1.Executive summary

Checks summary

2.AI crawler access

robots.txt: nothing blocked, but six dead rules

We did not assess robots.txt by reading it. We matched every rule against every URL in all three sitemaps using Google's matching semantics, because a rule that looks alarming often applies to nothing. That is exactly what happened here. Our own opening hypothesis — that Disallow: /blog/post/* was hiding the blog — was wrong, and we disproved it before it reached this report: articles live at /blog/<slug>, and 0 of 349 blog URLs match that prefix.

There is no per-bot section anywhere on any host for GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, Claude-SearchBot, Claude-User, PerplexityBot, Google-Extended or CCBot — all of them fall under User-agent: *, and * blocks no public content.

What we did find is decay. Of the 10 rules under User-agent: *, 6 point at URLs that now return 404: /app, /backend, /download, /translate, /case-studies/* and /blog/post/*. The remaining four (/join, /login, /settings, and a JS asset path) are legitimate. Today the dead rules cost nothing. They matter for two reasons: they already caused one analytical error in this audit, and at the next URL refactor a stale prefix can silently start matching live pages.

One rule is malformed rather than dead: the token User-agent: http://Mail.RU is not a valid user-agent string, so that block applies to no crawler at all. If blocking Mail.RU is intended, the token must be Mail.RU.

Live bot check: byte-identical responses

A robots.txt audit alone proves nothing — blocking usually happens at the CDN or WAF, below the file. So we requested the site as the bots themselves, comparing both status code and response size in bytes against a browser baseline. 13 AI user agents across 9–10 URLs, then a stability series of 3 runs × 7 UAs: all 63 measurements returned identical byte counts, with zero variance.

Common Crawl: the site is in the index, including the blog

Common Crawl is a training and retrieval corpus for many models, so presence there is a direct proxy for whether content has been collected. We queried the latest index, CC-MAIN-2026-30 (July 2026), via curl.

Content without JavaScript: no barrier on any platform

crowdin.com is not one system. We identified five stacks and checked each for whether content survives without JavaScript, since many AI crawlers do not execute it.

A raw-versus-rendered comparison first suggested a problem: the homepage went from 2,256 words to 3,999 after rendering, /pricing from 974 to 3,799. We checked what the extra text actually was, and it is the cookie-consent banner expanding into roughly 2,000 words of cookie tables. No substantive content block exists in the rendered DOM and is missing from raw HTML. The ratio was a false signal; we report the verified conclusion.

3.Technical audit

Worth stating plainly, because these are the checks that are expensive to fix later and are already right: every one of the 898 sitemap URLs returns 200, with no 4xx or 5xx among them. /this-page-does-not-exist returns a real 404 with a branded page rather than a soft-404. http redirects to https and www to non-www, both 301, with HSTS at max-age=31536000; includeSubdomains. There are 0 canonical conflicts across all 898 pages and 0 missing canonicals on core. All 898 titles are unique, there are 0 duplicate H1s and 0 pages carrying two H1s. Near-duplicate content across the 398 core and blog pages is 0.

Field data from real Chrome users (CrUX, 28 days to 15 Aug 2026) puts the origin outside threshold on three of four metrics on mobile: LCP 6,815 ms, INP 430 ms, FCP 6,718 ms, with 50.1% of views LCP-poor. Desktop is also failing: LCP 4,288 ms and CLS 0.16.

The aggregate hides where the damage is, which is why we broke it out by URL. Only five URLs have enough traffic for their own CrUX record, and the split is decisive.

So the marketing pages that matter for AI visibility are fine, and the origin-level failure is produced by the logged-in application. Two facts identify the cause precisely. Server response time across all 898 pages is 0.11–0.42 s, so the backend is fast. And lab testing of /project/ankidroid finds 1,920 KiB of unused JavaScript on a page whose entire text content is 562 words, with 7.2 s of main-thread work. The problem is client-side JavaScript weight, not infrastructure.

Crowdin publishes 21 language subdomains. We checked 20 of them on /pricing. 18 return HTTP 200 together with noindex, nofollow. At the same time 48 core pages publish 22 hreflang alternates pointing at those very URLs — the site tells search engines and models that translated versions exist, then forbids every one of them from being indexed. Localization spend is not converting into visibility.

