GEO Lifts AI Citations 37% on Perplexity: 3 Papers Marketers Need
The customer journey is being rerouted through AI, and three new papers give marketers the clearest map we've seen so far. Generative engine optimization — the practice of getting your content cited inside AI-generated answers — now has its first peer-reviewed playbook. At the same time, researchers are naming a bigger shift: the buyer on the other end of your funnel may not be a person at all.
Today's radar covers a Princeton team's KDD 2024 paper defining GEO, a 2026 Journal of Marketing Analytics piece arguing that AI systems are becoming customers themselves, and an arXiv preprint showing how sponsored content can be inserted into any chatbot response without retraining the model.
If you run content, SEO, media buying, or brand strategy, these papers point at the same underlying question: are you still marketing to humans, or to the algorithms making decisions on their behalf? Here's what the evidence supports — and where it stops.
Quick Takeaway
- Adding statistics, expert quotes, and credible citations lifted AI-engine citations by up to 37% on Perplexity in controlled tests.
- What works depends on query type: data-heavy writing wins factual queries; authoritative language wins opinion queries.
- AI agents are moving from assistants to autonomous buyers, meaning product pages must be legible to machines, not just humans.
- Native LLM ad insertion is technically solved in research settings — a preprint shows a plug-in layer works across seven commercial models.
- GEO does not replace traditional SEO. Your page still has to be retrieved before it can be cited.
What This Research Means for Marketers
The short version: AI search and AI-mediated shopping are creating a second layer of visibility that sits on top of, not instead of, your current SEO. The GEO paper gives you concrete content moves — stats, quotes, citations, authoritative tone — that measurably increase citation rates once your page is in the AI's context. The machine marketing paper warns that the reader on the other end may soon be an AI agent, which changes how product data, pricing, and brand claims need to be structured. And the PILA preprint signals that AI-native ad inventory is being engineered right now, before disclosure norms or pricing standards exist.
For most brands, the practical response is not panic — it's a parallel workstream. Keep doing SEO. Add a GEO layer for the pages that matter most. Start making your product data machine-readable. And watch native LLM advertising the way early digital marketers watched banner ads in 1996.
Papers Covered
Paper 1: GEO: Generative Engine Optimization
- Source / venue: ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD 2024)
- Source type: Peer-reviewed conference paper
- Method: Black-box optimization framework testing nine content-level interventions (statistics, citations, direct quotes, authoritative language, technical terms) against a purpose-built benchmark (GEO-Bench) and the live Perplexity.ai engine.
- Sample: Approximately 10,000 queries across nine domains (science, arts, politics, law, and others); real-world validation on Perplexity.ai.
- Main finding: Adding statistics and direct quotes from credible sources increased citation rates by up to 40% in benchmark testing and up to 37% on live Perplexity.ai. Optimal tactics varied by query type — data-heavy content won factual queries, authoritative language won opinion queries.
- Evidence strength: Peer-reviewed, real-world validated, foundational for the field.
- Limitation: Gains were measured after a source was already retrieved into the engine's context. The study does not prove GEO improves crawling, retrieval, organic traffic, clicks, or conversions. Techniques may shift as engines update.
- Practical implication: Audit high-value pages for citation density and add real statistics, external source citations, and expert quotes. Match content style to target query type.
Paper 2: Machine marketing: rethinking the customer in the age of generative AI
- Source / venue: Journal of Marketing Analytics (2026)
- Link: https://doi.org/10.1057/s41270-026-00521-y
- Source type: Peer-reviewed conceptual research note
- Method: Conceptual framework development. No empirical data. Reviews recent generative AI developments and real-world products (Expedia, OpenTable, OpenAI) to propose a research agenda.
- Sample: Not applicable — theoretical paper.
- Main finding: AI tools are shifting from assistants to autonomous agents that shop, compare, and transact on behalf of users. The authors argue a new discipline — 'machine marketing' — is needed because AI decision logic differs from human decision logic.
- Evidence strength: Peer-reviewed but conceptual only; no empirical validation of the proposed framework.
- Limitation: No data collected. The extent to which consumers will delegate purchase authority to AI agents is an open empirical question. Examples used may age quickly.
- Practical implication: Make product pages, pricing, and specs machine-legible. Test how your brand shows up when AI assistants answer category queries. Factor AI-assisted respondents into consumer research design.
Paper 3: PILA: Plug-and-Play Insertion for LLM-native Advertising
- Source / venue: arXiv (Cornell University)
- Link: https://doi.org/10.48550/arxiv.2607.25590
- Source type: Preprint — not peer-reviewed
- Method: System design and empirical evaluation. Fine-tuned Qwen backbone (4B and 8B) trained on a 25,000-sample synthetic corpus, evaluated across seven commercial LLMs (GPT, Claude, Qwen, Gemini, DeepSeek, others) against prompt-based, sampling-based, and fine-tuning baselines.
- Sample: Seven frontier commercial LLMs; 25,000 synthetic training samples; multiple ad-insertion settings including direct chat, tool-use/ReAct, RAG, and multi-agent workflows. Human evaluation panel details not fully specified.
- Main finding: A lightweight sidecar model can insert native ads into any chatbot's responses without retraining the base model, outperforming prompt-based approaches by 34.2%, sampling-based by 47.3%, and fine-tuning by 7.7% on a combined user-quality and ad-effectiveness score. Includes an 'ad intensity controller' for subtle-to-prominent placement.
- Evidence strength: Preprint, synthetic training data, not yet peer-reviewed. Metrics measure exposure and naturalness, not click-through or conversions.
- Limitation: Not tested in real-world deployment with actual users or advertisers. Ethical, disclosure, and regulatory implications are not addressed in the paper.
