How AI Search Engines Decide What to Cite: 3 Factors, Measured Across 15,883 AI Answers

AI search engines decide what to cite in two stages: retrieval finds pages relevant to the question, then source selection chooses passages that are clear, credible, current, and easy to attribute. Measured across 15,883 AI answer runs and the 22,213 domains those answers cited, the selection stage does not favour brand-name publishers. Half the top-ten slots in a buyer's question (51.5% of 880 slots across 88 measured segments) go to specialist sites outside every named publisher, platform, vendor, analyst and database class. Brands can improve three inputs: earned authority, entity clarity, and citation architecture. Rankings can aid discovery, but they do not guarantee citation.

OpenAI explains that ChatGPT search retrieves web sources and attaches citations to the answer. Google takes a related but distinct approach: its AI features documentation says a page must be indexed and eligible to appear with a snippet before it can be used as a supporting link. Discovery matters. The final source choice is a second decision.

That second decision is the one almost nobody measures. Every figure in the three measured sections below is read from the Machine Relations Index, release mri_score_v2.0+2026-09-19+0cad03121f60: 15,883 answer runs on 912 eligible buyer queries across six engines (ChatGPT, Claude, Gemini, Google AI Mode, Google AI Overviews, Perplexity), observed 2026-05-10 to 2026-09-19, producing 125,115 source events and 111,510 domain-run citations over 22,213 distinct cited domains. It measures the market's buyer questions, not our own brand terms.

The practical work maps to the first three layers of Machine Relations: earned authority creates independent evidence, entity clarity tells the engine who the evidence belongs to, and citation architecture makes the claim easy to extract.

Key Takeaways

• Retrieval eligibility comes first. Google requires ordinary indexability and snippet eligibility for pages supporting AI features.
• Selection comes second, and it is measurable. Across 88 published segments of the Machine Relations Index, 51.5% of the 880 top-ten slots are held by domains outside the eight named source classes: specialist sites, not brand-name publishers.
• Editorial publications are 1,224 domains and 13,389 domain-run citations of 111,510, and the leader of that class is medium.com at 6.0% of it. Independent coverage corroborates a claim; it is not a shortcut to the top of an answer.
• Question shape moves the answer's source mix more than subject does. Comparison questions give vendor-owned sources 13.8% of the top ten; the same brands get 6.4% when the buyer asks how to choose.
• Inside a segment the top is shallow: the most-cited source's rate is a median 2.5x the tenth-most-cited source's, so roughly ten sources share the answer surface.
• Attribution is the last checkpoint and it fails often. Of ChatGPT's 1,255 resolvable citations measured by Answer-Source Fidelity, 54.8% actually supported the claim beside them.
• The three controllable inputs are earned authority, entity clarity, and citation architecture.

Why Search Rankings Do Not Guarantee AI Citations

For two decades, getting cited online meant getting ranked. If your page appeared in the top 10, it got traffic and attribution. If it didn't, it was invisible. The signal and the outcome were inseparable.

A search ranking is an eligibility signal, not a citation contract. AI answer systems still need to decide which passage best supports the generated answer and whether that passage can be attributed cleanly.

• Google uses ordinary Search eligibility as the floor. Its AI-features documentation says no special AI file or schema is required, but the page must be indexed and eligible for a snippet.
• ChatGPT search exposes a separate source layer. OpenAI's product documentation describes web search and inline citations as part of the answer experience.
• Citation quality must be audited independently from answer quality. A 2026 PMLR study of Google AI Overviews evaluates whether citations actually support the claims they accompany, not only whether the prose sounds correct.
• Clicks are no longer the only outcome. Pew Research found that users clicked a traditional result in 8% of visits when an AI summary appeared, compared with 15% when one did not.

This is not a marginal shift. The systems operate on different selection criteria. A brand that has spent five years building SEO authority through backlink acquisition and keyword optimization may be completely invisible in AI answers, while a competitor with lower domain authority appears consistently, because they have earned media coverage in publications AI engines trust.

Understanding why this happens requires understanding what AI search engines are actually doing when they decide what to cite. The answer is more specific, and more actionable, than most guidance in this space suggests.

How AI Search Engines Make Citation Decisions

Most AI search products that use the live web follow a retrieval-and-generation pattern. The system searches an index for relevant material, filters candidate passages, synthesizes an answer, and attaches sources. Product details differ by engine, but the operating sequence is consistent enough to expose four checkpoints: eligibility, relevance, support, and attribution.