Two of the 21 are not language versions at all. xh.crowdin.com and zu.crowdin.com return 26,346 bytes of identical HTML titled "Crowdin Enterprise" — a single-page app for a different product, with no hreflang, no canonical and no meta robots. Their /sitemap.xml returns an HTML document with status 200 instead of XML, so a crawler asking for the sitemap gets an app page and is never told anything went wrong. Separately, every locale's sitemap returns the main-domain URL set, so even if noindex were lifted tomorrow the locale sitemaps would cover nothing.

13 URLs sit in the sitemap carrying noindex — 7 core and blog pages plus 6 closed projects. The site simultaneously requests indexing and forbids it. We confirmed each with a browser user agent, so this is served to everyone and is not a crawler artifact. Four are money pages.

29 internal link targets return 404. The cause is a single retired system: 28 of 29 point at store.crowdin.com app pages linked from "What's new at Crowdin" posts published between 2020 and 2026, where the apps were renamed or removed and the links never updated. The 29th is a slug typo — crowdinn with a double n — and it sits on the money page /customers. Note that the crawler's own inlink export reported zero broken links; that was a list-mode artifact, so we extracted 3,154 link targets from raw HTML and checked the 1,314 Crowdin-domain ones directly.

Beyond outright breaks, 162 internal links land on a redirect rather than the final URL, and 43 chains run 2 to 4 hops (no loops). Common cases: /ai-localization → /features/ai-translation, /alternatives → /, and legacy dated blog URLs such as blog.crowdin.com/2021/08/25/translation-memory/ taking three hops.

Two things stand out. br is the code for Breton, not Brazil; given the URL br.crowdin.com the intent was Brazilian Portuguese, whose correct code is pt-BR — and the /project/* template already uses it correctly. So the marketing pages advertise a Breton edition that does not exist while never declaring the Brazilian one. That the two templates disagree is itself the diagnostic: hreflang changes are shipping out of sync, and the core template is running an older version. Meanwhile the entire blog — the section that normally earns both search and LLM traffic — sits outside the international configuration completely.

249 URLs in sitemap-community0.xml canonicalise to a different address purely because their slugs contain capital letters (/project/My_F-droid, /project/Nyachi). No lowercase twin appears in the sitemap, so the file systematically submits the non-canonical variant. Also: lastmod is absent on 48 of 49 core URLs, and there are 2 duplicate <loc> entries.

4.Readiness for AI answers

The /project/* catalogue is 19,345 URLs, 98.0% of the site, and it is indexable, in the sitemap, and statistically a single document. We measured sameness three independent ways and all three agree:

Note the trap in that table: 562 words clears any sane thin-content threshold, so a word-count test would clear these pages. They are thin by substance, not by length.

They are also unreachable except by sitemap. The /projects listing emits zero links to any project page in raw HTML — the list is built client-side, and ?page=1 and ?page=2 return byte-identical HTML at 173,953 bytes. The crawler appears to show 6,490 inlinks to these pages, but every one is a self-link from the page's own hreflang and in-app navigation: cross-page links are 0, and no core or blog page links to a single project. All 19,345 are true orphans.

Crowdin's pricing is published only inside JSON-LD markup. We verified this from the primary source rather than a crawler report:

The "Compare Plans" table is empty in the markup: all 42 of its cells contain no text, each holding a loading skeleton instead — the page carries 69 animate-pulse placeholders and 23 hydration islands. Crucially this is not fixable by rendering: the crawler ran with JavaScript enabled and still recorded no prices, because the table hydrates from data that never arrives inside the render window. The plan names "Pro" and "Team+" likewise appear only inside the markup; visible text has only "Free plan" and "Enterprise".

Structured data must describe what a visitor sees. On two pages it does not:

On the first two, the words "FAQ" and "Frequently" do not appear on the page at all — we checked the live response, 434 KB and 220 KB respectively. Corpus-wide this is 15 of 285 questions (5.3%) not visible, and 11 of those 15 fall on those two pages. The remaining four are single omissions on blog articles and are ordinary content debt.

Worth knowing before prioritising: across all 898 pages Google generates no FAQ rich result whatsoever for this site. So the hidden markup carries policy risk with zero search upside; its only remaining value is LLM extraction, which is exactly why the questions belong in the visible text.

