Most of what AI cites about your brand lives on sites you do not control

Ask an AI assistant to recommend a company in your category, and most of the sources it uses to describe you will be pages you did not publish. Muck Rack's analysis of more than 25 million links cited by ChatGPT, Claude, and Gemini found that earned media accounts for 84 percent of AI citations, while paid and advertorial content made up 0.3 percent. Your homepage still matters, but it is rarely where the citation comes from.

The Total Graph Authority framework names this layer third-party validation, and our post on first-party authority covers the domain you own. Third-party authority works as a set of trust tiers. Each tier feeds AI systems in a different way and at a different point in the buying process, and the order in which you build them decides how fast the results add up.

Why AI cites other people's sites more than your own

AI assistants ground their answers in retrieved pages, and the retrieval step weighs how credible and independent each source looks as well as how relevant it is. A page you published about yourself is promotional by definition, however well it is written. A page someone else published about you can pass that test, which is why assistants reach for third parties when they describe a brand.

Several studies point the same way. When Stacker and Scrunch tracked 87 stories across eight AI platforms, stories distributed through third-party news outlets saw a median 239 percent increase in AI citations compared with brand-owned content alone. In the same study, 97 percent of distributed stories earned at least one AI citation, against 82 percent for owned content. Ahrefs found across 75,000 brands that branded web mentions correlate with AI Overview visibility at 0.664, against 0.218 for backlinks, which is roughly three times stronger.

Breadth matters too. 5W's 2026 citation research associates brands present on Wikidata, Wikipedia where notability allows, and four or more authoritative third-party platforms with 2.8 times higher citation likelihood. A model trusts a claim more when it finds the same claim in several places it does not think you control. Building that agreement across sources, so the model keeps finding the same brand described the same way, is the work of this layer.

The six tiers of third-party trust

Third-party sources do not carry equal weight. Editorial independence, the structured data attached to a source, and the density of user consensus all affect how much an AI system relies on it. Knowing where a source sits tells you where to spend first. A common mistake is pouring effort into Reddit threads and listicles while the entity foundation underneath them is still empty.

Tier 1 is less a citation source than the entity infrastructure the other tiers are credited to. Tiers 2 to 6 are where most direct citations happen, and each performs better when there is a solid entity record beneath it. Building them in that order is what makes the signals accumulate in one place.

Build the entity graph first

Before an AI system decides whether your content is worth citing, it has to work out which entity the content belongs to. If the system cannot tell who published a page, even an excellent page may not get cited. Brands skip this step more than any other, and it limits everything built on top of it.

Wikipedia plays a large part. Profound's analysis of AI citations found that Wikipedia accounted for 7.8 percent of all ChatGPT citations, more than any other source on that platform. When a model can check a brand claim against an independent entry, it has a reference point. Without one, it tends to hedge or leave the brand out. Wikipedia has strict notability requirements and most brands do not qualify, and you should never write or edit your own article.

Wikidata has no notability bar for verifiable organizations, and it feeds knowledge panels in Google and Bing and the entity systems AI search relies on. When the London School of Economics library added its theses to Wikidata and linked them from Wikipedia, the number of users referred to its theses site from Wikipedia rose 250 percent between 2019/20 and 2021/22. At minimum, a brand's Wikidata item should state that it is an organization and list the official website, founding date, headquarters, logo, and LinkedIn and X identifiers.

The sameAs property in your Organization schema connects these records. A Schema App test that linked pages to their entities through sameAs references to Wikipedia, Wikidata, and Google's Knowledge Graph recorded a 46 percent increase in impressions and a 42 percent increase in non-branded clicks over 85 days. The property tells a model that the brand on your site, on LinkedIn, on Crunchbase, and on Wikidata is one company, so signals from each are credited to a single entity.

