AI citation behavior is model-specific, not universal. Primary-source evidence from Yext, Pew Research Center, Gartner, Bain, SparkToro, Muck Rack, and academic GEO research points to the same conclusion: AI visibility is not one optimization problem. Different models and AI-mediated search environments reward different source types, different query structures, and different trust signals (Yext, 2026; Pew Research Center, 2025; Gartner, 2024; Bain & Company, 2025).
Brands keep talking about AI search as if it were one channel. The data says otherwise.
Yext’s 17.2 million citation dataset shows that model behavior diverges at the source-type level #
Yext analyzed 17.2 million distinct AI citations gathered during Q4 2025 across four major AI models and found consistent, model-specific sourcing patterns across seven sectors (Yext, 2026). The same study found that listings represented 54.53% of distinct cited URLs, while first-party websites generated 4.31 citation occurrences per URL versus 2.46 for listings, which means models revisit owned sources more often even when directories dominate URL count (Yext, 2026).
That is not a small implementation detail. It changes how a brand should allocate effort across owned content, directories, reviews, and earned media.
| Model | Reported citation tendency | Why it matters for brands |
|---|---|---|
| Gemini | Strongest full-control preference across most sectors, ranging from 22.4% to 54.0% | First-party authority and search-grounded credibility matter more (Yext, 2026) |
| Claude | Limited-control citations run 2-4x higher than competitors across all seven sectors studied | Reviews, social proof, and user-generated validation matter more (Yext, 2026) |
| Perplexity | Full-control preferences cluster between 37% and 50% across most sectors | Retrieval behavior is more stable and search-like across industries (Yext, 2026) |
| SearchGPT | Highest industry variance, with full-control ranging from 28.2% to 43.7% | Source strategy must adapt more aggressively by vertical (Yext, 2026) |
The practical implication is blunt: optimizing one source layer while ignoring the others will make a brand overfit to one model and underperform in the rest.
Claude’s source mix is materially different from Gemini’s #
Claude consistently leaned harder on limited-control sources than every other model in Yext’s study, with limited-control citation share ranging from 6.3% to 24.4% depending on sector (Yext, 2026). In Food & Beverage, Claude cited limited-control sources 24.35% of the time, versus 2.57% for Gemini, a nearly 10x gap inside the same sector (Yext, 2026).
That means a brand can look well-positioned in a first-party-heavy environment and still disappear in a model that weights reviews, communities, and user-validated sources more heavily.
This is exactly why a single-surface GEO program breaks. If the system is designed only to strengthen owned pages, it will miss the source environments that some models clearly trust more.
SearchGPT’s strongest divergence appears in hospitality, not in the aggregate #
SearchGPT did not just vary a little by industry. Yext found that in Hospitality, SearchGPT cited official hotel websites 38.1% of the time, while competing models ranged from 16.7% to 22.4% (Yext, 2026). Yext called that the largest single-model divergence observed in any sector in the dataset (Yext, 2026).
That matters because it breaks the lazy habit of talking about average AI behavior. Sector-level or model-level averages can hide the exact places where a visibility strategy fails.
A hotel brand optimizing for third-party review visibility alone could still lose ground in SearchGPT if its official property pages are weak. The opposite could happen in a model that leans harder on user-generated validation.
Earned media still matters because AI citation systems keep preferring third-party credibility #
Model divergence does not kill the earned-media thesis. It sharpens it.
Muck Rack reported in December 2025 that 94% of all citations in its dataset came from non-paid sources and 82% came from earned media alone, based on analysis of more than one million links cited by leading AI models (Muck Rack, 2025). The same report found only 2% average overlap between the journalists PR teams pitch most and the journalists AI engines cite most for a brand (Muck Rack, 2025).
Fullintel and the University of Connecticut found that 47% of citations in their AI response study came from journalistic sources, while another 48% came from corporate, university, health network, and association websites (Fullintel, 2026). The same piece cites supporting Muck Rack findings that more than 95% of citations were unpaid media and more than 89% of cited links were earned media in that study context (Fullintel, 2026).
