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 watch where the answer comes from. Most of the sources it pulls to describe you are not yours. Muck Rack's running analysis of roughly 25 million citations across ChatGPT, Gemini and Claude keeps landing around 84 percent of AI citations pointing at earned media and third-party pages rather than anything the brand published on its own domain. Your homepage still matters, but it is rarely where the citation comes from.

The Total Graph Authority post named that layer, third-party validation, as the one cited more often than a brand's own domain, and left four tactics on the page. The first-party post took apart the domain you own. This one takes apart the layer you do not, because third-party authority is not a single channel you switch on. It is a hierarchy of trust tiers, each feeding different AI systems in different ways at different points in the buyer journey, and the order you build them in decides how fast anything compounds.

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

The reason starts inside the model. Large language models ground their answers in retrieved data rather than generating from memory, and the retrieval step does not just look for relevant text, it weighs how credible and how independent the source is. Your own domain fails that test by definition, since a page you published about yourself is promotional no matter how well written it is. A page someone else published about you can pass it, which is why the retrieval layer reaches for third parties first.

The numbers around this are consistent enough to plan against. A controlled study from Stacker and Scrunch found that identical content, once distributed through third-party news outlets, produced a 239 percent median lift in AI search visibility compared with brand-owned distribution alone, with the brand's citation rate climbing from 82 to 97 percent across the platforms tested. When Ahrefs looked at 75,000 brands, plain web mentions correlated with AI Overview visibility at 0.664 against 0.218 for backlinks, roughly three times stronger, which is a quiet reordering of the thing most SEO budgets were built around. Even the platform-count effect points the same way, with 5W's citation research reporting that brands present on four or more third-party platforms are associated with 2.8 times the ChatGPT citation rate of single-platform brands.

Underneath all of it is corroboration. A model trusts a claim more when the same claim shows up independently in several places it does not think you control. Engineering that agreement across sources, so the model keeps arriving at the same brand described the same way, is the whole job of this layer.

The six tiers of third-party trust

Not every third-party source carries the same weight. AI systems apply a rough hierarchy based on editorial independence, the quality of the structured data attached, and how dense the user consensus is. Knowing where a source sits tells you where to spend first, and the most common mistake we see in audits is a brand pouring effort into Reddit threads and listicles while the foundation underneath them is still empty.

Tier 1 is not really a citation source in the ordinary sense, it is the entity infrastructure everything else gets credited to. Tiers 2 through 6 are where the direct citations happen. Every one of them performs better when there is a solid entity foundation beneath it, so the sequence is not optional, it is the difference between signals that accumulate and signals that scatter.

The entity graph comes before everything

Before an AI system decides whether your content is worth citing, it decides whether it understands what entity the content belongs to. Entity confidence sits upstream of content quality in the retrieval pipeline, so the most data-rich page on the web will not get cited if the system cannot confidently work out who published it. This is the piece brands skip most often, and it is the one that quietly caps everything above it.

Wikipedia does a lot of the heavy lifting here. Profound's study of hundreds of millions of citations put it at 7.8 percent of all ChatGPT citations, sitting as the top or second most-cited domain across nearly every major model. The deeper mechanism is anchoring. When a model can cross-reference a brand claim against a Wikipedia entry, it has something independent to check itself against, and when it cannot, it tends to hedge or leave you out.

The problem with Wikipedia is that it has real notability requirements, and most brands do not qualify. Wikidata does not have that bar. Any brand with verifiable identifying information can create a Wikidata item regardless of Wikipedia notability, and Wikidata feeds the knowledge panels in Google and Bing and pipes straight into the entity systems AI search relies on. After the London School of Economics library added its theses to Wikidata, downloads rose 47 percent over the following six months, a before-and-after jump from the Wikidata records alone with no Wikipedia article involved. At minimum, set the item to instance-of organization, add the official website, inception date, headquarters, logo, and the LinkedIn and X identifiers, so the record is unambiguous.

