A translated page can still supply English sharing text

Open Graph gives a page separate fields for its sharing title, description and image. These are useful source inputs to inspect, but our ordinary website requests did not show how any social platform would choose, cache or display them. We therefore describe supplied content throughout this study. ^[2]

We counted a narrow text event: a local-language page supplied a title or description that exactly matched the corresponding English field, contained at least five alphabetic words, and included an ordinary English promotional clause. A company name, product name or short label such as “Home” did not qualify.

Four of the 29 eligible pairs met that rule. One retained an English title; three retained an English description. The cases were concentrated in two website groups. A pair means one local-language homepage compared with its English counterpart, not an independently sampled company.

We summarize the same observations in two ways. Giving each of the 12 contributing groups equal weight produces an average retained-English share of 10.4%. Pooling all 29 language-page pairs produces 13.8%. A group with four eligible languages contributes four observations to the pooled rate, but still only one within-group share to the first average. ^[3]

Neither percentage describes the wider software industry. Only 12 of the original 60 selected groups contributed comparable pairs under our discovery rules. For the other groups, limited discovery, access or language evidence prevented this particular comparison; that is not a successful localization check.

The practical finding is specific. A local-language page can retain an English promotional sentence in separately supplied sharing copy. A review that stops at the page heading would not directly examine that field. The appropriate next step is to read the supplied title and description, rather than infer them from the page.

This event is narrower than finding any English anywhere. A differently worded English sentence would not meet the exact-match rule, and missing fields cannot establish retention. We also did not observe why a team chose a particular language. The finding identifies content to review against a campaign’s intended audience; it does not establish an error or its cause.

The image comparison starts with a different denominator. We selected the 18 eligible pairs whose Open Graph title or description contained local-language or neutral wording, with at least one nonempty field and no English or unresolved prose. Seven other pairs had neither text field; four had the retained-English event. Missing text was not treated as translated text.

Within those 18 pairs, six supplied a first Open Graph image with clear English promotional wording. The text beside an image could therefore be local while the image itself carried an English message. These six observations reused just two image files across two groups; they are not six independent creative examples.

Seven pairs supplied wordless or brand-only artwork. A reused symbol or proper company name did not become an English-language problem merely because several language pages used it. The study gives no reason to require a different picture for every locale when the existing image carries no message that needs translation.

Three pairs shared one image whose readable corporate-affiliation wording divided the reviewers. One read it as part of the brand lockup; the other read meaningful English beyond the proper names. We kept that disagreement unresolved. The resulting range is six definite to nine possible English-image cases out of 18, not a confidence interval. ^[4]

The remaining two pairs supplied no image property. That is a different observation from an image with the wrong language, and it does not establish what a platform would show instead. A useful review distinguishes supplied wording, neutral artwork, missing source material and unresolved interpretation.

For a marketing manager, that distinction changes the assignment. “Translate the preview” is imprecise when the title is already local and the image contains the remaining message. Naming the particular image and its wording gives a designer a clearer task. Keeping neutral artwork can also avoid creating unnecessary variants.

A language decision and a design decision can have different owners. The person approving local copy can identify whether a headline needs adaptation, while the designer can judge how to accommodate it. Keeping those questions distinct also helps with longer translations: the objective is readable wording in the chosen composition, not merely a different image filename.

Fifteen eligible pairs explicitly declared an Open Graph locale whose primary language matched the local page. None had our retained-English text event. Fourteen omitted that declaration. This pilot does not support a claim that matching locale tags concealed retained English prose. ^[5]

A declaration and the words it describes are nevertheless different observations. A reviewer can record both without assuming that one substitutes for the other. The image check also remains a separate task; an optional locale field does not itself describe every word visible inside an image.

We inspected explicit X title and description fields only when the page supplied them. Of 36 present fields, 35 exactly matched their Open Graph counterpart. The remaining one differed in wording but still used the target language. This is supporting context, not evidence about X fallback behavior or its displayed cards.

For a release handoff, keep a source-content check separate from a platform-specific preview check. The former tells the team what the website supplied at a stated time. The latter needs an actual observation on the intended platform. Our dataset completes the first kind of check and makes no claim about the second.

Consider a fictional scheduling app preparing a German-language launch. Its homepage explains the product in German. The team has also translated the supplied sharing title and description. Its generic promotional image still contains the English line “Make room for the work you love.” This example is invented to illustrate the decision, not drawn from an audited company.

The copy reviewer can mark the page and sharing text as checked, while assigning the image wording for review. The designer has at least two reasonable options: create a version with appropriate German wording, or use an approved wordless composition. The choice depends on the campaign message and brand requirements, not a rule that every language must receive unique artwork.