{
  "@context": "https://schema.org",
  "@type": "Organization",
  "@id": "https://yourbrand.com/#org",
  "name": "Your Brand Name",
  "url": "https://yourbrand.com",
  "sameAs": [
    "https://en.wikipedia.org/wiki/Your_Brand",
    "https://www.wikidata.org/wiki/Q######",
    "https://www.linkedin.com/company/your-brand",
    "https://www.crunchbase.com/organization/your-brand"
  ],
  "knowsAbout": ["your primary category", "key use case 1", "key use case 2"]
}

Earn mentions on the sources AI retrieves

With the entity record in place, the next step is presence on the tiers AI systems read when they answer buying questions. Each tier needs a different approach.

Review platforms

When a buyer asks an assistant which software to choose, the retrieval layer reads review platforms. SE Ranking's analysis of AI Overviews found that five platforms account for 88 percent of review-platform citations, led by Gartner Peer Insights at 26 percent, G2 at 23.1 percent, Capterra at 17.8 percent, Software Advice at 12.8 percent, and TrustRadius at 8.3 percent. G2 announced in January 2026 that it would acquire Capterra, Software Advice, and GetApp from Gartner, and Gartner reported completing the sale in February 2026. Omniscient Digital projected the combined group at 3.68 percent of bottom-of-funnel AI citations, second only to Reddit, and 12.69 percent for proof and evidence queries. Those are modeled projections, not post-acquisition data.

Having a profile at all makes the biggest difference. Seer Interactive's study of more than 800,000 AI responses found that brands with no Trustpilot profile had a median citation rate of 1 percent, against 53.5 percent for brands with even a minimal profile of a handful of reviews. Moving to an active profile added about 25 points more. Seer notes the findings show correlation, not causation.

AI reads a profile differently from a person skimming it. It parses the rating, review count, category tags, feature attributes, and the wording of individual reviews. A description such as "software for agencies of 10 to 50 people managing client projects with profitability tracking" matches real questions far better than "innovative next-generation project management". Fill every attribute field, including integrations, deployment, company size, and pricing model. Encourage reviews that describe outcomes, because "we cut onboarding from 14 days to 3 with the automated flow builder" gives a model a fact it can use and "great product" gives it nothing. Choose the platform your buyers use, such as G2 for B2B software or Trustpilot for consumer brands.

Earned media and analyst coverage

Earned media closes the gap between being found and being trusted. Forrester's State of Business Buying 2026, based on nearly 18,000 global business buyers, found that 94 percent use AI during their buying process. Coverage of the same research reported that 20 percent felt less confident after meeting unreliable or inaccurate AI information. Buyers discover options through AI and then check them against sources they trust. A brand that appears in the AI answer with no editorial footprint passes the first step and stalls at the second.

Analyst coverage sits at the high-trust end of this tier. Gartner says a vendor's status as a client does not affect inclusion or positioning in its Magic Quadrant, and Forrester's integrity policy says clients cannot buy coverage. Analysts select vendors against inclusion criteria, so you cannot submit yourself, but you can make yourself easy to include. Give one senior person ownership of analyst relations for a few hours each quarter, brief analysts when you meet the criteria, respond to research requests quickly, and keep a one-page vendor data sheet current. A line in a market guide or a quote in a trade article about your category also counts.

Press releases have become more visible in AI answers. Muck Rack reported in December 2025 that press release citations grew fivefold since July 2025, and a later summary of the same data put the share at roughly 1.2 to 6 percent of AI citations over that period. The releases that got cited contained twice as many statistics, 30 percent more action verbs, and a 30 percent higher rate of objective sentences. Write every sentence so it contains an entity, a verifiable number, or a named external source. For Indonesian brands, our post on multilingual GEO in Bahasa Indonesia covers how this works across languages.

Reddit and community threads

Reddit holds a privileged position in AI citations. Google signed a content licensing deal with Reddit reportedly worth about 60 million US dollars a year, according to sources who spoke to Reuters, and OpenAI signed its own agreement. Ahrefs' tracking of Google AI Overviews ranked reddit.com as the second most-cited domain in September 2026, with an 18.5 percent share behind YouTube.