Those findings fit together cleanly. Even when models diverge, they still reward credibility outside the brand’s four walls.
The academic GEO record supports multi-format, evidence-dense optimization, not a one-trick playbook #
Aggarwal et al. introduced GEO as a formal optimization problem and showed that optimization methods could improve visibility in generative engine responses by up to 40% (Aggarwal et al., 2024). The same paper found that different optimization tactics perform differently across domains, which is one of the clearest academic arguments against a universal template (Aggarwal et al., 2024).
That matters here because model-specific citation behavior and domain-specific optimization are the same structural story viewed from two angles. AI visibility is conditional.
A brand needs:
- strong first-party pages for models that favor full-control sources
- credible directory and profile coverage for models that rely on some-control surfaces
- review and user-generated proof for models that weight limited-control sources
- earned media and independent coverage for models that still treat third-party validation as the cleanest trust signal
That full system is closer to Machine Relations than to narrow GEO page-tweaking.
AI visibility measurement fails when teams collapse all models into one score #
Yext’s conclusion was direct: the source mix that makes a brand visible in Gemini is not the same mix that makes it visible in Claude (Yext, 2026). That means reporting a single blended “AI visibility” number can hide both strength and weakness.
A useful measurement system should separate at least four things:
| Measurement question | Why it matters |
|---|---|
| Which models cite the brand most often? | Prevents false confidence from blended averages |
| Which source types produce those citations? | Reveals which layer of the system is working |
| Which industries or query classes shift the mix? | Exposes vertical-specific vulnerabilities |
| Which independent publications repeatedly appear? | Informs earned-media targeting and citation architecture |
Without that separation, teams optimize the wrong source layer and then wonder why the next model update wrecks performance.
The Machine Relations implication is simple: brands need a source-portfolio strategy, not a page-portfolio strategy #
Machine Relations is the discipline of making a brand legible, credible, and citable across machine-mediated discovery systems. Yext’s model-level findings show why that frame matters. The challenge is not just publishing better pages. The challenge is building a portfolio of source types that different models can trust for different reasons.
That portfolio includes first-party content, high-quality listings, review surfaces, expert commentary, journalistic placements, and extractable formatting. A brand that treats AI visibility as a single-channel content optimization problem is designing for a market that no longer exists.
The stronger move is to treat model divergence as a routing problem. Different engines prefer different evidence. Brands that win will build enough credible evidence across enough surfaces that no single model preference can erase them.
Search behavior data reinforces the model-divergence thesis #
The model-level evidence sits inside a larger behavioral shift. Gartner said on February 19, 2024 that traditional search engine volume will drop 25% by 2026 as AI chatbots and virtual agents replace part of the classic search journey (Gartner, 2024). Bain reported on February 19, 2025 that 80% of consumers rely on AI-written results for at least 40% of their searches and that organic traffic is down 15% to 25% because answers are being resolved directly in the interface (Bain & Company, 2025).
Pew Research Center’s July 22, 2025 analysis of 68,879 unique Google searches found that users clicked a traditional search result on 8% of visits when an AI summary appeared versus 15% when one did not, and clicked a cited source inside the AI summary only 1% of the time (Pew Research Center, 2025). SparkToro’s 2024 clickstream study found that only 360 of every 1,000 Google searches in the United States produced a click to the open web, while the EU figure was 374 per 1,000 (SparkToro, 2024).
These are not identical datasets, and they should not be collapsed into one metric. But together they make the operating environment obvious: visibility is being decided earlier, inside the answer layer, and under different citation rules depending on the model and interface.