A 2025 cross-engine citation study audited 1,702 citations from Brave Summary, Google AI Overviews, and Perplexity. Its 16-part framework connected citation outcomes with content quality, semantic HTML, structured data, metadata, and freshness. The takeaway is operational: machines need a source they can retrieve, understand, and quote without inventing missing context.

Those four checkpoints run differently in each product, and the per-engine detail is worth reading separately from the cross-engine pattern. For how the major platforms differ in what they disclose versus what can be measured, see how AI platforms like ChatGPT and Perplexity decide what to cite. For one engine audited stage by stage against its own documentation, see how Perplexity decides which sources to cite. The circulating figures about that pipeline — a five-stage gate system, a 0.75 quality threshold, a 40% freshness decay — are checked claim by claim in the Perplexity citation numbers everyone quotes; most of them trace to agency write-ups with no published sample, so treat them as hypotheses rather than disclosure.

A citation has to do more than point to a real page. It also has to support the nearby claim. The VeriCite research on retrieval-augmented generation separates citation correctness from citation completeness and uses claim-level verification to test whether the evidence actually entails the answer.

Inside that study's own corpus, the pillars most strongly associated with citation were Metadata and Freshness, Semantic HTML, and Structured Data. Those are machine-readability signals rather than traditional SEO metrics, and they are the best outside evidence available on the question. They are also a 1,702-citation sample of three engines. The three sections that follow answer the same question from a larger and fresher measurement of our own.

Which Sources AI Engines Actually Cite, Measured

Across 15,883 answer runs, the sources AI engines cite are not concentrated in the publications brands compete for. They are concentrated in a handful of platforms and spread across a very long tail of specialist sites.

The Machine Relations Index classifies every cited domain into one of nine source roles. On release mri_score_v2.0+2026-09-19+0cad03121f60, the nine classes close exactly on 22,213 domains and 111,510 domain-run citations:

Source class	Cited domains	Domain-run citations	Citations per domain	Class leader, share of its own class
Other observed source	18,655	67,491	3.6	healthline.com, 0.4%
Editorial publication	1,224	13,389	10.9	medium.com, 6.0%
Vendor-owned source	823	11,203	13.6	microsoft.com, 2.7%
Market and company database	688	6,092	8.9	g2.com, 7.4%
Community and social platform	27	4,319	160.0	reddit.com, 46.6%
Academic and government source	413	3,880	9.4	nih.gov, 12.6%
Analyst and consulting research	364	3,390	9.3	gartner.com, 13.5%
Search or media platform	10	1,515	151.5	youtube.com, 94.6%
Wire and press-release distribution	9	231	25.7	businesswire.com, 42.4%

Three readings matter for anyone deciding where to spend:

• 60.5% of all citation goes to domains outside the eight named classes. The classified layer, every publisher, platform, vendor, analyst, database, academic and wire source the taxonomy recognises, accounts for 39.5% of citation. The rest is specialist sites the market has never heard of, each earning 3.6 citations on average.
• The smallest classes are the densest. Twenty-seven community and social domains carry 4,319 citations, 160.0 each. Ten search or media platforms carry 1,515, 151.5 each. The 1,224 editorial publications average 10.9. Density is not authority; it means the engines return to a few destinations constantly and to everyone else occasionally.
• The press-release lane barely exists in answers. Nine wire and distribution domains carry 231 domain-run citations in total, 0.2% of the measured citation, and businesswire.com is 42.4% of that sliver.

How to read the class figures. Two class figures above are read off the Machine Relations Index's source_role field, which in the current release combines source-type evidence with AuthorityTech's placement catalog; 96 of the 1,231 class members are admitted through the catalog, and the Index is moving editorial classification to source-type evidence alone. The class currently includes six platform domains alongside edited newsrooms: medium.com (class rank 1), amazon.com (7), prnewswire.com (8), apple.com (16), wordpress.com (239) and blogspot.com (330), while substack.com, beehiiv.com and dev.to — the same hosted open-publishing product as Medium — are filed as community platforms. Those six hold 1,358 of the editorial class's 13,525 cited runs, 10.04%, including its top slot, so the class count and its per-domain average include those platforms. The nine-domain wire and press-release class likewise excludes prnewswire.com; counted with it, the class is ten domains at 414 of 112,518 cited-domain observations, 0.37% rather than 0.21%. Independent of class membership, per-domain citation rates, cited runs, days cited, engine breadth and confidence are direct observations over a fixed run set, and are the primary measure. See the MRI methodology and update log.

Comparing concentration across these classes uses each leader's share of its own class rather than a top-ten share, because two classes hold ten or fewer domains and any top-ten column would read 100% for them by construction.