When a model is asked to recommend a localization platform it does not read one page; it decomposes the request into follow-up questions and looks for each. We ran the twelve questions a real buyer asks against each of the top ten money pages — 120 checks — and graded each by reading the page.

Read the pattern rather than the total. Crowdin covers how the product works and answers poorly on why to choose it — price, security, comparison and fit. Those are the selection criteria, which is precisely why the model names Crowdin in 11 of 12 answers and recommends it in none. SELF at 0 of 5 reproduces the one prompt where Crowdin was absent entirely: asked for a self-hosted Lokalise alternative, the model returned Tolgee, Weblate, Traduora and Pontoon. The words "self-hosted", "on-premise" and "private cloud" appear on none of the ten money pages, even though Crowdin Enterprise offers it.

/alternatives/lokalise-alternative is one of the two most important pages for AI visibility, because "X vs Y" and "X alternatives" are exactly the queries models fire when asked to recommend a tool. Its comparison tables have 168 cells, of which 83 (49%) contain no text at all — the value is carried by an inline SVG marked aria-hidden="true", with no label and no title: 55 checkmarks and 28 crosses, plus 97 aria-hidden attributes on the page. Extracted as text, a feature row reads:

| Continuous AI fine-tuning |  |  |

Two whole tables — "Similarities Between Lokalise and Crowdin" and "Crowdin Enterprise Unique Features", 15 rows each — contain not one value in text form. A model reading this page cannot tell who wins any comparison.

The fix is already written. The neighbouring /alternatives/phrase-alternative uses the same component and renders 40 of 43 cells as text — only 3 are icon-only, and it carries 18 aria-hidden attributes against 97. It spells values out in words ("Available on all plans at no extra cost from us") instead of drawing a checkmark. So this is one component change, reusing a pattern that already works on the adjacent page.

Across the whole corpus, E-E-A-T scoring splits cleanly by template:

44 of 46 commercial pages score below 60, with a median of 48, against a blog median of 87. Only two clear 60: the homepage at 88 and /learn at 62. The signal string attached to nearly every failing page is identical — no author; external links 0; no dates — and the pages sharing it are the /features/* and /solutions/* templates. The same repeated cause means the same single fix: adding an author or reviewer, a visible date and outbound references to the two templates raises all 44.

What makes the gap notable is that Crowdin clearly knows how to do this. The blog carries a named author on 100% of 311 articles. The capability exists; it just stops at the templates that sell.

Three commercial pages are orphans: exactly one file in the entire 898-page crawl links to each of them, and that file is the page itself. Not the homepage, not /features, not /solutions, not any article.

Two of the three are the alternatives pages — the highest-value comparison content Crowdin owns, and the exact page type models look for. A further two money pages (/features/workload-management and /solutions) have inlinks from navigation only and zero editorial links.

{count} appears in the visible body text of 12 pages, including the homepage, where a visitor and a model both read:

You have content in {count}+ disconnected tools (GitHub, Figma, CMS)

The full list: /, /features/quality-assurance, /features/translators-workbench, /learn/continuous-localization-ebook, /learn/mobile-app-localization-ebook, /product/agency-partners, /solutions/customer-support-translation, /solutions/e-commerce-translation, /solutions/game-localization, /solutions/software-localization, /solutions/website-translation and /webinar/mobile-app-localization. It also leaks into 3 meta descriptions and 6 JSON-LD nodes (where {name} and {user_name} also appear) — three layers of the same unsubstituted variable.

The damage is specific. Every one of these slots is where a number belongs: "100+ file formats", "40+ MT engines", "700+ integrations". A model extracting facts gets a broken token exactly where the quantified claim should be — the sales argument is lost precisely where it was meant to land.

One related item, separate because it is genuinely isolated: a live H2 on the money page /features/in-context-translations reads "Highlight the specific "Web" pain points here" — a copywriter's instruction shipped to production. We checked scope across all 898 pages: this phrase appears on 1 page, and Lorem ipsum, TODO and FIXME appear on 0. A single incident, not a release process problem.