- Practical implication: AI-native ad inventory is technically feasible today. Publishers and platforms should begin planning for tiered native ad products; brands should start thinking about creative built for LLM insertion.
Plain-English Payoff
AI is no longer just a channel — it's becoming the audience, the shopper, and the media surface all at once. The evidence shows you can measurably improve how often AI engines cite your content by adding stats, quotes, and credible sources. But the bigger shift is that the entity reading your product page may increasingly be an AI agent, and the ad next to your brand may increasingly live inside a chatbot's answer.
Money Move
The most immediate opportunity is a GEO content audit service: score client pages on citation density, statistics count, and authoritative language markers, then rewrite the weakest high-traffic pages to match what the Princeton research shows AI engines actually reward. It's the Yoast model applied to AI search — a defensible new service line while the category is still unnamed by the big players. Agencies that build a repeatable audit template now can sell it as a standalone engagement or bolt it onto existing SEO retainers.
Evidence Check
- GEO paper: peer-reviewed KDD 2024, full text reviewed. Gains measured post-retrieval — the study does not prove pages get found or drive traffic, only that they get cited more once in context.
- Machine marketing paper: peer-reviewed but conceptual. No empirical data. Framework is early-stage and untested.
- PILA paper: preprint on arXiv, not peer-reviewed. Training data is synthetic. No real-world advertiser or user deployment reported.
- PILA ad effectiveness is measured by exposure and naturalness, not by clicks, conversions, or ROI.
- None of the three papers establishes causal impact on business outcomes like revenue, retention, or brand lift.
- Generative engines change quickly. Tactics validated on Perplexity in 2024 may perform differently on ChatGPT search or Google AI Overviews today.
What to Test Next
- Action step. Pick your three highest-traffic pages and count the statistics and externally cited sources on each. If the average is under two, those pages are underweight for AI citation and should be rewritten with credible data and quotes.
- Action step. Run five category-relevant queries through Perplexity, ChatGPT search, and Google AI Overviews. Log which brands get cited and whether yours appears — this becomes your GEO baseline.
- Action step. Audit one product page for machine legibility: structured specs, consistent pricing, clear comparison attributes. Ask whether an AI agent doing comparison shopping could parse it without guessing.
- Action step. If you operate a chatbot or LLM-powered product, brief your engineering team on the PILA architecture so you can evaluate native ad insertion before industry pricing norms are set.
How This Connects to AI Marketing Strategy
These three papers form a single arc: discovery, decision, and monetization are all being reshaped by AI. GEO addresses discovery — how content gets cited when AI synthesizes rather than lists. Machine marketing addresses decision — who is actually evaluating your product when an agent shops on someone's behalf. And PILA addresses monetization — how AI-native ad inventory will be built and priced.
For Big Plans Media's coverage, this is the clearest signal yet that AI marketing strategy is bifurcating into two audiences that must be served in parallel. Human-facing brand work continues. Machine-facing work — structured data, citation-worthy content, agent-legible pricing, LLM-native creative — is the new layer. The brands that treat these as one workstream will underinvest in both.
FAQ
What is generative engine optimization (GEO)?
GEO is the practice of structuring web content so that AI-powered search engines like Perplexity, Bing Chat, and Google AI Overviews cite it in their generated answers. The foundational KDD 2024 paper defines it as a distinct discipline from traditional SEO — one focused on citation visibility inside AI responses rather than link rankings.
How much can GEO tactics actually improve AI citations?
In the Princeton benchmark, adding statistics and direct quotes from credible sources increased citation rates by up to 40%. On the live Perplexity.ai engine, the same tactics produced up to 37% more citation visibility. Note that those gains were measured after the source had already been retrieved into the engine's context.
Does GEO replace traditional SEO?
No. The research is explicit that GEO operates on top of traditional SEO. Your page still has to be crawled, indexed, and retrieved before an AI engine can cite it — that's what traditional SEO does. GEO improves what happens once your content is in play.
What are AI customers, and should marketers care yet?
AI customers are AI systems — assistants and agents like ChatGPT-based shopping tools — that search, compare, and increasingly transact on behalf of human users. The Journal of Marketing Analytics paper argues marketers should start preparing now by making product data machine-legible, though widespread consumer delegation of purchase authority is still an open empirical question.
Is native LLM advertising real, or still theoretical?
Technically, it's real. The PILA preprint demonstrates a working sidecar model that inserts sponsored content into any chatbot's responses without retraining the base model, tested across seven commercial LLMs. It has not been deployed at scale with real advertisers, and it's a preprint rather than peer-reviewed work, but the engineering blueprint exists.
How should small businesses respond to these findings?
Focus on GEO first — it's the most actionable and requires no new tools. Audit your highest-value pages, add real statistics and expert quotes, and cite credible external sources. Then check how your brand appears when someone asks an AI assistant to recommend products in your category. Those two moves are low-cost and directly evidence-supported.
What are the risks of native LLM advertising for brand trust?
The PILA paper does not address disclosure or regulation, which is a meaningful gap. If sponsored content is inserted into AI answers without clear labeling, consumer trust in AI recommendations could erode quickly. Brands considering LLM-native ad buys should push for disclosure standards before, not after, the format scales.
Listen to the Episode
Sources and Further Reading
- GEO: Generative Engine Optimization
- Machine marketing: rethinking the customer in the age of generative AI
- PILA: Plug-and-Play Insertion for LLM-native Advertising
Related Big Plans Media:
About Big Plans Media
Big Plans Media helps marketers, educators, entrepreneurs, consultants, and business leaders translate AI marketing research into practical strategy. AI & Marketing Research Radar is produced by Big Plans Media and hosted by Evita, an AI-generated research briefing avatar trained on Dr. Eva Wolf's research framework.
