match_all(
1356	'/\b(' . $attributes . ')\s=\s(["\'])(.*?)\2/s',
1357	$matches[1],
1358	$found,
1359	PREG_SET_ORDER
1360	)) {
1361	return ' ';
1362	}
1363
1364	$parts = [];
1365	foreach ($found as $attribute) {
1366	$value = trim($attribute[3]);
1367
1368	// An attribute holding markup or a JSON fragment is
1369	// configuration that happens to share a name with a copy
1370	// field, not something a visitor reads.
1371	if ('' === $value \|\| preg_match('/^[\[{<]/', $value)) {
1372	continue;
1373	}
1374
1375	$parts[] = $value;
1376	}
1377
1378	return empty($parts) ? ' ' : ' ' . implode(' ', $parts) . ' ';
1379	},
1380	$source
1381	);
1382
1383	// Oxygen escapes square brackets in an element's copy before writing
1384	// it between the tags, so that "Best tools [2026]" cannot be mistaken
1385	// for a shortcode (`oxygen_vsb_filter_shortcode_content_encode()`).
1386	// Decoded only now, after the tag scan, for the same reason; left
1387	// encoded, the placeholders were scored as words of their own.
1388	$text = str_replace(
1389	['_OXY_OPENING_BRACKET_', '_OXY_CLOSING_BRACKET_'],
1390	['[', ']'],
1391	$text
1392	);
1393
1394	// Entities are stored encoded in attributes (&, ’), and would
1395	// otherwise be counted as words.
1396	$text = html_entity_decode($text, ENT_QUOTES \| ENT_HTML5, 'UTF-8');
1397
1398	return trim((string) preg_replace('/\s+/u', ' ', $text));
1399	}
1400
1401	/**
1402	* A node's children, whether it stores them as an array or an object.
1403	*
1404	* The walker used to return immediately on `!is_array($node)`, so an
1405	* object node was dropped along with its entire subtree — silently, as
1406	* `''`, which the caller reads as "this builder stored nothing" rather
1407	* than "this walker cannot read this shape".
1408	*
1409	* Beaver Builder stores `_fl_builder_data` as an array of stdClass nodes,
1410	* each with a stdClass `settings` object, so every node would have been
1411	* dropped and adding its meta key alone would have looked like it worked
1412	* and changed nothing. Not BB-specific: any builder storing objects hits
1413	* this, and that shape will come up again (#449).
1414	*
1415	* @since 2.1.0
1416	*
1417	* @param mixed $node Candidate node.
1418	* @return array<string\|int,mixed>\|null Traversable children, or null.
1419	*/
1420	private static function as_children($node): ?array {
1421	if (is_array($node)) {
1422	return $node;
1423	}
1424
1425	// Deliberately not is_object(): a builder can store a value object
1426	// (DateTime, a WP_Post) whose properties are not content, and
1427	// get_object_vars() on those yields noise. stdClass is what the
1428	// JSON/serialize round-trip produces, which is the shape we want.
1429	if ($node instanceof \stdClass) {
1430	return get_object_vars($node);
1431	}
1432
1433	return null;
1434	}
1435
1436	/**
1437	* Walk a builder node tree and collect the user-visible text.
1438	*
1439	* Values are joined with block-level markup so downstream heading, link and
1440	* image detection keeps working on the result.
1441	*
1442	* @param array $tree Decoded builder tree.
1443	* @return string Collected HTML.
1444	*/
1445	private static function text_from_tree(array $tree): string {
1446	$collected = [];
1447
1448	// Strings already represented inside reconstructed markup, so the plain
1449	// sweep below doesn't emit a link label or heading a second time and
1450	// double it in the word count.
1451	$consumed = [];
1452
1453	// Pass 1 — rebuild <a>, <img> and <hN> from node shape. This has to
1454	// happen per node rather than per leaf: a link's label and its
1455	// destination are separate sibling fields, so once the tree is
1456	// flattened to leaves the pairing is gone.
1457	$reconstruct = static function ($node) use (&$reconstruct, &$collected, &$consumed): void {
1458	$node = self::as_children($node);
1459	if (null === $node) {
1460	return;
1461	}
1462
1463	$markup = self::markup_for_node($node, $consumed);
1464	if ('' !== $markup) {
1465	$collected[] = $markup;
1466	}
1467
1468	foreach ($node as $child_key => $child) {
1469	// A `link` / `image` sub-object is a destination descriptor the
1470	// parent has already folded into its markup. Descending into it
1471	// would emit the same URL a second time as a bare link, and
1472	// would turn an image's own `url` field into a spurious <a>.