The connective tissue is the sameAs array, and it does more work than its size suggests. A Schema App study that linked pages to their underlying entities, partly through sameAs, recorded a 46 percent lift in impressions and 42 percent lift in clicks on non-branded queries over 85 days. The property acts as an entity canonical, telling a model that the brand on your site, on LinkedIn, on Crunchbase and on Wikipedia is one company, so signals from all of them get credited to a single entity instead of splitting into four weaker ones.

{
  "@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"]
}

Review platforms answer the buying questions

When a buyer asks an assistant which software to pick, the retrieval layer does not reason it out from first principles, it goes to review platforms and reads what is already structured there. SE Ranking's analysis of AI Overviews found that five platforms account for 88 percent of all review-platform citations, led by Gartner Peer Insights at 26 percent, G2 at 23.1, Capterra at 17.8, Software Advice at 12.8, and TrustRadius at 8.3. That concentration is why a thin or missing review profile costs you the entire purchase-question surface.

G2's agreement to acquire Capterra, Software Advice and GetApp, expected to close in Q1 2026, would tighten this further. Modeling the combined ecosystem, Omniscient Digital projected it at 3.68 percent of all bottom-of-funnel AI citations, second only to Reddit, rising to 12.69 percent for proof-and-evidence queries about reviews and comparisons. Those are projections rather than post-acquisition data, but the direction is hard to argue with. And a Seer Interactive study of 800,000 AI responses found that brands with fully optimized review profiles were several times more likely to show up in AI answers than those with sparse ones.

What counts as an optimized profile is not what a human buyer would assume, because AI does not read the page the way a person skimming reviews does. It parses the structured parts, the star rating, the review volume, the category tags, the feature attributes, and the language inside individual reviews. A description that answers a real question, something like software for agencies of 10 to 50 people managing client projects with profitability tracking, gets matched to queries far better than innovative next-generation project management solution, because the model is looking for descriptions that fit the precise question it was handed. Filling every attribute field, integrations, deployment, company-size fit, pricing model, feeds the structured data the model pulls for comparison answers, and an incomplete profile simply loses those comparisons to a competitor with a complete one.

Volume matters less than most teams think. The returns come from recency, specificity, and use-case density rather than raw count. A review that says the team cut onboarding from 14 days to 3 using the automated flow builder hands the model an extractable fact, while great product, highly recommend hands it nothing. Below roughly 50 reviews there is not enough signal to work with, and past 100 the mentions climb noticeably. Which platform to prioritize depends on where your buyers go, so a B2B SaaS audience skewed to ChatGPT points to G2, a B2C audience to Trustpilot, a Google AI Mode audience to Gartner Peer Insights, and a Perplexity audience to the combination of G2 and Reddit.

Earned media is what buyers check before they trust you

There is a gap between finding you and trusting you, and earned media is what closes it. Forrester's State of Business Buying 2026, drawn from nearly 18,000 global B2B buyers, found that 94 percent of B2B buyers now use generative AI while purchasing, and 20 percent feel less confident because of it, thanks to answers that are incomplete or wrong. The result is a two-stage journey, AI-assisted discovery first, then human validation against trusted editorial second. A brand that surfaces in the AI answer but has no real editorial footprint clears the first stage and stalls in the second, because the buyer went looking for a publication that had written about you and found nothing.

This is where the Ahrefs correlation gets specific, since brand web mentions, the direct output of earned coverage, correlate 0.664 with AI visibility against 0.218 for backlinks. That roughly three-to-one gap is the argument for shifting budget toward being written about rather than only building links.

Analyst coverage is the high-trust corner of this tier, and you do not need a paid Magic Quadrant to benefit from it. Firms like Gartner and Forrester shape model recommendations by labeling vendors as Leaders or Strong Performers, a kind of independent validation no self-published page can imitate, and a single Magic Quadrant mention gets cited in enterprise answers for months. The accessible version is being mentioned in a category context at all, a line in a market guide, a quote in a trade piece about your space, an entry on a vendors-to-watch list. The budget-efficient path is to assign one senior person as analyst-relations owner for a few hours a quarter, respond to every research call for proposals quickly, keep a one-page vendor data sheet current, and prioritize the free vendor submissions like Forrester Now Tech, Gartner Critical Capabilities and IDC MarketScapes.

