Home / Research / India AI Citation Study / Classification rules
Domain classification rules
Frozen 22 August 2026 before analysis. Revised 24 August 2026 after a 100-row sample review found three error classes. Both the original rules and the corrections are published here.
Why there are two axes
A single owner axis hides the most actionable pattern in the data. Domains are coded twice. Axis A is the owner, who controls the domain. Axis B is the page type, whether the cited page is editorial content or a product page.
Axis B matters because a brand cited for a blog post and a brand cited for a product page are different outcomes with different implications for anyone trying to earn a citation.
Axis A, owner buckets
Brand-owned (462 citations). A company’s own domain, where that company sells the product class in question.
Google-owned (357). google.com Shopping links and youtube.com.
Publisher, medical or institutional (156). Medical authorities, hospitals, universities, government health bodies.
Marketplaces (248 across four types): pharmacy (1mg, PharmEasy, Apollo), horizontal (Amazon, Flipkart, Myntra), specialist (Nykaa, Purplle, Tira) and nutrition (HealthKart, Nutrabay).
Social platform (62). Instagram, Pinterest, Facebook, Medium.
Publisher, media (35). Independent editorial not owned by a seller.
UGC forum (27). Reddit and Quora, reported separately rather than pooled.
Clinic or practitioner (21), Reference (10, Wikipedia), Brand-owned media property (8, publisher-styled sites owned by a brand such as skincare.com and bebeautiful.in) and Off-topic corporate content (5).
Decisions made, and why
YouTube is its own bucket, not merged with google.com. Shopping links are a SERP feature with no publisher and no way for a third party to earn one. YouTube videos have creators and are earnable by anyone. Merging them conflates opposites; merging YouTube with Reddit and Instagram instead would lose the fact that Google owns it.
Known limitation: exported YouTube URLs are bare watch IDs, so brand channel versus independent creator cannot be distinguished. No claim is made about who wins in video.
International brands’ Indian sites count as international. cetaphil.in, nivea.in, garnier.in, centrumshop.in, dove-india.com. Parent company decides. The alternative would inflate the Indian-brand share with subsidiaries of multinationals.
Marketplaces are split four ways rather than pooled. Buckets can be merged after the fact, not un-merged. Stated plainly: the split did not produce a headline. Shares are broadly balanced within categories and it is reported for completeness.
Nationality is hand-verified per domain, never derived from the TLD. giva.co is Indian and icecartel.com is American; both are generic TLDs. Indian D2C overwhelmingly uses .com, which is why a ccTLD cut fails in both directions.
Axis B, page type
Three values. editorial is a content page: blog post, article, guide, advice. product / other is a product page, category page, homepage or search results page. not applicable is platform content, where the distinction does not exist.
The rule, applied in this order. One: an explicit editorial marker in the path makes it editorial. Markers include /blog/, /blogs/, /article/, /guide/, /journal/, /tips/, /advice/, /learn/, /insights/, /tutorials/, plus vendor-specific equivalents such as /skin-smart/, /health-library/, /everyday-care/ and /cancerwise/. Two: otherwise a commerce marker makes it product/other. Markers include /collections/, /products/, /shop/, /category/, /c/, /otc/, /q/, /search/, /buy/. Three: otherwise, for publishers and clinics only, it is editorial, because they do not sell anything. Four: otherwise, for sellers, it is product/other.
Editorial markers are checked before commerce markers, so a path like /blogs/category/how-to-apply-sunscreen reads as editorial rather than being captured by the word “category”.
Page type is not applicable for platform content — YouTube, Instagram, Reddit, Quora, Wikipedia, Facebook. 281 citations, 20% of the dataset. google.com stays product/other, because a Shopping link genuinely is a product listing.
Corrections made after the sample review
A 100-row random sample was drawn and reviewed against the coded output. It surfaced three distinct error classes. All three were corrected and the full dataset recoded.
One. Publisher articles were mis-coded as product pages. The original rule required an explicit editorial marker, then fell back to a long-hyphenated-slug test. Publishers use flat URLs without markers, so healthline.com/nutrition/best-multivitamins-for-men and webmd.com/diet/protein-shakes were both coded product/other. The rule assumed editorial content announces itself in the path; publishers have no reason to do that, because everything they publish is editorial. Fixed by rule three above. Publisher editorial share moved from 17% to 97%.
A failed intermediate attempt, recorded because it matters: the first fix applied the flat-slug test to every domain. That called flipkart.com/q/moisturizer-for-oily-skin (a search page) and 1mg.com/otc/centrum-men-vegetarian-tablets (a product page) editorial, and pushed marketplace editorial from 23% to 38%. Marketplaces name product pages with long hyphenated slugs exactly as publishers name articles, so the test cannot separate them. Restricting it to non-sellers is what made it safe.
Two. Platform content was forced into a distinction that does not apply to it. YouTube videos, Instagram reels and Reddit threads were coded product/other. Neither value is right. A Reddit thread is not a product page, and calling it editorial implies a contrast with a product page that does not exist. 281 citations are now marked not applicable. This is the more serious of the two page-type errors, because a wrong-but-plausible value survives review in a way a blank does not.
Three. The Indian / international brand flag was incomplete. The original flag was a hand-built list of recognised international brands and it missed several, including icecartel.com (US), blueland.com (US), nathanalanjewelers.com (US), stockholm.com.hk (Hong Kong), queensmith.co.uk (UK), onecklace.com (Israel) and alpinloacker.com (Austria). This one could not be fixed by rule, because brand nationality is not derivable from a domain. Every brand-owned domain was reviewed by hand against the parent-company rule: 52 of 163 are now flagged international, up from 20. One correction ran the other way, denverformen.com was assumed foreign on the strength of its name and is in fact Indian. Effect on published results: none, because every reclassified domain sits below the reporting threshold.
Verification status
100-row random sample reviewed. Three error classes found, all corrected.
Corrected page-type rule re-tested against 16 known-answer URLs: 16 correct.
Brand nationality hand-verified domain by domain, 163 domains.
m.youtube.com merged into youtube.com. The raw export does not normalise it.
Exact-match denominator excludes Google Shopping. 0 of 175 google.com citations are Exact URL, by construction, while 95 of 182 youtube.com citations are. The two must never be pooled for rank-match cuts.
A residual error rate was not counted. The sample was used diagnostically, to find error classes rather than to estimate a rate, and the classes it found were fixed rather than reported. A rate measured against the corrected coding would be a reasonable addition to edition two.
What this taxonomy cannot answer
Who owns a cited YouTube video. The channel is not resolvable from the export.
Whether a brand-owned media property is editorially independent.
Whether a clinic site is a practice or a product business in disguise.
Edge cases where a publisher uses a single-word slug, for example examine.com/supplements/creatine/, which reads as product/other under rule two.