Key takeaways
Compare raw and rendered hreflang annotations using the same pages and conditions. HTTP Archive records them as separate detection flags. [1]
The benchmark below uses four page groups from July 2025, each with its own page-row denominator. [1]
A detected annotation or changed value set is a starting point for checking the actual alternate URLs and their relationships. [1] [2]
Google supports HTML, HTTP headers and sitemaps for declaring localized alternates. Identify the method your site uses before judging a missing HTML annotation. [2]
When hreflang differs before and after rendering, check which annotations changed, which localized URLs they name and whether those relationships match your intent. Start by identifying your implementation method. Then compare the pages under consistent conditions and document what needs correcting.
This review is useful when a content team maintains several language versions of a product, help or documentation page. It gives the team a concrete technical question to investigate alongside the localized URLs it sees in saved AI answers.
Start with the method your site uses
Google Search accepts three ways to declare alternate language or regional versions of content, using HTML annotations, HTTP response headers or a sitemap. It treats those methods as equivalent choices. For an HTML implementation, the annotations belong in the page’s head. [2]
Ask the team that maintains localization which method is intended and where it is controlled. The answer helps you choose the evidence to inspect. An HTML review needs the response and rendered page. A header or sitemap implementation needs its corresponding declaration. Record that choice in the review so another person can repeat the check.
Choose a small group of pages that represents a real content decision. A translated installation guide is a more useful starting point than an undefined collection of international URLs. Write down which versions are meant to be alternates and what a reader should find on each one.
What the raw and rendered benchmark measures
HTTP Archive’s published SEO results contain separate flags for hreflang detection in raw HTML, rendered HTML and headers, plus flags for rendered-only detection and changed values. The chart uses the four desktop/mobile and homepage/secondary-page rows from the initial result grid linked by figure 5.57 of the 2025 Web Almanac. [1]
HTTP Archive’s July 2025 results, grouped by desktop/mobile and homepage/secondary-page samples.
- Raw HTML
- Rendered HTML
Full data table
| Client and sampled page group | Raw HTML | Rendered HTML |
|---|---|---|
| D home | 10.2415 | 10.4315 |
| D secondary | 10.092 | 10.3813 |
| M home | 9.9992 | 10.1996 |
| M secondary | 9.7049 | 9.9922 |
Ansvra multiplied the published fractions by 100. Each series uses all query page rows in its own stratum as the denominator. Detection does not validate the alternate URLs or demonstrate AI citations. D means desktop, M means mobile. Secondary means sampled secondary pages. [1]
In each group, the query divides the number of rows with a positive detection flag by all page rows in that group. Ansvra converted the published fractions to percentages. The chart therefore gives context for what the collector detected under its July 2025 conditions. The group totals remain separate throughout the analysis. [1]
Complete supporting results
The following tables retain all five selected metrics for each of the four groups, with the page-row denominator given above each table. Percentages are Ansvra calculations from the published fractions, rounded to four decimal places. The final row in each table is rendered prevalence minus raw prevalence, expressed in percentage points. [1]
Desktop homepages
Each percentage uses 12,155,025 page rows as its denominator.
Metric | Value |
|---|---|
Raw HTML detection | 10.2415% |
Rendered HTML detection | 10.4315% |
HTTP-header detection | 0.0757% |
Rendered without raw | 0.1956% |
Recorded-value change | 0.2315% |
Net rendered minus raw | 0.1901 pp |
Desktop secondary pages
Each percentage uses 10,834,572 page rows as its denominator.
Metric | Value |
|---|---|
Raw HTML detection | 10.0920% |
Rendered HTML detection | 10.3813% |
HTTP-header detection | 0.0799% |
Rendered without raw | 0.2938% |
Recorded-value change | 0.3211% |
Net rendered minus raw | 0.2893 pp |
Mobile homepages
Each percentage uses 15,425,851 page rows as its denominator.
Metric | Value |
|---|---|
Raw HTML detection | 9.9992% |
Rendered HTML detection | 10.1996% |
HTTP-header detection | 0.0683% |
Rendered without raw | 0.2072% |
Recorded-value change | 0.2429% |
Net rendered minus raw | 0.2004 pp |
Mobile secondary pages
Each percentage uses 13,192,188 page rows as its denominator.
Metric | Value |
|---|---|
Raw HTML detection | 9.7049% |
Rendered HTML detection | 9.9922% |
HTTP-header detection | 0.0731% |
Rendered without raw | 0.2948% |
Recorded-value change | 0.3235% |
Net rendered minus raw | 0.2873 pp |
On desktop homepages, raw detection is 10.2415% and rendered detection is 10.4315%. Their net difference is 0.1901 percentage points. The separately recorded rendered-without-raw rate is 0.1956%. These are different calculations. The first compares the prevalence of two flags across the group, while the second counts pages meeting a specific combination of flags. [1]
For a page review, use the benchmark to frame a question about where annotations appear. To explain a particular change, inspect that page’s actual declarations and targets.