Organization markup is present on 359 pages and, to Crowdin's credit, all of them use one stable identifier — crowdin.com/#organization, with no competing variants. But sameAs exists on exactly one page: the homepage. The other 358 carry a stub of name, URL and logo with no sameAs, no contactPoint, no address and no founding date. Measured against the full site, the company's verifiable identity is published on 1 URL in 19,743.

The cause is mechanical and therefore cheap: only 4 unique payloads exist across the entire site, byte-identical within each template, and the same stub appears on the locale subdomains too. The defect is in a shared Astro partial, not in individual pages — one edit closes all 358.

Two further graph defects compound it. Critical: on 311 articles #organization is declared twice with different fields — one node carrying description, the other legalName — because two markup generators in the theme both emit it; #website is duplicated the same way. A consumer consolidating that graph gets a non-deterministic result. High: on 43 of 48 core pages, about and mainEntity point at a #webapplication node that is not present on those pages — a dangling reference.

Authorship is Crowdin's strongest E-E-A-T asset and its most under-published. 311 of 311 articles (100%) have a named author, confirmed visible on the page in a random sample of 40. There are 42 authors. There are 0 author pages: /blog/author/* returns 404, and no such URL appears anywhere in the crawl.

The markup is equally incomplete. Across 312 Person nodes, the fill rate for @id, url, jobTitle, worksFor, image and knowsAbout is 0%. Job titles exist — as prose inside the bio ("currently leads a marketing team at Crowdin") rather than as a field a machine can read. So the author is a name, not an addressable entity.

Coverage is cheap to fix: 7 author pages would cover 274 of 311 articles (88%) — Diana Voroniak (108 articles), Yuliia Makarenko (70), Iryna Namaka (39), Khrystyna Humenna (25), Julia Herasymchuk (15), Yana Feshchuk (7), Andrii Bodnar (2).

One item needs correcting before any other authorship work, because it is not a broken link but a false statement about a person: on /blog/translation-quality-without-speaking-language the author Ronak Ganatra's sameAs points at linkedin.com/in/akanksha-babbar-14583420/ — a different person, who writes a different article on the same site. The markup asserts that two people are one.

Freshness is a ranking and citation factor for models, and Crowdin is losing credit for work it has actually done. 98 articles display "Last updated: <date>" to readers. Only 52 also carry dateModified in markup. For the other 46, the sitemap reports the original publication date. The clearest example:

A five-year-old timestamp on freshly updated content is worse than no timestamp. Separately, 0 of the 10 top money pages have a determinable date by any source, and lastmod is missing from 48 of 49 core sitemap URLs. Two smaller markup defects sit alongside: unrendered markdown leaked into 293 Person.description fields, and 6 pages ship placeholder tokens inside JSON-LD.

13 keyword clusters spanning 46 pages compete internally, and 7 of those pit a commercial page against an informational one. Two matter commercially:

The second is really a differentiation problem. The /solutions/* pages share a template, so the generic category term surfaces on all of them rather than "ecommerce translation software" or "game localization services".

We also found 5 page pairs above 0.85 semantic similarity — and this is a case where the obvious action is the wrong one. Literal text overlap on those pairs is only 0.095–0.146: they discuss genuinely different subjects (CakePHP versus Django localization; academic versus open-source programmes) in near-identical phrasing. Do not merge them with redirects or canonicals. They need differentiated copy. Merging would delete real content.

Finally, facts contradict each other across pages, and any of these could end up quoted:

Two of our checks scored the same ten pages on the same 0–100 scale and disagreed: 88.9 and 65.1. Rather than average them or pick the friendlier one, we reread the pages. Both are correct, and they measure different things.

/pricing shows why the gap exists. On structural signals it scores 90: one table, six H2s, nine marked-up questions. Read it and all 42 table cells are empty, behind 69 loading skeletons, with the prices only in JSON-LD and the FAQ questions absent from the text. The container is there; the content is not. The same pattern hits /alternatives/lokalise-alternative (92 structural, 54 on content). We use 65.1 throughout this report and treat 88.9 as the ceiling the existing structure would reach once filled.