The Top Ten Is Where Citation Lives, and Half of It Is Specialists

An answer cites a handful of sources, so the operative question is not "am I in the index" but "am I in the top ten for the question my buyer asks." The Index publishes 88 segments, each a subject category paired with a buyer question shape, and ranks every domain inside each one. That is 880 top-ten slots on this release:

Source class	Top-ten slots held	Share of 880 slots
Other observed source	453	51.5%
Editorial publication	118	13.4%
Community and social platform	103	11.7%
Vendor-owned source	95	10.8%
Search or media platform	46	5.2%
Academic and government source	29	3.3%
Market and company database	21	2.4%
Analyst and consulting research	15	1.7%
Wire and press-release distribution	0	0.0%

Two things follow. A specialist page that answers one question precisely holds more of the answer surface than the entire editorial, analyst and database layer combined, which is the strongest available argument for publishing depth on your own domain. And press-release distribution holds none of it: zero of 880 slots.

The top of a segment is also shallow. Across all 88 published segments, the rank-1 domain's citation rate is a median 2.5x the rank-10 domain's: the narrowest gap is 1.61x (HR and talent news questions, where unleash.ai at 11.35% leads linkedin.com at 7.07%) and the widest is 5.5x (AI visibility problem-first research, where youtube.com at 33.3% leads yotpo.com at 6.06%). Ranking first in an answer set is worth roughly two and a half times the tenth position, not a hundred times it. Ten sources genuinely share the answer.

Question Shape Reallocates the Answer

The same subject, asked in a different shape, hands the answer to a different class of source. Holding the segment population constant and splitting the 880 slots by the buyer's question shape:

Question shape	Segments	Specialist sites	Editorial publications	Vendor-owned	Community and social
How buyers choose	14	64.3%	10.0%	6.4%	10.0%
Best tools	14	56.4%	17.1%	7.1%	8.6%
Top lists	14	60.7%	11.4%	9.3%	7.1%
Is it worth it	13	49.2%	14.6%	9.2%	13.1%
Problem-first research	13	54.6%	11.5%	6.9%	16.9%
Comparisons (X vs Y)	13	49.2%	10.8%	13.8%	13.1%
News topics	7	0.0%	22.9%	34.3%	15.7%

Four of the nine classes are shown; the remaining slots in each row go to search and media platforms, academic and government sources, databases and analyst research.

The operational reads:

• Comparison questions are the vendor's best shape. Vendor-owned sources take 13.8% of top-ten slots on X-vs-Y questions and 6.4% on how-buyers-choose questions, a bit over double, on the same engines in the same window. If your own domain is going to be cited anywhere, it is on the comparison.
• Problem-first research is the community's shape. Community and social platforms hold 16.9% of those slots, their highest of any buyer shape, against 8.6% on best-tools questions.
• News questions are the exception that proves the rule. In the seven news-topic segments, not one top-ten slot goes to a source outside the eight named classes: brand-name publishers, vendors, analysts and databases take all 70. When the buyer asks for news, recognised institutions win. When the buyer asks how to choose, they mostly do not.

This is the same mechanism as the rank-1 source role by buyer segment, measured a layer deeper: not who wins, but how the whole top ten is allocated. It is why a single "AI visibility" content plan underperforms. The shape of the question decides which kind of source the engine reaches for, and your earned coverage, your own pages and your community footprint are answering different questions.

Three factors govern what a brand can do about all of this. They correspond to Layers 1, 2, and 3 of the Machine Relations framework.

Factor 1: Earned Authority, the Publication Trust Signal

Earned authority is trusted third-party coverage in publications that AI systems already recognize as credible, and its job is corroboration. It is not the case that engines refuse to cite a brand's own domain: vendor-owned sources average 13.6 domain-run citations each on this release, above the editorial class's 10.9, and hold 10.8% of every measured top-ten slot. What a brand's own page cannot do is corroborate itself. Independent coverage is what lets an engine compare two accounts of the same claim, and its share of the answer rises where the buyer wants an assessment rather than a spec sheet: 22.9% of top-ten slots on news questions and 17.1% on best-tools questions, against 10.8% on comparisons.

Independent coverage matters because it changes the evidentiary status of a claim. A statement on your own site is self-description. The same statement reported, checked, and contextualized by a credible publication becomes third-party corroboration. Search-augmented systems can compare those independent accounts before they attribute the claim. The signals an engine reads to decide a source is credible in the first place — entity consistency, cross-source corroboration, evidence density — are broken out in how AI search engines evaluate source credibility.