Press releases are worth a mention because their citation rate in AI grew fivefold in the back half of 2025, from 1.2 percent to 6 percent of all AI citations, but the releases that got cited were structurally different from the ones that did not. They carried nearly double the citations, more verbs, and more objective sentences, which is another way of saying AI cites releases that are genuinely newsworthy and data-rich and ignores the generic ones. The structural rule that holds up is that every sentence should contain either an entity reference, a verifiable number, or a named external validation, because a sentence with none of those is not citable regardless of how it reads.

Reddit answers the questions buyers actually type

Reddit sits in a structurally privileged spot in AI citations, and the licensing money confirms it. Google signed a data deal with Reddit reported at 60 million dollars a year, and OpenAI signed its own agreement on undisclosed terms, both buying access to the same real-time discussion. Perplexity leans on Reddit more than almost any other source, and Ahrefs ranked reddit.com as the second most-cited domain in Google AI Overviews in June 2026 with a 19.6 percent share.

The reason is query-format matching. When a buyer asks whether a tool is right for a 10-person team, the model decomposes that into sub-questions about team-size fit, pricing, real-world limits and day-to-day experience, and Reddit threads happen to answer exactly those sub-questions in ordinary conversation. What actually gets cited is worth being precise about, because comments get pulled far more often than original posts, medium-length comments in the 150 to 400 word range dominate while short declarative answers still get picked when they are precise, and high upvote counts are not required since the model weighs semantic relevance over popularity. Product recommendation and comparison threads, the X versus Y and best-tool-for-a-niche and annual megathread formats, are the highest-yield targets.

A comment built to be cited tends to follow the same shape, a direct answer to the question, then a measurable outcome or piece of specific evidence, then an honest note on the trade-off or the ideal-fit limit, then a disclosure of affiliation if the product is yours. That reads as credible rather than promotional, which is exactly the register the model wants to borrow when it synthesizes an answer, and marketing language gets dropped while experience-based specifics get kept. The things to avoid are the ones moderators already punish, vote manipulation, hidden affiliations, promo-only accounts, copy-pasted answers and engagement pods, and AI systems are starting to pattern-match that artificial activity as a negative signal too. Reddit's own 90/10 guideline, ninety percent genuine contribution to ten percent or less on-topic brand mention, is not just etiquette, it reflects how AI reads credibility, since an account that is mostly promotion looks like marketing and the model favors what looks human.

The comparison pages you are not allowed to publish

A large share of buyer queries are comparative, the best-X and X-versus-Y and alternatives-to-Z formats, and you cannot answer them with your own content because you cannot credibly publish a neutral top-ten that puts you first. What you can do is earn a place in the ones AI already cites. Bottom-of-funnel analysis keeps finding that comparison pages carry outsized citation weight at the decision stage, and the pages that win share a structure, an answer-first recommendation up top, a section per product with real specs, honest pros and cons, and an actual HTML comparison table. A page that admits where a rival is stronger gets published and cited, and one that pretends to win every row gets ignored.

Earning your way in is mostly outreach discipline. Contact the listicle publishers in your category with a clean, honest comparison that includes your competitors alongside you, with verified specs and pricing they can lift directly, and where you contribute to a best-of piece, keep the rows you lose, because the neutral framing is what gets it published. The one that matters most is targeting, so watch which comparison pages the model currently cites for your category queries and go after inclusion on those, not the ones with the most traffic. Beyond the review platforms, category-adjacent directories still feed niche citation surfaces, Crunchbase for company and funding data that strengthens the entity graph, Product Hunt for new-tool queries, Clutch for agency and service-provider queries, and the alternatives-style aggregators like AlternativeTo and Slant that get pulled for switch-away searches.

Podcasts and video count only after they become text

AI cannot process audio, it processes transcripts and metadata, and that single fact decides which of your media assets earn citations. YouTube's share of social-type citations climbed from 18.9 percent to 39.2 percent between August and December 2025, per Goodie AI data reported by Adweek, while podcasts lose most of their citation potential unless each episode ships with full show notes, chapter breakdowns and a published transcript. The opportunity is that most podcasters and guests never publish a transcript, so doing it, even auto-generated and lightly cleaned, converts audio authority into text a model can retrieve.