5.Visibility in model answers

How often each vendor's domain is actually cited as a source in ChatGPT answers, measured on the LLM Mentions dataset for the US in English:

Set that against brand demand and the picture sharpens. Crowdin's AI search volume is 643 — higher than Lokalise's 441 and Transifex's 259. Lokalise earns 58 citations on 441 of demand; Crowdin earns 15 on 643. People are asking about Crowdin more than about Lokalise, and the models are citing it four times less. Demand is not the constraint.

A second, independent signal points the same way. Asked for the brands associated with "translation management", the dataset returns 42 brands — Phrase, DeepL, Trados, memoQ, Wordfast, OmegaT, Transifex, Localizely, Smartcat, Lokalise, TransPerfect and others. Crowdin is absent entirely. We treat that as a fact rather than a glitch for a specific reason: the query returned 42 results and Crowdin's own competitors are among them, so the dataset is neither empty nor broken. Had it failed, the competitors would be missing too.

We ran twelve prompts through ChatGPT in clean sessions, fixed in advance and shuffled in order, covering three types: keyword-mapped, community-style and buyer-scenario. All results are reported, including the misses.

Crowdin appears in 11 of 12 at an average position of 2.9, and takes first place only in the branded "crowdin vs lokalise". For comparison, Lokalise appears in 12 of 12 and Phrase in 10. And in none of the twelve answers was crowdin.com cited as a source — the mentions come from model memory, with no link back to the site. Crowdin is known and not quoted.

The one complete absence is the most instructive. Asked for a self-hosted Lokalise alternative, the model named Tolgee, Weblate, Traduora and Pontoon — Crowdin Enterprise offers self-hosting, and the model does not know it, because as section 4 showed the phrases "self-hosted", "on-premise" and "private cloud" appear on none of the ten money pages.

Across the twelve answers the model consistently frames Crowdin as "open source, community, developer-heavy". Verbatim from one answer: "Choose Crowdin if your content lives in GitHub, GitLab, or similar development workflows, or if you rely on community translation", with the table row reading "Software + open-source/community". "Best overall" goes to Phrase; "best developer-first" to Lokalise. In the SaaS scenario Crowdin gets "Good for larger translation programs" while Phrase gets "Excellent" and the "best overall" verdict — a difference of adjectives that decides the shortlist.

This is not a distortion to be corrected with PR. It is an accurate paraphrase of the pages. /solutions/software-localization carries the H2 "Built by developers, for developers"; /pricing has a dedicated "Crowdin for Open Source" section; /product/for-open-source and /product/for-academic exist as standalone pages. The phrases "for product teams" and "for SaaS companies" appear on none of the ten money pages. The model is reading the site correctly and reporting what it says. That means the label is fixable only by changing the copy — and, encouragingly, that the copy is the whole of what needs changing.

One bright spot to build on: where Crowdin is cited, the blog does the work. Of the 15 prompts in which a Crowdin URL was cited at all, four cite a real page — three of them blog articles ("is crowdin a cat tool?" cites /blog/cat-tools; "what are the three types of localization?" cites /blog/software-localization). The fourth is /project/minecraft, a lucky win on someone else's brand out of 19,345 UGC pages. The editorial content earns citations; the commercial pages do not.

6.Semantics & content plan

Advertisers do not pay for clicks that do not convert. That makes the top-of-page bid an independent check on intent, and one that volume alone cannot give. The contrast is stark:

Applying that as a rule — volume above 5,000 with a top-of-page bid under $2 is presumed not to be a buyer — removed 88,500 searches a month, 41% of what first looked like the gap. Our own initial analysis had multilingual user interface as the single biggest unclaimed opportunity on the site. It is Windows language packs. We withdrew it, and the control confirms the withdrawal: seeding the keyword tool with that term returns exactly two keywords, itself and "windows multilingual user interface" — it has no semantic neighbourhood in this category because it does not belong to it.

The same discipline killed a second false opportunity. types of computer-assisted translation showed 90,500 searches a month. Its bid is $0.02–0.07, and /blog/cat-tools already ranks eighth for it. Had we skipped the subtraction step, this report would have recommended building a page Crowdin already has, for a query with no commercial value, as priority one.

The honest consequence of all this filtering: our gap figure came down from 361 keywords / 234,320 searches to 297 keywords / 126,880 — roughly halved. The smaller number is the one worth acting on.