Reddit threads match how buyers ask questions. When someone asks whether a tool suits a ten-person team, the model breaks that into questions about team size, pricing, limits, and daily use, and Reddit threads answer those questions in ordinary language. Product recommendation threads, comparisons, and annual megathreads are the most useful places to contribute.

A useful comment gives a direct answer, then specific evidence, then an honest note on the trade-off or who the product does not suit, and discloses any affiliation. That reads as credible, and models keep experience-based detail while dropping marketing language. Avoid what moderators already punish, such as vote manipulation, hidden affiliations, promotion-only accounts, and copied answers. Reddit's old self-promotion guidance suggested that 10 percent or less of your activity should point to your own content, and each subreddit now sets its own rules.

Comparison pages and directories

Many buyer questions are comparative, in the form of best X, X versus Y, or alternatives to Z. You cannot credibly publish a neutral top ten that puts you first, so the task is to earn a place in the comparison pages AI already cites. The pages that get cited share a structure, with a recommendation near the top, a section per product with real specifications, honest pros and cons, and an HTML comparison table.

Earning inclusion is outreach work. Send listicle publishers in your category an honest comparison that includes your competitors, with verified specifications and pricing they can use directly. When you contribute to a best-of piece, keep the rows you lose, because neutral framing is what gets it published. Target the comparison pages that AI assistants currently cite for your category queries, which are not always the ones with the most traffic.

Category directories feed narrower citation surfaces. Crunchbase holds company and funding data that strengthens the entity graph, Product Hunt appears for new tool queries, Clutch for agency and service queries, and sites like AlternativeTo for people looking to switch.

Podcasts and video transcripts

AI systems work from transcripts and metadata, so audio and video only count once they become text. Goodie's study of AI citations found that YouTube's share of social-type citations rose from 18.9 percent in August 2025 to 39.2 percent in December 2025, with December covering data through the 15th. Podcasts lose most of their citation value unless each episode ships with show notes, chapters, and a published transcript. Most podcasters and guests never publish one, so doing it gives you an advantage.

The text around the media is what gets cited. YouTube titles and descriptions are processed before any footage, and a linked transcript gets indexed. A description that works reads like an answer, with a question-shaped title such as whether a tool is worth it for a ten-person agency, a summary of results with real numbers, a line on who it suits, and timestamps that mark the key result and the comparison.

Keep every source consistent

When sources agree on the same brand and the same facts, the model reaches the same answer whichever path it takes through them. When they disagree, it may skip citing you instead of repeating a conflict it cannot resolve. If your G2 profile lists different pricing from your website, or Reddit threads describe features your press coverage never mentions, the model's confidence in you drops.

Three checks keep the story consistent. Your brand name, canonical URL, and company description should be identical everywhere, and any rename should reach every surface at once. Your category language should match how buyers search, because if buyers type "freight management software" and every profile says "logistics platform", the model may never connect the two. Your facts, including pricing ranges, customer counts, coverage, and named integrations, should match across every profile.

Run these checks before any outreach campaign. Clean, consistent records let new mentions add to the same entity, while inconsistent records turn each new mention into more noise.

Each AI engine pulls from a different set of sources

AI engines draw on visibly different pools of sources, so third-party work should be weighted toward the engines your buyers use. The exact shares move from month to month, but the broad pattern in the published data is stable enough to plan with.

Bing deserves attention beyond Copilot. Microsoft announced in 2023 that Bing would be ChatGPT's default search experience, and OpenAI describes ChatGPT search as using third-party search providers alongside its own crawler and partners. Getting pages cleanly indexed in Bing Webmaster Tools widens your reach, and our guide to setting up the Bing Webmaster MCP makes that monitoring easier. Public data on where Claude, Gemini, and Grok source citations is thinner, so track those engines with your own prompt runs.