1473	if (is_string($child_key)
1474	&& (in_array(strtolower($child_key), self::URL_KEYS, true)
1475	\|\| in_array(strtolower($child_key), self::IMAGE_KEYS, true)
1476	\|\| in_array(strtolower($child_key), self::VIDEO_KEYS, true))
1477	) {
1478	continue;
1479	}
1480
1481	$reconstruct($child);
1482	}
1483	};
1484	$reconstruct($tree);
1485
1486	// Pass 2 — remaining visible text.
1487	$walk = static function ($node, $key = null) use (&$walk, &$collected, &$consumed): void {
1488	$children = self::as_children($node);
1489	if (null !== $children) {
1490	foreach ($children as $child_key => $child) {
1491	$walk($child, is_string($child_key) ? $child_key : $key);
1492	}
1493	return;
1494	}
1495
1496	if (!is_string($node) \|\| '' === trim($node)) {
1497	return;
1498	}
1499
1500	// Already inside a reconstructed tag.
1501	if (in_array($node, $consumed, true)) {
1502	return;
1503	}
1504
1505	$is_content_key = is_string($key)
1506	&& in_array(strtolower($key), self::CONTENT_KEYS, true);
1507
1508	// Markup is content wherever it appears; bare strings only count
1509	// when their key says they are content, so slugs and class names
1510	// stay out of the word count.
1511	if ($is_content_key \|\| strpos($node, '<') !== false) {
1512	$collected[] = $node;
1513	}
1514	};
1515
1516	$walk($tree);
1517
1518	if (empty($collected)) {
1519	return '';
1520	}
1521
1522	// De-duplicate: builder trees often repeat a value across responsive
1523	// breakpoints, which would otherwise multiply the word count.
1524	$collected = array_unique($collected);
1525
1526	return implode("\n", $collected);
1527	}
1528
1529	/**
1530	* The video source a node is actually playing, if any.
1531	*
1532	* @since 2.3.1
1533	*
1534	* @param array $node Builder node.
1535	* @return string Video source, or '' when the node carries none.
1536	*/
1537	private static function video_from(array $node): string {
1538	foreach ($node as $key => $value) {
1539	if (!is_string($key) \|\| !is_string($value)) {
1540	continue;
1541	}
1542
1543	if (!in_array(strtolower($key), self::VIDEO_TYPE_KEYS, true)) {
1544	continue;
1545	}
1546
1547	$keys = self::VIDEO_KEYS_BY_TYPE[strtolower(trim($value))] ?? null;
1548	if (null === $keys) {
1549	continue;
1550	}
1551
1552	// A recognised video_type settles it, including when that
1553	// provider's own field is empty. Falling through to the flat sweep
1554	// there handed back whichever sibling key happened to come first in
1555	// node order — the stale youtube_url left behind after switching
1556	// the widget to a hosted file, which is exactly what keying on the
1557	// declared type is meant to prevent.
1558	$declared = self::url_from($node, $keys);
1559
1560	return self::is_video_source($declared) ? $declared : '';
1561	}
1562
1563	$url = self::url_from($node, self::VIDEO_KEYS);
1564
1565	return self::is_video_source($url) ? $url : '';
1566	}
1567
1568	/**
1569	* Whether a value can be a video source.
1570	*
1571	* `looks_like_url()` also accepts `#anchor`, `mailto:` and `tel:`, which a
1572	* link node may legitimately hold but a video cannot: `<iframe src="#top">`
1573	* is not a video and would reach a video sitemap as one.
1574	*
1575	* @since 2.3.1
1576	*
1577	* @param string $url Candidate source.
1578	* @return bool
1579	*/
1580	private static function is_video_source(string $url): bool {
1581	return '' !== $url
1582	&& (1 === preg_match('#^(https?:)?//#i', $url) \|\| str_starts_with($url, '/'));
1583	}
1584
1585	/**
1586	* Whether a video source points at a file rather than a provider page.
1587	*
1588	* @since 2.3.1
1589	*
1590	* @param string $url Video source.
1591	* @return bool
1592	*/
1593	private static function is_video_file(string $url): bool {
1594	$path = (string) wp_parse_url($url, PHP_URL_PATH);
1595	$ext = strtolower((string) pathinfo($path, PATHINFO_EXTENSION));
1596
1597	return in_array($ext, self::VIDEO_FILE_EXTENSIONS, true);
1598	}
1599
1600	/**
1601	* Rebuild the HTML a single builder node represents, if any.