• Editorial coverage supplies independent corroboration. The publication, author, date, and surrounding reporting give the engine more context than an isolated brand claim.
• Primary reporting carries source ownership. Reuters, Bloomberg, government filings, academic papers, and original company announcements let the engine trace a fact back to the party that produced it.
• Repeated claims across independent sources improve confidence. The goal is not a pile of copied mentions. It is consistent attribution across sources with distinct editorial control.

What earned authority is worth is measurable, and the measurement is less flattering than the pitch. Editorial publications are 1,224 cited domains carrying 13,389 of 111,510 domain-run citations, and the leader of that class is medium.com at 6.0% of it. The tier brands actually buy is far narrower than the class: on the same release, the measured peer table of ten major business publications spans 59x from forbes.com at 4.14% of the runs it appeared in to inc.com at 0.07%. "Get covered in a credible publication" is not one decision. It is a rate schedule, and most of the names on the usual media list sit at the bottom of it.

The mechanism still holds where it holds. A claim reported by an outlet with editorial control is independent evidence an engine can compare against the brand's own page, and news-shaped questions hand recognised publishers 22.9% of every top-ten slot, their strongest shape by a wide margin. What the data does not support is the idea that a placement anywhere in the editorial class buys a position in a buyer's answer. Named publishers hold 13.4% of the answer surface across all measured segments; the specialist long tail holds 51.5% of it.

For enterprise AI vendors, source credibility also has a procurement consequence: inference trust becomes the control plane through which buyers verify security, provenance, observability, and operating limits before a model reaches production.

This is why PR's original mechanism, earned media in respected publications, has not become obsolete in the AI era. It has become structurally more important. The publications themselves haven't changed. What changed is who is reading them. Machines are now the primary reader of the content that determines brand citation, and they respond to the same editorial trust signals that shaped human perception for decades. The practice of briefing for that reader has its own name and definition: PR for machine readers.

A December 2025 arXiv study (Zhang et al., arXiv:2512.09483) found that LLM-based search engines source a smaller set of URLs than traditional search results and show different domain-selection patterns. A separate 2025 GEO-16 study of B2B SaaS citations cautioned from its corpus that even high-quality pages may lose visibility when they exist only on vendor-owned blogs. The practical point is narrower than a universal ranking rule: citation architecture helps, but outside corroboration still matters.

The practical implication: earned authority is a rate schedule, not a badge. Pick the outlets measured as cited for the questions your buyers ask, and treat everything else in the media list as reach rather than citation. Not ad spend. Not link building. Not keyword density. Measured placement, then the two layers below.

Factor 2: Entity Clarity, Machine Legibility

Entity clarity is the degree to which AI systems can unambiguously identify, categorize, and relate a brand to its category. A brand that AI systems cannot confidently resolve will not be cited confidently, even if that brand has relevant content and earned media coverage.

This is the factor most marketing teams have never addressed, because it was irrelevant in keyword-based search. Keywords didn't require machines to understand who you are, they required machines to match text patterns. Entity clarity requires something fundamentally different: consistent, structured signals across multiple platforms that allow AI systems to build a coherent internal model of who your brand is, what category it belongs to, and what it is known for.

When a user asks ChatGPT "which agency specializes in AI brand visibility," the AI system doesn't just retrieve relevant pages. It attempts to resolve which entities, specific named companies, are credibly associated with that category. Brands that have built clear entity signals appear as resolved candidates. Brands that haven't built those signals are absent from the resolution, regardless of how much content they've published.

The signals that build entity clarity:

• Consistent naming across platforms, the brand name, founder name, and company description use the same language everywhere: website, LinkedIn, Crunchbase, Wikipedia, press mentions, and any other platform where the brand is mentioned
• Cross-platform presence, the brand exists as a named entity in multiple independent sources, not just its own web properties; each independent mention strengthens the entity signal
• Structured data markup. JSON-LD Organization and Person schema on web properties that explicitly defines the company, its founder, its category, and its relationships to other entities. Google's structured-data documentation explains that markup gives Search explicit clues about a page and its meaning.
• Category association, consistent linkage between the brand and the specific category it wants to own in AI responses, established across multiple independent sources

The Kumar et al. arXiv study (2509.10762) found that Structured Data and Semantic HTML were among the pillars most strongly associated with citation across all three AI engines studied. The December 2025 arXiv study (Zhang et al., 2512.09483) found that domains favored by LLM-based search engines exhibited "more structured, hierarchical HTML, easier-to-read text, and more outlinks to reputable sources", the structural signals that enable entity resolution.