What a live check across four assistants showed

On 15 July 2026 we ran one prompt across four assistants and logged what each cited and recommended. The prompt asked for the best agency to get a brand cited by AI search and ChatGPT in Indonesia, the kind of question a business shopping for GEO help would type. The four assistants used four different sets of sources and reached two different conclusions about whether our brand existed.

Google AI Mode built its answer mostly from indexed pages on the agencies' own domains, and it cited our site and the Total Graph Authority post. ChatGPT, running a live web search, returned a short list that opened with Search Agency, mixed with local marketing agency profiles and a few news and directory pages. On both, a brand with a strong domain surfaced on that strength.

Perplexity leaned on third-party roundups, Indonesian media and blog listicles ranking local agencies, a Clutch ranking, and a few YouTube videos, and it named the agencies that recur across those lists. Gemini named another shortlist drawn from aggregated agency comparisons. Neither cited our domain. A strong owned domain earned citations from the two assistants that read it, and only third-party presence would have earned the other two.

The order to build third-party authority

Spend the first two weeks on the entity graph. Create or verify the Wikidata item, check any existing Wikipedia article for accuracy without editing it yourself, add sameAs schema linking your site, LinkedIn, Crunchbase, Wikidata, and Wikipedia, and make the brand name, category description, and founding details identical everywhere.

Use weeks three to eight for the review layer. Pick one primary platform based on where your buyers go, complete the profile with question-answering descriptions and every attribute filled, ask satisfied customers for reviews that describe specific outcomes, and reply to all of them before adding a second platform.

After that, move to a steady rhythm. Aim for one citable editorial placement each quarter by building a data asset nobody else has, such as an original study or a benchmark, and pitch it to the outlets AI already cites for your category. Take part in the three to five subreddits where your buyers ask questions, as a genuine expert and not in campaign bursts. Around month three, work on inclusion in the comparison pages AI cites and claim the relevant directory profiles. By month four, publish a transcript for every podcast appearance and give your videos question-shaped titles and timestamped descriptions.

Third-party authority is not always the right place to start. If AI crawlers cannot read your website, or your domain has no clear page explaining what you do, mentions elsewhere point toward a brand the models still cannot resolve. Our free Agent View extension shows how much of your site AI crawlers can read. A local service business with no review-platform category will get little from G2, and an early-stage company with no data or distinctive view is better off creating something worth citing before trying to distribute it.

How to measure third-party authority

Ranking positions and organic sessions do not show whether this layer is working. Track how often AI answers cite an external source that names you, the share of your citations that come from third-party sources compared with your own domain, your review count and rating against competitors, editorial placements from outlets AI retrieves, and traffic arriving from AI assistants.

Run your prompt set weekly across branded, category, comparison, problem, and eligibility questions on at least ChatGPT, Gemini, Perplexity, and Google AI Mode. Log the cited sources as well as whether your name appeared. The share of third-party sources among those citations is the clearest single number for this layer, and the live check above is the same method run once.

Third-party authority for Indonesian brands

For Indonesian and Southeast Asian brands, the third-party layer often matters more because the gap starts wider. Multinationals competing in these markets usually arrive with Wikipedia entries, Wikidata items, Crunchbase profiles, and review listings built over years. Local competitors often have none of these, so models have far more independent material about the foreign brand than the local one. The gap is in entities and reviews more than in content.

Closing it means building the same records on the surfaces Indonesian buyers and AI systems reach, in Bahasa Indonesia as well as English. That includes coverage in major Indonesian news outlets such as Kompas, Detik, and Bisnis, participation in communities like Kaskus where candid user opinions play the role Reddit plays elsewhere, and complete Wikidata, Crunchbase, and LinkedIn records. The live check above ran on this market and showed the split clearly, with the same brand visible on assistants that reward its own domain and missing from those that rely on third-party roundups.

If you want to see where your brand stands across these tiers and which gaps to close first, request an audit and we will map your third-party footprint against the sources AI engines cite in your category.