This is the cheapest work in the report and it is not content creation. 89 keywords worth 25,720 searches a month belong to articles Crowdin has already published, and competitors outrank them. The pattern is consistent: Crowdin has the article, someone else has the position.

Note this survived our own filtering: it was 112 keywords in the first pass and is 58% smaller now — and it is still priority one, because strengthening a published URL costs a fraction of launching a new one.

Crowdin publishes two comparison pages, for Lokalise and Phrase. There are 13 direct competitors. Missing: Smartling, Transifex, Smartcat, POEditor, Weglot, Centus, OneSky, memoQ, Trados, XTM and Memsource.

The case for building them is not volume — phrase alternatives is 260 searches a month. It rests on two things. First, competitor brand queries where Crowdin has no presence total 13,510 searches a month at bids reaching $49 (weglot 3,600, smartling 2,400 at $49.09, lokalise 1,000 at $19.64, transifex 390 at $22.16). Second, and more important here: "X alternatives" and "X vs Y" are precisely the fan-out queries a model issues when asked to recommend a TMS. This is the page type that gets read by machines at decision time. The two pages that exist are currently orphans (section 4), which should be fixed before more are built.

18 keywords validated by two or more competitors, difficulty 2–20, 3,630 searches a month. This has the highest competitor validation of anything in the set at the lowest difficulty, which makes it the most reliable demand signal we have: react-intl (480, KD 3), flutter localization (320, KD 4), flutter translate (210), angular languages (210), translate wordpress plugin (390, bid $14.92). It overlaps heavily with priority 1, which is a point in its favour — the same work serves both.

Two smaller terms deserve explicit mention because volume-based planning would discard them: lqa at 1,300 a month, KD 5, bid $22.16 and mtpe at 260 a month, KD 0, bid $8–12. Both are validated by Lokalise and Smartling, both are unmistakably ICP vocabulary, and both are ideally shaped for LLM citation — a defined process term with a precise answer. They are top-three by commercial value despite tiny volume.

7.External placements

We took eight direct competitors and pulled every domain linking to two or more of them but not to Crowdin. The pool was exhausted — all 2,247 available rows — giving 2,026 unique donors, of which 4 link to all eight competitors and 26 to six. After filtering, 1,074 are editorial, 689 have authority rank 40 or above, and 249 rank 60 or above.

Only a minority can be bought. Probing the 689 editorial donors for guest-post and contributor pages found buyability signals on 44 of them — 6.4% (a random sample of 80 predicted around 7.5%, consistent with the full sweep). The conclusion matters for budgeting: this niche monetises through editorial reviews and analyst coverage, so most of the gap closes by pitching, not purchasing.

github.com/oh-jon-paul/awesome-i18n — 408 stars, last updated 24 July 2026 — does not mention Crowdin at all. It is the largest and only actively maintained awesome-list in the category (the runner-up has 192 stars). It already has an open section, "Apps and extensions for translation management", which lists POEditor, and SimpleLocalize appears 6 times across the file. Crowdin appears zero times. The entry format is a single line, and 4 of the last 5 pull requests were merged.

Awesome-lists are heavily represented in model training and retrieval data, which is why a one-line addition to a 408-star list is the best effort-to-effect ratio in this report. We surveyed all five relevant lists so the picture is complete:

Crowdin's review presence is strong in two places and weak in the one that is structurally built to answer brand queries:

Two reasons this matters more than the numbers suggest. Trustpilot is structured precisely for "<brand> reviews" queries — the URL is /review/crowdin.com and the page title is "Crowdin Reviews" — so it surfaces on exactly the query a buyer runs before deciding, whereas G2 competes against hundreds of its own comparison pages. And for a model, a small low-rated sample is more dangerous than a large one: 3.8 gets weighted alongside 4.4, and the 3.8–4.7 spread across platforms reads as uneven quality rather than as a sampling artifact.

890 domains are excluded, each with a stated reason: 578 for spam score at or above 30, 540 for authority below 20, 111 on risky TLDs, 106 on free subdomain hosts, 44 with spam terms in the hostname, and 34 as non-editorial noise (job boards, URL shorteners, CDNs).