The measured scale of the resolution problem: 18,655 of the 22,213 cited domains, 84.0%, sit outside every named source class, and they average 3.6 citations each across 126 days of observation. Being retrievable is cheap and being resolvable is not. The domains that compound are the ones an engine can place: a category, a function, a named author, a consistent description across the properties where it is mentioned.

Being cited once is also not the same as being cited durably. Measured on the head of the Index, the 25 most-cited domains are absent on 42.4% of their observed domain-days, with a median presence of 69 days out of 125. Resolution has to hold up on a random Tuesday, not once.

Entity clarity is not a technical SEO exercise. It is the process of making your brand legible to machine readers who are trying to determine whether you are the right source for a given query. A brand that is unambiguously associated with a specific category and function, across its own properties, its earned media placements, and third-party data sources, as a brand that AI systems can cite with confidence. A brand that sends inconsistent signals, or that exists only on its own domain, is a brand that AI systems resolve with uncertainty. Uncertain resolution produces fewer citations.

Factor 3: Citation Architecture, Structural Extractability

Citation architecture is the structural formatting of content, data density, FAQ sections, tables, and answer-first structure, that makes specific claims independently extractable by AI retrieval systems. A page that reads well for humans but lacks these structural signals is a poor candidate for AI citation, regardless of how it ranks or how authoritative its source publication is.

The foundational research on this factor is the GEO paper (Aggarwal et al., KDD 2024), which tested how content changes affected visibility across generative engines and 10,000 queries. Its strongest methods added source citations, quotations, and statistics instead of stuffing more keywords into the page.

• Cite sources. A direct link to the primary evidence gives the engine an attribution path it can inspect.
• Add specific statistics and quotations. Concrete evidence gives the engine a bounded passage to extract instead of asking it to infer the claim from broad prose.

The mechanism is specific to how RAG pipelines work. An AI search engine that retrieves a page doesn't read it the way a human does, it identifies passages that directly answer the query and extracts them as citation candidates. A page that states its key claim in the first sentence, supports it with a named statistic, and presents related data in a table provides clean extraction targets. A page that buries its main point in the seventh paragraph, uses passive voice throughout, and presents data in flowing prose requires the system to do interpretive work that reduces extraction confidence, and therefore reduces the likelihood of citation.

Source controls also shape retrieval eligibility. Google's robots meta tag documentation defines how publishers allow or restrict indexing and snippets. A blocked page cannot become a reliable citation target, no matter how strong the prose is.

The specific elements of citation architecture that the research supports are qualitative signals, not universal numeric thresholds:

• Answer-first structure, the main claim appears early, stated declaratively, self-contained, and usable by an AI system without surrounding context
• Data density, named statistics and quotations from primary sources, inline-cited with direct URLs to the original document; each sourced fact gives the AI a specific, attributable claim to extract
• FAQ sections, question-answer pairs that AI systems can treat as direct extraction targets; the format mirrors how AI systems receive queries, making extraction more reliable than parsing body prose
• Tables, structured comparison data that presents relationships between entities, metrics, and claims in a format AI systems can extract and represent without paraphrase
• Metadata and freshness signals, the GEO-16 study found Metadata and Freshness to be the pillar most strongly associated with citation in its B2B SaaS corpus, including datePublished and dateModified markup, which signals to AI systems that the content is current and worth surfacing

Citation architecture is the layer where most "GEO optimization" efforts concentrate, and it's the most accessible layer for brands to address without external relationships. But the Kumar et al. study is clear that architecture alone does not guarantee citation: content that lives exclusively on brand-owned properties, regardless of how well-structured it is, faces a citation ceiling that earned authority breaks through.

Architecture without authority is optimization without foundation. The three factors work as a system, not as alternatives.

The Last Checkpoint: Does the Citation Actually Support the Claim?

Getting cited is not the end of the decision. The engine still has to attach your page to a sentence your page supports, and measurably it often does not. Our second instrument, Answer-Source Fidelity, grades whether an engine's cited source actually supports the claim beside it, engine by engine, over the same population of answers.

On the current published aggregate, reading each engine over its own resolvable set: of the 1,255 resolvable citations for ChatGPT, 54.8% were Supported, with resolvable coverage of 77% (1,255 of 1,630). Of the 1,298 resolvable citations for Perplexity, 50.1% were Supported, at 47.6% coverage (1,298 of 2,726). Claude and Gemini remain below the instrument's evidence floor on resolvable coverage, so it reports no rate for them and neither do we. The instrument also discloses its own bias: against the AttrScore reference set it agrees on 6 of 18 extrapolatory cases, which is where a human grader and a machine grader disagree most.