1602	*
1603	* Looks only at the node's own fields (plus one level of nesting, because
1604	* builders commonly wrap a destination as `{ url: … }`). Returns an empty
1605	* string for the vast majority of nodes, which are layout or configuration.
1606	*
1607	* Any leaf string folded into the returned markup is appended to $consumed
1608	* so the plain-text sweep doesn't count it twice.
1609	*
1610	* @param array $node Builder node.
1611	* @param array $consumed Collects strings represented in the returned markup.
1612	* @return string Reconstructed HTML, or '' when the node carries none.
1613	*/
1614	private static function markup_for_node(array $node, array &$consumed): string {
1615	$text = self::first_value($node, self::CONTENT_KEYS);
1616	$url = self::url_from($node, self::URL_KEYS);
1617	$image = self::image_from($node);
1618	$video = self::video_from($node);
1619	$tag = self::heading_tag_from($node);
1620
1621	$parts = [];
1622
1623	// Video: an embed shape rather than a link, so the video detector can
1624	// see it while the link counters do not mistake it for an outbound
1625	// link. A file source becomes <video src>, anything else an <iframe>,
1626	// matching how the builder itself renders the two cases.
1627	if ('' !== $video) {
1628	$parts[] = self::is_video_file($video)
1629	? sprintf('<video src="%s"></video>', esc_url_raw($video))
1630	: sprintf('<iframe src="%s"></iframe>', esc_url_raw($video));
1631	}
1632
1633	// Image: alt text matters as much as the tag, since alt checks run over
1634	// whatever this returns.
1635	if ('' !== $image['url']) {
1636	$alt = '' !== $image['alt'] ? $image['alt'] : (string) self::first_value($node, self::ALT_KEYS);
1637	if ('' !== $alt) {
1638	$consumed[] = $alt;
1639	}
1640	$parts[] = sprintf(
1641	'<img src="%s" alt="%s" />',
1642	esc_url_raw($image['url']),
1643	htmlspecialchars($alt, ENT_QUOTES)
1644	);
1645	}
1646
1647	if ('' !== $text) {
1648	$inner = $text;
1649
1650	if ('' !== $url) {
1651	$consumed[] = $text;
1652	$inner = sprintf('<a href="%s">%s</a>', esc_url_raw($url), $text);
1653	}
1654
1655	if ('' !== $tag) {
1656	$consumed[] = $text;
1657	$parts[] = sprintf('<%1$s>%2$s</%1$s>', $tag, $inner);
1658	} elseif ('' !== $url) {
1659	$parts[] = $inner;
1660	}
1661	} elseif ('' !== $url) {
1662	// A destination with no label still counts as a link for link
1663	// checks; the URL doubles as its anchor text.
1664	$parts[] = sprintf('<a href="%1$s">%1$s</a>', esc_url_raw($url));
1665	}
1666
1667	return implode("\n", $parts);
1668	}
1669
1670	/**
1671	* First non-empty scalar value under any of the given keys.
1672	*
1673	* @param array $node Builder node.
1674	* @param string[] $keys Candidate keys.
1675	* @return string Trimmed value, or '' when none match.
1676	*/
1677	private static function first_value(array $node, array $keys): string {
1678	foreach ($node as $key => $value) {
1679	if (!is_string($key) \|\| !is_string($value)) {
1680	continue;
1681	}
1682	if (in_array(strtolower($key), $keys, true) && '' !== trim($value)) {
1683	return trim($value);
1684	}
1685	}
1686
1687	return '';
1688	}
1689
1690	/**
1691	* Link destination held by a node, as a bare string or a `{ url: … }` object.
1692	*
1693	* @param array $node Builder node.
1694	* @param string[] $keys Candidate keys.
1695	* @return string URL, or '' when the node holds none.
1696	*/
1697	private static function url_from(array $node, array $keys): string {
1698	foreach ($node as $key => $value) {
1699	if (!is_string($key) \|\| !in_array(strtolower($key), $keys, true)) {
1700	continue;
1701	}
1702
1703	if (is_string($value) && self::looks_like_url($value)) {
1704	return trim($value);
1705	}
1706
1707	// Elementor and Breakdance both nest the destination one level down.
1708	$nested_values = self::as_children($value);
1709	if (null !== $nested_values) {
1710	foreach ($nested_values as $nested_key => $nested) {
1711	if (is_string($nested_key)
1712	&& in_array(strtolower($nested_key), ['url', 'href', 'permalink'], true)
1713	&& is_string($nested)
1714	&& self::looks_like_url($nested)
1715	) {
1716	return trim($nested);
1717	}
1718	}
1719	}
1720	}
1721
1722	return '';
1723	}
1724
1725	/**
1726	* Image URL and alt text held by a node.