Spam score alone was not enough, and here is the case that proves it. 24 domains link to five or more competitors at once while ranking below 40 — the signature of a link network or a scraper rather than editorial interest. Several would have passed a spam-score filter comfortably: stackreaction.com (rank 34, all 8 competitors, spam 18), changelogwp.com (rank 35, 7 competitors, spam 9), hubbry.com (rank 34, 7 competitors, spam 24). Linking to every competitor in the category is what exposes them, so we required two signals rather than one.

A second exclusion group is the opposite trap — very high authority, useless as targets: 14 Glassdoor locales at rank 85–90, indeed.com (91), tinyurl.com (90), aliyuncs.com (86, a CDN linking to all 8), and site builders like Wix and Squarespace at 89–91. Sorting purely by authority puts these at the top of the list.

8.Roadmap

A.Methodology & limitations

A. Methodology, sources and limitations

Crawl. Screaming Frog 24.3 on a dedicated server, list mode with JavaScript rendering enabled, 898 URLs: all 49 core pages, all 349 blog URLs, and a random sample of 500 from the 19,345 /project/* pages — 2.6%, drawn with random.seed(20260817) for reproducibility. Proportions from that sample are representative; absolute figures for the UGC section are extrapolations and are labelled as such. We sampled randomly rather than taking the first or most prominent URLs, because prominent pages are systematically biased — they are where fixes get applied first, so testing a rule there reports "works" when the site-wide answer is different.

Site size is measured from the sitemap, not the crawl. The site is 19,743 URLs (19,345 UGC + 349 blog + 49 core), not 898. Page counts derived from a crawl or from pagination are false derivatives: crawl scope is a decision we made, not a property of the site.

Search volume and bids come from Google Ads Keyword Planner — the primary source, exact rather than modelled, and free. Third-party volume estimates were used for nothing in the final figures. Keyword difficulty is on the DataForSEO scale and is not comparable with Ahrefs figures in earlier reports.

Page speed verdicts come from CrUX field data (real Chrome users, 28 days to 15 Aug 2026). PageSpeed Insights lab data is used only to identify causes. Where the two diverge by a factor of several, that is normal and the field data governs. Per-URL field data exists for only 5 URLs because CrUX withholds low-traffic records — a property of the source, not a measurement failure.

The HTTP Last-Modified header was deliberately excluded. It is populated on 397 pages but holds just 8 distinct values, all inside a 31-second window matching our crawl time — it reports CDN cache age. Using it would have marked 397 pages as fresh and erased the freshness findings entirely.

Embeddings are TF-IDF with LSA over 200 components (explained variance 0.634), not neural embeddings — no external API was involved. Cosine similarity was computed over 79,003 page pairs across the 398 core and blog pages; only 8 exceeded 0.85. The UGC class was handled as a class rather than as 124,750 pairs. Boilerplate above 30% document frequency was stripped, and 75 release-note articles plus 23 pagination pages were excluded as template repetition — without that filter, one keyword alone would have produced a false 22-page cannibalisation cluster.

Ahrefs was not used in this run by agreement, so there is no Brand Radar share of voice, no Ahrefs difficulty or content gap, and no cross-check of index completeness. Backlink data comes from DataForSEO, whose index trails Ahrefs by 5–29% on live domains — so every absolute link figure in section 7 is a lower bound. Total third-party data spend for the audit was $2.58.

Corrections we made to our own work, disclosed because they changed conclusions. Our opening hypothesis that robots.txt was blocking the blog was disproved at the primary source and never reached this report. An automated fan-out pass scoring 66% was discarded for 53% after we read all 120 evidence strings. A 90,500-a-month "gap" was withdrawn once we checked that /blog/cat-tools already ranks eighth for it. A 33,100-a-month "opportunity" was withdrawn on bid evidence. Two extractability scores that disagreed were reconciled by rereading the pages rather than averaged. And a Wikipedia "gap" was rejected after reading the article in full. In each case the derived signal was wrong and the primary source corrected it.

B. Blocked checks and what closes them

Eight checks were not performed. None of them is reported as a zero result.

C. Data artifacts

This report is a thin layer over the data. Every finding is traceable to a row in the client spreadsheet, and the tabs below are linked individually from the sections above.

▧ Open the full audit spreadsheet