A note about the published figure
The chapter’s figure description uses implementation labels that conflict with its surrounding prose. The initial linked result grid also has raw and rendered rates that differ from the headline percentages in that prose. This article uses a new analysis of the first four query-result rows and their named fields. The numerical comparison can be checked against those rows and the query definition. [1]
Open “View data” and “View query” under figure 5.57 in the source chapter to inspect the original files. The grid includes additional presentation sections beyond the initial query results. Our scope is the four rows and five fields identified above. [1]
Compare the actual annotations on your pages
For an HTML implementation, record the hreflang values and target URLs from the raw response and the rendered page. Keep the page URL, time and relevant browser conditions with the comparison. If a preference or session state affects what the page returns, document that condition too.
Read each value and target as a pair. A list of language codes can look stable even when a target points to the wrong localized page. An annotation added during rendering also needs to name the intended version. The purpose of the comparison is to find the exact declaration that changed and decide whether the change was expected.
HTTP Archive’s change flag compares recorded raw and rendered value arrays after sorting them. Its presence checks test whether the relevant arrays contain values. A full site review goes further by checking the target URLs and their relationships. [1]
If the declarations differ, capture both versions before proposing a fix. Ask which template or localization rule owns the final annotation. A useful handoff includes the affected page, the expected language and target pair, the two observed versions and the intended implementation method.
Check the alternate relationships
Google’s guidance says each language version should list itself and its alternates, with fully qualified URLs. It also requires reciprocal links between the relevant pages. Use those rules to check the chosen group of localized pages together. [2]
The following fictional mapping illustrates a two-version guide. It shows the declarations an editor would expect to inspect, without representing measured results.
Fictional page being checked | Expected alternate declarations |
|---|---|
https://example.com/en/guide | en → https://example.com/en/guide |
https://example.com/fr/guide | en → https://example.com/en/guide |
Check the corresponding declarations on each page, using the method the site actually maintains. Confirm that the target contains the intended localized content. If an alternate has moved or a version was removed, update the intended mapping before reviewing its implementation.
Keep annotation and content questions in the same review. For example, a guide may have a correct alternate relationship while its translated instructions need an editorial update. Recording both observations helps assign the right work to the right owner.
Ansvra’s suggested workflow from intended alternates through implementation checks to saved answer evidence.
- Map alternates
Choose a real group of localized URLs. Record which versions are intended to be alternates before reviewing their implementation. [2]
Next: Choose method
- Choose method
Identify HTML, HTTP headers or a sitemap as the implementation method. Google treats these as equivalent choices. [2]
Next: Compare pages
- Compare pages
For an HTML implementation, compare recorded raw and rendered annotations under documented conditions. The benchmark flags describe detection and value changes, not a full validation. [1]
Next: Check links
- Check links
Inspect language codes and target URLs, self-references and return links for the chosen page group. Keep Google’s requirements specific to Google Search. [2]
Next: Review answers
- Review answers
Read the relevant saved answers and cited localized URLs. Keep citation evidence distinct from annotation changes. [1] [2]
This is an original review workflow. HTTP Archive supports the benchmark definitions. Google supports its implementation guidance. [1] [2]
Review recorded access and cited URLs
When the review also concerns page access, examine the recorded activity and responses for the URLs in your chosen group. Ansvra’s AI Agent Analyzer helps teams with supported connected sites inspect recorded crawler activity, pages and responses. Carry those observations into the review with the dates they cover. AI Agent Analyzer
For a citation question, open the relevant saved answer and record the localized URL it references. Read the answer alongside the page. Ask whether the language, facts and level of detail fit the customer question you are investigating.
Citation Source Analysis helps teams inspect available references and recurring cited pages in saved answers. Keep the question and answer context next to the cited URL so the content team can explain why that page deserves attention. Citation Source Analysis
Suppose an answer references the English guide for a question your team reviews in French. Record the answer and source first. Then inspect the French guide, its intended relationship to the English version and the implementation evidence. This creates a specific investigation. A change in an annotation and a change in a saved answer each need their own recorded evidence.
Close the review with a repeatable check
Assign a concrete action, such as correcting a target URL in the localization template or updating the content on one guide. Include the affected page group, the expected result and the evidence needed to verify the action. Save the before-and-after declarations under comparable conditions.
After the change, inspect the chosen implementation again across the page group. If citations were part of the original question, review the relevant saved answers separately. Keep the conditions and dates visible so a later reviewer can understand what changed and which questions remain open.
Keep the intended mapping with the content brief when a page group gains another language version. That lets the next editor check the new version against the existing relationships and identify which template controls the declarations. If the review finds an unexpected target, preserve the captured evidence before changing the mapping. A documented expectation helps the team distinguish a planned localization update from an accidental change during rendering.
Method and limits
The analysis retains 20 published fractions from four initial result rows. It uses each row’s own page count, converts fractions to percentages and calculates four net differences. The figures describe HTTP Archive’s July 2025 crawl. The source grid gives fractions and page-row denominators, with no numerator counts for these fields. [1]
Detection rates and recorded-value changes provide implementation context. They do not establish correct alternate relationships or measure citation outcomes. Google’s documentation supplies the implementation guidance used here, including its distinction between localized-alternate declarations and its own language-detection process. [1] [2]
Summary
Identify your hreflang method, map the intended localized versions and compare the relevant declarations under consistent conditions. Check the targets and return relationships before assigning a technical fix.
Use recorded page activity and saved answers to investigate the access or citation question, with each observation documented on its own terms.
Discussion
Join the discussion