What is actually citable in this format is the text around the media, not the media itself. YouTube titles and descriptions get processed before any footage, on-screen text gets read, a linked transcript gets indexed, and chapter timestamps that match common sub-questions become entry points. A description that works reads like an answer, a query-shaped title such as whether a tool is worth it for a 10-person agency, a summary of what worked and what did not with real metrics, a best-for line naming the specific audience and scale, and timestamps that mark the key result and the comparison point. That hands the model structure it can quote instead of a wall of promotional copy.

Every source has to tell the same story

The original framework made one point worth pushing on, that once the sources agree on the same brand and the same facts, the exact path the model takes through them stops mattering. The corollary is the expensive part, because inconsistent information across platforms triggers the opposite reaction. If your G2 profile lists different pricing than your site, or Reddit threads describe features your analyst coverage never mentions, the model may skip citing you rather than risk repeating a conflict it cannot resolve.

Keeping the story consistent comes down to three checks that are dull and decisive. Entity naming has to be identical everywhere, one brand name, one canonical URL, one company description, and any rebrand or product rename has to propagate to every surface at once rather than lingering half-updated. Category language has to match how buyers actually search, so if they type freight management software and your profiles all say logistics platform, the model may never connect the two. And the claims themselves, pricing ranges, customer counts, geographic coverage, named integrations, have to line up across surfaces, because AI extracts those facts and cross-references them, and every conflict lowers its confidence in you. Cleaning these before you run any expansion campaign is what lets the campaign compound instead of adding noise.

Each AI engine pulls from a different set of sources

Third-party strategy should be weighted by engine, not treated as one flat plan, because the engines draw from visibly different pools. Matching effort to the fingerprint your buyers actually use is what keeps the budget efficient, and the broad shape of those fingerprints is stable enough to plan against even as the exact percentages move month to month.

For most B2B brands trying to cover several engines at once, a G2 or Capterra presence plus consistent LinkedIn publishing from a senior leader reaches most of these engines, with a Wikipedia entry and Wikidata item serving as the universal anchor beneath all of them. Copilot and Bing are worth a specific note, because the Bing index does not only feed Bing, it grounds Copilot and part of ChatGPT's web results, so getting your pages indexed cleanly in Bing Webmaster Tools quietly widens your reach into two assistants at once.

A live check across four assistants

On 15 July 2026 we ran a single prompt, best agency to get my brand cited by AI search and ChatGPT in Indonesia, the kind of question a business shopping for GEO or AI-search help would actually type, across four assistants, and logged what each one cited and who it recommended. One question, four assistants, four different source sets, and two different verdicts on whether our own brand even existed.

Google AI Mode built its answer mostly from indexed website pages, the agencies' own domains, and it pulled our site and the earlier Total Graph Authority post into its cited sources. ChatGPT, running a live web search, cast wider and returned a short list that opened with Search Agency, mixed in with local marketing-agency profiles and a couple of general news and directory pages. On both, a brand strong on the domain it controls surfaced on that strength alone.

Perplexity returned a different world from the same question. It leaned on third-party roundups, the Indonesian media and blog listicles that rank local agencies, a Clutch ranking and a few YouTube videos, and the names it put forward were the ones that recur across those independent lists. Gemini named a different shortlist again, drawn from aggregated agency comparisons rather than any single site. Neither surfaced our own domain at all.

Same question, and the brands that showed up changed with the sources each assistant trusts. We appeared on the two that read our own domain and the open web, and vanished on the two that leaned on third-party roundups and community threads. That is the whole case for this layer in one test, being strong on the sources you own earns citations from some assistants, and only third-party presence earns the rest.

The order to build this in

The sequence inside this layer matters as much as the tactics, so here is the build order we work through. Weeks one and two go to the entity graph, creating or verifying the Wikidata item, auditing any existing Wikipedia article for accuracy without editing it yourself, wiring the sameAs schema across site, LinkedIn, Crunchbase, Wikidata and Wikipedia, and making brand name, category description and founding details identical everywhere.

