How AI Engines Decide Which Brands to Cite: Evidence and Mechanisms

Abstract. Generative engines — ChatGPT, Perplexity, Google AI Overviews, Gemini, Claude and Copilot — now sit between buyers and vendor websites, and the behavioral data is unambiguous: in a Pew Research Center study of 68,879 Google searches, users clicked a traditional result on 8% of visits when an AI summary appeared, versus 15% without one [1]. Visibility is migrating from rankings to answers. This paper explains how answer selection works inside generative engines, summarizes the peer-reviewed and industry evidence on what moves it, and sets out a measurement-led program for brands. Its core claim: citation is a content property, not an accident — it can be engineered and monitored.

1. Background: from ranked lists to written answers

Two 2025 datasets quantify the shift. Pew Research Center analyzed the browsing behavior of 900 U.S. adults across 68,879 Google searches in March 2025: 18% of searches produced an AI summary, and those pages were far more likely to be triggered by question-shaped and longer queries — 53% of queries with ten or more words produced one [1]. Ahrefs, analyzing 300,000 keywords with aggregated Google Search Console data, found the click-through rate of the top-ranking result 34.5% lower where an AI Overview appeared (April 2025), and approximately 58% lower by December 2025 [2].

The traffic rationale for classic SEO is eroding at the same time as a new surface — the answer itself — has become the decision point. In this environment the unit of competition is no longer position but mention: whether the brand is named, how early in the answer, and described how accurately.

2. Scope and method

This paper synthesizes three classes of evidence:

1. Peer-reviewed research — principally the GEO-bench study from Princeton University, IIT Delhi and the Allen Institute for AI, published at KDD 2024 [3], and follow-on work from Carnegie Mellon University [4].
2. Large-sample industry measurement — Pew Research Center’s behavioral dataset [1] and Ahrefs’ keyword and brand studies [2][5].
3. Our monitoring practice — daily neutral prompt sets run across six engines, which we use to check whether published findings generalize to live engines.

Every figure is cited to its source, and vendor studies are marked as such. We do not claim access to engine internals; the selection logic in Section 3 is inferred from observable inputs and outputs.

3. The selection pipeline

Despite different interfaces, generative engines follow a broadly similar pipeline:

1. Understanding the question. The engine rewrites the user’s prompt into one or more search-style queries, adding context the user never typed — location, product category, budget signals.
2. Retrieval. Those queries run against a search index or the engine’s own crawl corpus. Only documents that rank well enough for the rewritten queries reach the next stage; retrieval failure is invisible and unrecoverable.
3. Selection and compression. The model reads the retrieved documents and decides which sources support each claim. Everything else is discarded, regardless of rank. Citation slots per answer are scarce — a handful of sources survive out of dozens retrieved.
4. Composition. The answer is written, and the surviving sources become the visible citations — when citations are shown at all.

The practical consequence: ranking in classic search is necessary but not sufficient. A page can rank first and still never be cited if the model cannot extract attributable claims from it.

4. Findings

F1 — Content changes move visibility, measurably

The GEO-bench benchmark (roughly 10,000 queries; nine content-modification strategies; multiple generative engines) found that adding quotations, statistics and citations to source pages lifted their visibility in generated answers by up to approximately 40% in relative terms, with quotations and named statistics among the strongest levers; the one legacy tactic tested — keyword stuffing — performed negatively [3]. Complementary work from CMU found that cooperative, semantically explicit content earned 35–60% higher citation rates than terse or adversarial equivalents [4]. The direction is consistent: engines reward sources that make claims easy to verify and extract.

F2 — Corroboration outweighs size

Ahrefs’ study of 75,000 brands found that branded web mentions correlated 0.664 with AI Overview presence — a stronger association than backlinks or domain rating — and that brands in the bottom half of web-mention volume were largely invisible in AI answers [5]. Being described consistently across independent third-party sources matters more than site authority.

F3 — Visibility and traffic have decoupled

Pew found users clicked a source inside the AI summary itself in just 1% of visits, and ended the browsing session after 26% of summary pages [1]. Ahrefs found even tenth-position results lost roughly a fifth of their clicks [2]. Increasingly, being cited in the answer is the entire exposure — the click cannot be assumed to follow.

F4 — Engines disagree, so averages mislead

In our daily monitoring, the same neutral prompt regularly produces different cited brands across ChatGPT, Perplexity and AI Overviews, and the differences move over time. This is consistent with the GEO-bench authors’ observation that strategy efficacy varies across queries and engines [3]. A single-engine audit is therefore structurally misleading: it measures one sample of one distribution.

F5 — Freshness matters, selectively

Engines weight recency for time-sensitive questions and tolerate older dates for stable definitions. Blanket “update everything” advice wastes effort; freshness investment should follow the questions where staleness changes the answer.

5. Implications for brands

• Treat the answer as the landing page. The brand description inside an AI answer does the persuading that a title tag used to do.
• Invest in extractable claims and named sources. These are the levers with published effect sizes behind them [3][4].
• Grow consistent third-party descriptions. Corroboration across independent sources is the strongest brand-level association with AI visibility [5].
• Measure per engine, per prompt, over time. Single audits conflate engine differences with real movement.
• Do not expect clicks to follow citations. Report visibility as its own outcome [1][3].

6. Limitations

Benchmarks measure relative visibility inside controlled experiments, not pipeline revenue; the mapping from visibility to business outcomes is still undocumented. Vendor studies carry disclosure asymmetries — methodologies and sample frames differ, and engines change faster than studies replicate. The figures cited here are directional and dated; none should be read as a permanent constant. Our own monitoring data is observational, drawn from six engines and our clients’ categories.

7. Recommendations

1. Build a stable prompt set from the questions your buyers actually ask engines; freeze it before changing anything on the site.
2. Audit key pages for extractable claims and entity clarity — phrase content so passages can stand alone.
3. Identify the third-party sources cited in your category and treat their editorial standards as the benchmark for your own content.
4. Re-run the prompt set daily across engines; track mention rate, position and cited sources as separate trends.
5. Review the visibility trend alongside classic search metrics, and expect the two to diverge.

References

1. Pew Research Center, “Google users are less likely to click on links when an AI summary appears in the results,” July 2025. pewresearch.org
2. Ahrefs, “AI Overviews Reduce Clicks by 34.5%” (April 2025) and December 2025 update (300,000 keywords, Google Search Console data). ahrefs.com/blog/ai-overviews-reduce-clicks
3. Aggarwal, P., Murahari, V., Rajpurohit, T., Kalyan, A., Narasimhan, K., Deshpande, A., “GEO: Generative Engine Optimization,” KDD 2024. arXiv:2311.09735
4. Chen, et al., “What Generative Search Engines Like,” Carnegie Mellon University, 2025. arXiv:2510.11438
5. Ahrefs, “An Analysis of AI Overview Brand Visibility Factors (75K Brands Studied),” 2025. ahrefs.com/blog