Suppose another language page uses only the app’s logo. Replacing it merely to make the files different adds work without addressing a language-bearing message. The reviewer should describe it as brand-only artwork, preserving the reason it can be reused.

After an image change, the team checks the exact URL supplied by the local page. Approving a design file does not establish that the page points to it. If the campaign also requires confirmation on a particular social platform, the team records that separately, including the date and URL observed. The resulting handoff distinguishes approval of the source content from confirmation of a particular platform result.

The following brief can sit beside a localized landing-page or campaign review. It is a proposed workflow drawn from the measured distinctions, not an observed process used by the sampled companies. Give each line an owner and a dated observation so that a broad “localization complete” label does not hide an unchecked component.

To gather the source fields, ask the person who maintains the website: “For this exact page URL, please send the sharing title, sharing description and full image URL the page supplies. Mark missing fields as missing.”

Page and audience: [exact landing-page URL], [intended language], [campaign or audience], [review date]. Describe the page actually reached after redirects. A local-language link that reaches a different product page needs a comparability decision before its wording is assessed.

Main content: [reviewed heading or substantive page copy], [language observation], [any mixed-language sections relevant to the campaign]. Do not turn the presence of some translated prose into approval of every part of the page.

Sharing text: [supplied title], [supplied description], [reviewer and decision]. Read the actual wording. Separate proper names and short labels from promotional sentences. If a field is missing, record that absence instead of copying the page heading into the review as though it had been supplied.

Image: [exact supplied image URL], [wording or brand-only observation], [decision and owner]. Inspect the file that the page names. State whether the work concerns a translated headline, an approved neutral image, a missing asset or an unresolved interpretation.

Release evidence: [which source fields were checked], [which image was inspected], [time of observation]. Add [platform, exact shared URL and observation time] only when someone actually checks a platform-specific result. Do not merge that observation with an assumption about every other platform.

Follow-up: [specific component], [responsible person], [next check]. “Review the English image headline” is actionable. “Fix localization” leaves the next person to rediscover the issue.

If one person owns the entire page, that person can complete each observation in the brief. If work is split across an agency, designer and internal marketer, each can see the exact component assigned to them. A dated, specific observation is more useful than an unexplained green check beside a language name.

We selected 60 website groups from a frozen Wikidata frame on October 17, 2026. The frame covered software-company entities with an official website, at least 10 Wikimedia sitelinks and no recorded dissolution. Shared website domains were grouped before selection. The source had 443 entities in 420 groups; the fixed hash-based selection was not replaced when a website proved inaccessible. ^[6] The frame includes gaming and other software businesses; it is not an exclusively SaaS sample or a sample of creators.

For each group, the requested comparisons were English against German, Spanish, French and Japanese. We followed published language links under fixed request limits, rather than guessing paths or searching until a suitable page appeared. Of 240 requested slots, 29 were eligible, 123 had no target-language counterpart discovered, 52 had an unavailable baseline, 16 had no English baseline discovered, 16 had uncertain baseline language, and four redirected to the baseline language.

Eligibility required comparable homepage purpose and actual language-bearing source content. Before sharing outcomes were exposed, we clarified that a literal HTML main element or heading tag was not required when the main homepage content could be identified in the source. All readable pages were reviewed uniformly. The original stricter subset remained a sensitivity check: it contained 13 pairs across five groups and no retained-English text events. Its local-or-neutral-text subset still had three English-image cases among nine pairs, reusing one asset. ^[7]

We used ordinary public requests, respected robots and access restrictions, and read the initial response without executing page scripts. Real metadata in the whole response was retained, with head-only and first-occurrence comparisons. Those alternatives did not change the observed title, description or locale values. The study did not test a social crawler, a logged-in user or a rendered page.

Two independent AI-assisted readings classified all sharing text and the original fetched images. These were not human or native-speaker certifications. Text normalization decoded HTML entities once, standardized Unicode and whitespace, and preserved case and punctuation. Unresolved image wording stayed unresolved; the public files retain its effect on the counts. ^[8]

The download reproduces the arithmetic from all 240 coded slots, with fresh pseudonymous group and image-cluster identifiers. It does not redistribute original website copy, HTML, images or an identity mapping. Readers can recalculate every published denominator and comparison; they cannot independently recreate the original language judgments from the public bundle alone. A later website fetch would also be a new snapshot.

Earlier research already examines language differences between visible page content and accessibility text. Our narrower contribution is to compare the same homepage role across languages, then distinguish exact English sharing-prose retention from the wording inside its supplied image. That distinction supports a concrete review decision without turning this small selected sample into a general failure rate. ^[9]