The operational consequence for citation architecture is direct. Close to half of the attributions an engine makes do not hold up against the page it points at, which means the page that states its claim in one sentence, with its own number, its own denominator and its own date, is doing the engine's verification work for it. That is not a style preference. It is the difference between being cited and being cited correctly.

The Scale of the Shift: Why AI Citation Matters Right Now

Two measured changes explain why citation now matters as much as the click:

Google said AI Overviews had reached more than 1.5 billion monthly users by May 2025. That figure came from Google's own I/O announcement. AI-generated synthesis is already a mainstream discovery surface.

Pew Research found that users clicked a traditional search result in 8% of visits when an AI summary appeared, versus 15% when it did not. The 2025 browsing-behavior study also found that clicks on sources inside the AI summary were rare. The brand can influence a buying decision before the buyer ever reaches the site.

The buyer who asks Perplexity or ChatGPT which company leads a category is already building a shortlist. If the engine cannot retrieve credible, attributable evidence about your brand, your sales team never sees the lost opportunity.

The stakes extend beyond individual queries. Brands that build AI citation infrastructure now, earned authority, entity clarity, citation architecture, accumulate a compounding advantage. Each placement in a trusted publication strengthens entity resolution. Each well-structured piece of content adds another extraction target. Each FAQ section planted in a credible source is a direct input to AI answers. Brands that wait to address these factors are not maintaining the status quo. They are falling behind competitors who are actively building the infrastructure that AI citation requires.

How Different Disciplines Approach AI Citation

The AI citation challenge is being addressed from multiple disciplines simultaneously. Each discipline has real value, and each has a gap. Understanding where each approach sits clarifies why a full-system response is necessary.

Discipline	Optimizes for	Success condition	Scope
SEO	Ranking algorithms	Top 10 position on SERP	Technical + content
GEO	Generative AI engines	Cited in AI-generated answers	Content formatting + distribution
AEO	Answer boxes / featured snippets	Selected as the direct answer	Structured content
Digital PR	Human journalists/editors	Media placement	Outreach + storytelling
Machine Relations	AI-mediated discovery systems	Resolved and cited across AI engines	Full system: authority → entity → citation → distribution → measurement

The gap in every discipline except Machine Relations: each optimizes for one layer of a three-layer problem. SEO addresses some citation architecture signals (technical structure, schema) but ignores earned authority entirely. GEO addresses citation architecture but typically treats earned media as an optional amplification layer rather than the foundational requirement it is. AEO and Answer Engine Optimization agencies are the closest to the full picture, but most operate at the content formatting level without the earned media infrastructure that determines whether that content gets cited in the first place. Digital PR builds earned authority but was not designed for machine extraction, placements in the right publications get AI citations, but the content itself may not be structured for reliable extraction.

Machine Relations is the discipline that addresses all three factors as a single integrated system rather than separate optimizations. Earned authority provides the trust foundation. Entity clarity provides the resolution signal. Citation architecture provides the extraction mechanism. All three must be present for AI citation to occur reliably.

What Brands Get Wrong About AI Citations

Three misconceptions account for most of the strategic errors brands make when trying to improve their AI citation rates:

Misconception 1: More content automatically produces more citations. Volume creates inventory. It does not create support. A source still needs a relevant passage, a traceable claim, and enough context for the engine to attribute it correctly.

Misconception 2: A strong ranking guarantees selection. Ranking can put a page into the candidate set. The answer system still has to find a passage that directly supports its claim. A broad page can rank well and lose the citation to a narrower source with cleaner evidence. The measured version of this is the shallow top ten: across 88 segments the most-cited source's rate is a median 2.5x the tenth's, and 51.5% of those slots belong to specialist sites rather than the domains that dominate search rankings for the same terms.

Misconception 3: AI citations are purely a content problem. This is the most costly misconception in practice. Brands invest in content formatting, FAQ sections, and structured data, all useful, while ignoring the earned authority layer that sits beneath them. A page with strong citation architecture but no third-party corroboration can still lose visibility to sources with stronger independent authority. The GEO-16 study makes that limitation explicit for its B2B SaaS corpus: on-page quality matters, but vendor-owned pages still face a citation ceiling without broader authority. Architecture without authority is optimization without foundation.