1727	*
1728	* @param array $node Builder node.
1729	* @return array{url:string,alt:string}
1730	*/
1731	private static function image_from(array $node): array {
1732	foreach ($node as $key => $value) {
1733	if (!is_string($key) \|\| !in_array(strtolower($key), self::IMAGE_KEYS, true)) {
1734	continue;
1735	}
1736
1737	if (is_string($value) && self::looks_like_url($value)) {
1738	return ['url' => trim($value), 'alt' => ''];
1739	}
1740
1741	$nested_values = self::as_children($value);
1742	if (null !== $nested_values) {
1743	$url = '';
1744	$alt = '';
1745	foreach ($nested_values as $nested_key => $nested) {
1746	if (!is_string($nested_key) \|\| !is_string($nested)) {
1747	continue;
1748	}
1749	$nested_key = strtolower($nested_key);
1750	if ('' === $url && in_array($nested_key, ['url', 'src'], true) && self::looks_like_url($nested)) {
1751	$url = trim($nested);
1752	}
1753	if ('' === $alt && in_array($nested_key, self::ALT_KEYS, true)) {
1754	$alt = trim($nested);
1755	}
1756	}
1757	if ('' !== $url) {
1758	return ['url' => $url, 'alt' => $alt];
1759	}
1760	}
1761	}
1762
1763	return ['url' => '', 'alt' => ''];
1764	}
1765
1766	/**
1767	* Heading tag a node asks for, normalised to h1–h6.
1768	*
1769	* Accepts both the `h2` form and a bare level like `2`.
1770	*
1771	* @param array $node Builder node.
1772	* @return string Tag name, or '' when the node is not a heading.
1773	*/
1774	private static function heading_tag_from(array $node): string {
1775	foreach ($node as $key => $value) {
1776	if (!is_string($key) \|\| !in_array(strtolower($key), self::HEADING_TAG_KEYS, true)) {
1777	continue;
1778	}
1779
1780	if (is_string($value) && preg_match('/^h([1-6])$/i', trim($value), $m)) {
1781	return 'h' . $m[1];
1782	}
1783
1784	// A bare level only counts under a key that unambiguously means one;
1785	// `size` and `tag` carry values like "large" or "div" far more often.
1786	if (is_numeric($value)
1787	&& in_array(strtolower($key), ['level'], true)
1788	&& (int) $value >= 1 && (int) $value <= 6
1789	) {
1790	return 'h' . (int) $value;
1791	}
1792	}
1793
1794	return '';
1795	}
1796
1797	/**
1798	* Whether a string is plausibly a link or asset destination.
1799	*
1800	* Deliberately permissive about relative paths — builders store internal
1801	* links that way — but rejects the option slugs and CSS values that make up
1802	* most of a builder tree.
1803	*
1804	* @param string $value Candidate.
1805	* @return bool
1806	*/
1807	private static function looks_like_url(string $value): bool {
1808	$value = trim($value);
1809
1810	if ('' === $value \|\| strlen($value) > 2048) {
1811	return false;
1812	}
1813
1814	if (preg_match('#^(https?:)?//#i', $value) \|\| str_starts_with($value, '/')) {
1815	return true;
1816	}
1817
1818	// Protocol-ish destinations a link node can legitimately hold.
1819	return (bool) preg_match('#^(mailto:\|tel:\|\#)#i', $value);
1820	}
1821
1822	/**
1823	* Whether a value carries nothing worth analyzing.
1824	*
1825	* Readable text is the usual signal, but not the only one: a page can be
1826	* made entirely of media. A builder section holding just a gallery
1827	* reconstructs to `<img>` tags and one holding just a video widget to a
1828	* single `<iframe>` — both strip to an empty string, so a text-only test
1829	* discarded them here and the page fell through to the next builder key,
1830	* then to the raw markup, and finally reported as having no content at all.
1831	*
1832	* Comments are dropped before the tag test: the raw markup this class falls
1833	* back to on a builder page is unrendered block comments, which must stay
1834	* blank rather than be mistaken for reconstructed media.
1835	*
1836	* @param string $value Candidate content.
1837	* @return bool
1838	*/
1839	private static function is_blank(string $value): bool {
1840	if ('' !== trim(wp_strip_all_tags($value))) {
1841	return false;
1842	}
1843
1844	$without_comments = (string) preg_replace('~<!--.*?-->~s', '', $value);
1845
1846	return 1 !== preg_match('~<(?:a\|img\|iframe\|video\|source)\b~i', $without_comments);
1847	}
1848	}
1849