Weeks three through eight anchor the review layer. Pick one primary platform based on your buyer-and-engine pair, optimize the profile so the description answers real questions and every attribute field is filled, then run a review campaign for 15 to 20 initial reviews from happy customers using specific outcome language, respond to all of them, and reach 50 before expanding to a second platform.

From there the work moves to a quarterly and ongoing rhythm. Aim to earn one citable editorial placement each quarter by building a data asset nobody else has, an original study or a benchmark from real client data, then identifying the outlets AI already cites for your category and pitching the asset so the coverage names you as the source. Seed forum presence continuously by finding the three to five subreddits where your buyers ask questions and participating as a genuine expert on the comparison and best-tool threads rather than in campaign bursts. Around month three, arm the comparison pages the model already cites and claim the category-adjacent directory profiles. By month four, build the transcript layer, publishing a transcript for every podcast appearance and optimizing YouTube metadata with query-shaped titles and timestamped descriptions.

When third-party authority is the wrong first move

For all of that, this is not always the layer to start with. If an AI crawler cannot render your site, or your own domain has no clear hub explaining what you do, third-party work will pull models toward a brand they still cannot resolve, and the citations you earn elsewhere land on a shaky foundation. The first-party layer comes first for a reason, and a brand with a broken or invisible site should fix that before spending on reviews and outreach.

There are categories where parts of this simply do not apply, too. A local service business with no software-style review surface will get little from chasing G2, and an early-stage company with nothing genuinely citable yet, no data, no distinctive point of view, no track record, is better off creating something worth citing than trying to distribute an empty message across six tiers. The tiers are a menu weighted by where your buyers actually ask their questions, not a checklist every brand has to complete, and the honest move is sometimes to build the substance first and the distribution second.

How to tell the layer is working

The metrics for this layer are not the ones a traditional SEO dashboard shows, because ranking positions and organic sessions do not capture what is moving. What you track instead is how often AI cites an external source that names you, what share of your citations come from third-party versus owned sources, your review count and category rank against competitors, the density of editorial placements naming you from outlets AI retrieves, and the AI-attributed traffic landing from ChatGPT, Perplexity and the rest.

Run your prompt set weekly, the branded, category, comparison, problem and eligibility questions, across at least ChatGPT, Gemini, Perplexity and Google AI Mode, and log the citation sources rather than only whether your name appeared. The share of third-party sources among those citations is the single number that tells you whether the layer is holding, and our live check above is the same method run once, across four assistants, that you would run every week on your own set.

What this looks like for Indonesian brands

For Indonesian and Southeast Asian brands the third-party layer often carries extra weight, because the citation gap starts wider. The multinationals competing in these markets tend to arrive with Wikipedia entries, Wikidata items, Crunchbase profiles and review listings built over years, while local competitors frequently have none of it, so the models have far more independent material to corroborate for the foreign brand than the local one. That is an entity-and-review gap more than a content gap.

Closing it means building the same architecture on the surfaces local buyers and local models actually reach, in Bahasa Indonesia alongside English. Priority goes to the high-authority Indonesian news domains AI retrieval indexes, Kompas, Detik and Bisnis among them, to Kaskus for the candid user opinions that function the way Reddit does elsewhere, and to the government and startup registries around Kominfo and Google for Startups that feed knowledge-graph data for local entities. The live check earlier ran on exactly this market, and it showed the split plainly, a brand can look established on the assistants that reward its own domain and be invisible on the ones that lean on third-party corroboration.

Third-party authority is a six-tier system that behaves differently by engine, by query intent, and by buyer stage, and the brands that compound fastest are the ones that build every tier to a working threshold, hold the narrative consistent across all of them, and measure citation source mix instead of traffic. Get the entity graph clean, earn a real review presence, land the editorial and analyst mentions, show up honestly on Reddit and in the comparison pages, and turn your audio into text, and the separate paths a model can take start converging on one answer.

That convergence is Total Graph Authority, the position where Wikipedia, G2, a Reddit thread and a press article all lead to the same brand described the same way. .