How to Get Cited by AI Search Engines: The Practical Sequence

Based on the three citation factors and the research supporting each, here is the sequence of actions that produces measurable improvement in AI citation rates:

1. Build earned authority against the measured rate, not the brand list. Secure placements in the publications measured as cited for your category's questions, and sequence the target list by that rate: on the current release forbes.com is cited in 4.14% of the runs it appears in, techcrunch.com 1.03%, entrepreneur.com 0.11% and inc.com 0.07%, so a list that treats those four as one interchangeable tier is off by 59x. Each placement in a measured outlet is a trust transfer from the publication's credibility to your brand. This is the layer no amount of on-page optimization can replicate. You can read your brand's current AI answer footprint at any point using AI monitoring tools, but fixing it starts with earned media, not with monitoring.
2. Establish entity clarity across the graph. Audit your brand's presence across the entity graph: Wikipedia, Crunchbase, LinkedIn, your own website schema. Verify consistent naming, category association, and structured JSON-LD markup. Inconsistent descriptions across platforms create entity resolution uncertainty. Uncertainty produces fewer citations, even from brands with strong earned authority.
3. Structure every piece of content for machine extraction. Open with a clear answer to the question the page is meant to satisfy. Use named statistics from primary sources when the claim needs evidence. Add tables for comparative data and FAQ sections with declarative question-answer pairs. These are not arbitrary style preferences; they are structure choices that make content easier for AI retrieval pipelines to parse, verify, and extract.
4. Publish original research. This is the step the measurement argues for hardest: if half the top-ten slots in a buyer's question belong to specialist sites, then one page carrying a number nobody else holds outcompetes ten pages restating the same borrowed study. A primary dataset, transparent methodology, or repeatable analysis gives other publications and answer engines something they cannot source elsewhere. State the sample, date, method, and limitation beside the finding. OpenAI describes deep research in ChatGPT as a workflow that searches, analyzes, and synthesizes sources into a documented report, which makes source-owned evidence the raw material of the answer.
5. Maintain recency signals. The GEO-16 framework found Metadata and Freshness to be the pillar most strongly associated with citation. Pages with current datePublished and dateModified markup, regularly updated content, and recent references consistently outperform stale content, even when the underlying claims remain accurate. Use the content freshness framework to decide which cited pages need monthly, quarterly, or annual refreshes.

The sequence matters. Brands that attempt citation architecture before establishing earned authority are optimizing the packaging of content that lacks distribution rights. Brands that earn media placements but fail to structure the content for extraction leave attribution on the table. The three factors work in sequence: authority enables extraction, entity clarity enables attribution, and citation architecture enables reliable, repeatable citation at scale.

Why Earned Media Is the Mechanism That Makes It All Work

Machine Relations (MR) is the discipline of earning AI citations and recommendations for a brand by making that brand legible, retrievable, and credible inside AI-driven discovery. I coined Machine Relations in 2024 after watching the same editorial placements built for human trust become evidence for machine-generated answers. AuthorityTech operationalizes the discipline through earned media, entity clarity, citation architecture, distribution, and measurement.

The mechanism behind Machine Relations is the mechanism that made traditional PR valuable: earned media placements in publications that trusted audiences, and now trusted machines, rely on for authoritative information. A Forbes placement that generates an AI citation is not a new phenomenon requiring a new technology stack. It is the original PR mechanism, now operating with machine readers as the primary audience alongside human ones.

PR got one thing exactly right: earned media. A placement in a respected publication, secured through a real editorial relationship, as the most powerful trust signal available for AI citation. It was true when buyers were human readers. It is true now that AI systems perform the first cut of research on a buyer's behalf. The publications that shaped human brand perception for decades are the same publications AI engines treat as authoritative sources. The editorial relationships that produced those placements are the same relationships that now produce AI citations.

What changed is not the mechanism. What changed is the reader. Machine Relations is the name for understanding this shift, and for building the infrastructure that makes earned media produce both human and machine attribution simultaneously.

Brands that have been consistently building earned media infrastructure are discovering that their existing placements translate directly into AI citations. Brands that optimized exclusively for SEO rankings are discovering those rankings produce no equivalent benefit in AI answers. The investment thesis for which brands dominate is being rewritten by this shift.

Start your visibility audit →

Frequently Asked Questions

How do AI search engines decide what to cite?

AI search engines decide what to cite by retrieving relevant pages, selecting passages that support the answer, and attaching sources the system can attribute. Google requires ordinary indexability and snippet eligibility for pages used by its AI features. Measured across 15,883 AI answer runs on release mri_score_v2.0+2026-09-19+0cad03121f60, the selection stage favours precision over brand: 51.5% of the 880 top-ten slots across 88 buyer segments are held by specialist sites outside every named publisher, platform, vendor, analyst and database class, and the rank-1 source's citation rate is a median 2.5x the rank-10 source's. Brands can improve selection through earned authority, entity clarity, and citation architecture.