Yext's earlier October 9, 2025 location-context study sharpens the same point from another angle. It analyzed 6.8 million source citations from 1.6 million responses and found that websites accounted for more than 40% of citations for objective unbranded queries across Gemini, OpenAI, and Perplexity, while OpenAI shifted to 46.3% directory reliance for branded subjective queries and Perplexity pushed industry-specific directories to 24% for unbranded subjective queries (Yext, 2025). That is more evidence that model behavior changes with both interface and query frame.
The credibility side matters too. Reuters Institute's Digital News Report 2025 says evidence-based journalism is competing in an environment of declining engagement and low trust, but it remains one of the few content systems built around attribution, updates, and editorial process (Reuters Institute, 2025). That helps explain why journalistic sources keep showing up so heavily in AI citation studies.
Google's May 14, 2024 Search announcement is also strategically important because it explicitly describes a custom Gemini model that summarizes the web, expands AI Overviews, and introduces AI-organized results pages for more complex discovery tasks (Google, 2024). That is not a traffic dataset, but it is first-party confirmation that one of the most important search interfaces is intentionally moving more discovery into answer-layer synthesis.
The Verge's contemporaneous reporting captured the same product shift in plainer language: Google was making search "AI all the way down," using Gemini to both answer and organize result pages (The Verge, 2024). Taken together with the citation studies, that matters because it means source selection is being mediated both by model behavior and by interface design.
Key takeaways #
- Major AI models do not share one citation logic. Yext's 17.2 million-citation study shows persistent model-specific source preferences across sectors.
- Query context changes the source mix. Yext's earlier 6.8 million-citation study found objective and subjective queries trigger different source dependencies.
- User behavior is moving toward answer consumption, not link traversal. Pew, Bain, Gartner, and SparkToro all point in that direction from different datasets.
- Earned media still matters because third-party credibility keeps winning citations even when model preferences diverge.
- Measurement has to be model-specific and source-type-specific or it will hide the real failure point.
What brands should actually do #
| Source layer | Why it exists | Which evidence supports it |
|---|---|---|
| First-party pages | Needed for models and sectors that favor full-control sources | Yext found Gemini leading full-control reliance across sectors and SearchGPT spiking to 38.1% hotel-site citations in hospitality (Yext, 2026) |
| Listings and managed profiles | Important where some-control directories shape discovery | Yext found listings made up 54.53% of distinct cited URLs across the dataset (Yext, 2026) |
| Reviews and UGC | Critical for models that lean harder on limited-control evidence | Yext found Claude’s limited-control citations ran 2-4x higher than competitors across all seven sectors (Yext, 2026) |
| Earned media and independent reporting | Needed because AI systems still overweight third-party credibility | Muck Rack found 94% of citations were non-paid and 82% were earned media; Fullintel-UConn found 47% of citations came from journalistic sources (Muck Rack, 2025; Fullintel, 2026) |
| Structured, evidence-dense formatting | Needed because visibility in generative interfaces is partly extractability | GEO research found optimization could improve visibility by up to 40% and varied by domain (Aggarwal et al., 2024) |
FAQ #
Who coined Machine Relations? #
Jaxon Parrott, founder of AuthorityTech, coined Machine Relations in 2024 as the parent discipline for AI-mediated brand discovery, citation, and recommendation systems.
Is Machine Relations just GEO rebranded? #
No. GEO focuses on optimization inside generative engines, while Machine Relations covers the full system of earned authority, entity clarity, citation architecture, distribution, and measurement.
Where do GEO and AEO fit inside Machine Relations? #
GEO and AEO sit inside the distribution layer of Machine Relations. They matter, but they do not replace the authority, source diversity, and measurement layers required to win across models.
How do AI models decide what to cite? #
The current evidence suggests they do not all decide the same way. Yext’s Q4 2025 dataset found model-specific source preferences, while academic GEO research shows that formatting, statistics, and domain context materially affect visibility outcomes (Yext, 2026; Aggarwal et al., 2024).
For a broader definition of the discipline behind this shift, see Machine Relations and Jaxon Parrott’s Machine Relations framework on Medium.