Technical investigation
International SEO investigation
Finding two technical causes behind 612 multilingual pages Google barely saw.
Don't accept the first diagnosis
- 612pages Google barely saw; 538 never had an impressionSearch Console
- 2independent causes: a template rule and two generatorsBuild and source
- 6 of 6sampled URLs indexed; all 612 in the submitted sitemapsSearch Console
- 335 → 3,029organic sessions in equal 51-day windows. Association, not proof of causeGA4
- Challenge
Pages that existed, but Google barely saw.
I found 612 multilingual pages that Google barely saw: 538 of them had never had a single impression.
- My contribution
Index the right pages, on purpose.
I defined an approved list of 102 country-and-field pages, each in 6 languages, and implemented the indexing fix in the site's code, reviewed by the developer.
On a multilingual site of 4,800+ URLs in 6 languages, the country-guide section had lost about 70% of its Google clicks year on year.
- Delivery and measurement
Every page submitted. Every sampled page indexed.
Afterwards, all 612 pages were in the submitted sitemaps and 6 of 6 sampled URLs were indexed.
What this doesn't prove
This shows association, not proof of cause: results were still settling when I left, and seasonality and other site changes are not isolated.
Three surfaces disagree
Live: 200, self-canonical. Build: noindex. Prerender: none of the 612.
Two independent causes
A template rule and two generators, traced before any fix was written.
One small, reviewed change
Approved pages, template, sitemap generator and prerender step.
Coverage and a sample
All 612 in the submitted sitemaps; 6 of 6 sampled URLs indexed. A sample, not an indexing census.
The live pages looked fine. The build said otherwise.
Live URL /
200 · self-canonical
Neither observation establishes that Google indexed the pages.
Build output /
noindex
This is evidence about the build, not a documented trace of the directive Google received for every URL.
Prerender coverage /
None of the 612
Absence from prerender does not, by itself, establish how every page was crawled or rendered.
The pages returned 200 and were self-canonical, but the build marked them noindex and none of the 612 was prerendered.
I can't show the order in which I ran these checks.
Prove the cause before writing the fix
- 612 country-level field pages were missing from the sitemap and prerender output, and some carried a noindex tag.
- Patch the indexing rule at once, or trace the cause first and prepare a reviewable change.
- Trace first. The noindex came from the isCityLevel condition in the page template; the missing sitemap and prerender entries had a separate cause, because neither generator emitted country-level field pages. I prepared the change in an isolated copy.
- Changing indexing rules on an unverified report risks deindexing more pages.
- On 31 July the change existed only in that copy and was held for review. In the copy, the sitemap rose from 4,229 to 4,841 URLs and the prerender output from 4,250 to 4,862 (+612 each). What followed is recorded further down this page.
Don't accept the first diagnosis
- The first explanation for the missing pages named a single cause.
- Build the fix on that explanation, or test it against the build output before writing anything.
- I tested it, and it did not hold. There were two independent causes, so I corrected the explanation before the change was prepared.
- A fix aimed at one cause would have left the pages missing for the other.
- The correction is recorded in the working session before the change was prepared. This page does not show how long the first explanation stood.
An explicit list across six languages.
102approved country-and-field pages
×6languages per selection
How to read the 612
The repeated count of 612 in the visibility finding, prerender observation and later sitemap check should be read with its stated context. The public narrative does not provide a row-level reconciliation of those populations.
Four components, distinct responsibilities.
Approved-page list
Page template
Sitemap generator
Prerender step
Small change, checked before it shipped.
Before release I ran a safety audit that compared 164 files and corrected 57 links, and the developer reviewed the change.
Release-check limits
The comparison baseline, link types, audit methods and review acceptance criteria are not documented publicly. These counts cannot establish exhaustive regression coverage.
Submission coverage is not an indexing census.
612in submitted sitemaps
All 612 pages were represented in the submitted sitemaps. Submission coverage does not establish that all were indexed.
6 of 6sampled URLs indexed
Indexing was confirmed for the six sampled URLs. The sampling method and representativeness are not documented; this result cannot be extended to all 612.
Equal windows; bounded interpretation.
In equal 51-day windows before and after the fix, these pages went from 143 to 340 pages reached, 335 to 3,029 organic sessions and 4 to 134 key events.
Pages reached
143 → 340 pages reached
Organic sessions
335 → 3,029 organic sessions
Key events
4 → 134 key events
Each pair uses its own scale; bar lengths cannot be compared across metrics. Equal durations do not make this a controlled experiment.
This shows association, not proof of cause: results were still settling when I left, and seasonality and other site changes are not isolated.
Measurement limits
The windows are 11 June to 31 July 2026 and 1 August to 20 September 2026, counting Organic Search landing sessions on the 612 pages in the Piktalent GA4 property. “Pages reached” means pages with at least one organic landing session. The 134 key events are 129 contact-form submissions, 3 premium applications and 2 city waitlist sign-ups.
The split is the 31 July audit, not the deployment. I don't have the deployment date or the GA4 configuration history, and other site changes may overlap, so this is an association.
The earlier approximately 70% year-on-year click decline describes the country-guide section. These intervention-page outcomes have a different scope and comparison period; they do not establish recovery of the whole section.
What the record leaves unresolved.
The evidence supports an implemented code change with developer review, documented pre-release checks, sitemap coverage, a small indexing sample and associated before/after improvements. It does not establish complete indexing coverage or an isolated causal effect.
The selection rationale, population reconciliation, audit method, sampling method and measurement definitions remain undocumented publicly. Those details would require approved supporting evidence before a more specific technical account could be published.
I challenge a first diagnosis, separate two causes that looked like one, and prepare a reviewed change to the system rather than the symptom. The limit is real: the results show association, and they were still settling when I left.
Reflection
A page that answers isn't a page that's found.