Which kinds of sources do AI engines cite most?

Across 111,510 domain-run citations on the current Machine Relations Index release, 60.5% of citation goes to domains outside the eight named source classes, and those specialist domains average 3.6 citations each. Of the classified layer, editorial publications carry 13,389 citations across 1,224 domains (10.9 each), vendor-owned sources 11,203 across 823 domains (13.6 each), market and company databases 6,092, community and social platforms 4,319 across just 27 domains (160.0 each), academic and government sources 3,880, analyst research 3,390, search and media platforms 1,515 across 10 domains, and wire and press-release distribution 231 across 9 domains.

How to read the class figures. Two class figures above are read off the Machine Relations Index's source_role field, which in the current release combines source-type evidence with AuthorityTech's placement catalog; 96 of the 1,231 class members are admitted through the catalog, and the Index is moving editorial classification to source-type evidence alone. The class currently includes six platform domains alongside edited newsrooms: medium.com (class rank 1), amazon.com (7), prnewswire.com (8), apple.com (16), wordpress.com (239) and blogspot.com (330), while substack.com, beehiiv.com and dev.to — the same hosted open-publishing product as Medium — are filed as community platforms. Those six hold 1,358 of the editorial class's 13,525 cited runs, 10.04%, including its top slot, so the class count and its per-domain average include those platforms. The nine-domain wire and press-release class likewise excludes prnewswire.com; counted with it, the class is ten domains at 414 of 112,518 cited-domain observations, 0.37% rather than 0.21%. Independent of class membership, per-domain citation rates, cited runs, days cited, engine breadth and confidence are direct observations over a fixed run set, and are the primary measure. See the MRI methodology and update log.

Does the wording of the question change which sources get cited?

Yes, measurably. Splitting the same 880 top-ten slots by question shape, vendor-owned sources hold 13.8% of the slots on comparison (X vs Y) questions and 6.4% on how-buyers-choose questions. Community and social platforms peak at 16.9% on problem-first research questions. In news-topic segments, every one of the 70 top-ten slots goes to a recognised class: publishers, vendors, analysts and databases, with no specialist sites at all.

What are the AI citation factors that brands can actually control?

Brands can directly influence three citation factors. Earned authority comes from independent coverage. Entity clarity comes from consistent names, relationships, and structured data across the web. Citation architecture comes from direct answers, primary-source links, specific evidence, descriptive headings, tables, and clean question-and-answer blocks.

Who coined Machine Relations?

Jaxon Parrott, founder of AuthorityTech, coined the term Machine Relations in 2024 to name the discipline of earning AI citations and recommendations for brands. He published the origin story and the five-layer MR stack at machinerelations.ai. Machine Relations extends Public Relations into a world where machines mediate how brands are surfaced, compared, and recommended, creating the machine as the primary reader of content that determines brand visibility and citation.

Is Machine Relations the same as SEO or GEO?

No. SEO improves discovery and ranking in search results. GEO improves how content performs in generative engines. Machine Relations addresses the full system: earned authority, entity clarity, citation architecture, distribution across answer surfaces, and measurement. GEO and AEO are useful tactics inside the distribution layer.

Where do GEO and AEO fit inside Machine Relations?

GEO (Generative Engine Optimization) and AEO (Answer Engine Optimization) are distribution tactics within Layer 4 of the five-layer Machine Relations stack. GEO optimizes content formatting for generative AI engines; AEO optimizes content structure for answer boxes and featured snippets. Both tactics produce significantly better results when Layers 1-3 are already in place. A brand doing GEO without earned authority is optimizing the extractability of content that lacks the trust foundation needed for AI engines to cite it with confidence.

How is Machine Relations different from digital PR?

Digital PR optimizes for human journalists and editors, the success condition is a media placement that reaches human readers. Machine Relations uses the same mechanism (earned media in trusted publications) but structures content and entity signals so that machine readers can also parse, extract, and cite what has been earned. Digital PR was designed for human audiences. Machine Relations extends it to work for both audiences simultaneously, because AI engines are now doing the first cut of research that precedes human buyer decisions. The PR mechanism (earned media in respected publications) is correct. What Machine Relations adds is the entity and architecture layers that make that earned media produce AI citation, not just human readership.

Does schema markup guarantee an AI citation?

No. Google's AI-features documentation says no special schema is required to appear in AI Overviews or AI Mode. Valid structured data can help machines understand a page and its entities, but the page still needs to be indexable, relevant, supported by evidence, and useful enough to select.