How Should I Classify Owned, Earned, Neutral, and Competitor Citations?

What counts as a citation?

A citation is a source the assistant links to or lists with its answer, inline or in a sources panel. It is different from a mention, where a brand is named in the answer text. The two often diverge. A Semrush and Growth Memo study found that in 61.7% of domain appearances, a site was cited without its brand being named in the answer. Classify citations and mentions separately, and record both.

What are the classes?

Decide "whom it mentions" from the cited page's text, not the answer's, by checking the page for your brand and your tracked competitors' names.

Which edge cases need rules?

• Your press release on a newswire or syndication site: owned-syndicated, not earned. You wrote it.
• Guest articles written by your staff: owned-contributed.
• Profiles you manage on review platforms, marketplaces and directories: managed profiles. The platform controls the page, but you supply much of the content, and customers add reviews. If reviews dominate the page, you can treat it as earned; apply the rule consistently.
• Customer reviews and community threads: earned, sub-labelled "community".
• Partner pages that list you: earned, sub-labelled "partner", since the relationship is commercial.
• Affiliate reviews with disclosed commissions: paid, or earned-affiliate if you want to track them within earned.
• Pages that mention you negatively: still earned, with a stance flag. Earned is about who wrote it, not whether it is kind.
• A competitor's page that mentions you: competitor, since the competitor controls the framing. How to judge such pages is covered in how to check whether a competitor-controlled comparison page is treated as neutral.

Write the rules down, with examples, and apply them consistently. The labels matter less than using the same ones every month.

What is Source-type mix in AI search citations, and why does it matter for my business?

Source-type mix is the share of citations in your category's AI answers that falls into each class. It matters because it tells you which lever moves answers:

• Owned-heavy mix: assistants rely on brands' own sites. Your content and site clarity are the main levers.
• Earned- and gap-heavy mix: assistants rely on reviews, lists and coverage. PR, reviews and list inclusion matter most, and gaps show exactly where to start.
• Competitor-heavy mix: competitors' pages are framing the category. Your own comparison content and third-party corrections are priorities.
• Neutral-heavy mix: reference sources define the topic. Being cited alongside them for your specific expertise is the realistic goal.
• Paid share: usually small. Muck Rack's analysis of more than 25 million links in AI answers found paid and advertorial content made up about 0.3% of citations.

Mixes differ a lot by category and question type. Muck Rack found earned media, broadly defined, made up 84% of citations overall. For local businesses, BrightLocal found about 93% of the unique cited domains were business websites, while Google Business Profiles and directories supplied many of the individual citations. An audit of mental health answers found government, commercial health and academic sources at roughly 22% each. So measure your own mix, rather than relying on a general figure.

What is an earned citation in AI search, and how do you measure it?

An earned citation is a source cited in an AI answer that a third party controls, that you did not pay for, and that mentions your brand. It is evidence from someone else that the assistant used when answering.

Measure it several ways:

• Earned citation rate: the share of answers that cite at least one earned source mentioning you.
• Earned share: earned citations mentioning you, divided by all citations in your tracked answers.
• Earned versus gap ratio: earned citations against gap citations. A low ratio means third-party sources cite your competitors more than you.
• Earned stance: the share of earned citations that are positive, neutral or negative about you.
• Top earned domains: which third-party sites cite you most, and which gap domains cite competitors most.
• Named rate: of answers with an earned citation mentioning you, how many also name you in the text.

Track these per assistant and per question type. Recommendation questions and "how to" questions usually have very different mixes.

How do you classify at scale?

1. Keep a domain list of your owned domains and each competitor's domains, including subdomains and old domains.
2. Fetch each cited page and check its text for your brand and competitors' names, including aliases.
3. Flag paid content by looking for sponsorship labels and affiliate links.
4. Apply the rules automatically where the domain list decides it, and use a reviewer or a rubric-driven AI classifier for the rest.
5. Audit a sample by hand each month. How to check an AI classifier against people is covered in human evaluation versus LLM-as-a-judge for measuring AI recommendations.

What does a monthly report look like?

Show the sample under the table, for example "based on 420 citations across 150 answers this month, 390 last month". At that size, changes under about seven points are within normal month-to-month variation, so none of these moves is yet a trend. The gap row is the action list: the lists and articles to pursue next. Whether competitor names belong in tracked prompts is covered in whether competitor names should be allowed in prompts for fair category visibility reporting.

What do people ask about classifying citations?

Should review-platform profiles count as owned or earned?

Pick one rule and keep it. If customer reviews make up most of the page, you can count it as earned and track it as a separate sub-label.

Is a negative earned citation still earned?

Yes. Record stance separately. The class describes who controls the source, not whether it helps you.

How many answers do I need for a reliable mix?

A few hundred citations per assistant is a workable minimum, but even then a class's share can move five to seven points between months by chance, and more when a handful of answers supply most of the citations. Show the number of citations in each report and treat smaller changes as noise unless they hold for two or three months.