For 25 years the web ran on attention. Get the click, run the funnel, win the user. That game is ending. Buyers now ask an AI agent to do the searching for them, and the agent decides who to mention. Your job moved from earning clicks to earning citations.
I build AI-native products and consult on AI adoption for a living. Most of my week sits on the other side of this equation: wiring agents into retrieval pipelines and watching them pick which sources to trust. The selection logic I see inside those systems also runs the public-facing surfaces of ChatGPT, Claude, Gemini, and Perplexity. This is the playbook I hand founders, brands, and solo operators asking where do I start.
A peer-reviewed academic foundation for this exists now, plus several large-scale industry studies, so the high-leverage tactics are knowable. Most online advice still treats GEO as recycled SEO with a fresh coat of paint. The moves that earn citations look different.
From the Attention Economy to the Interpretation Economy
The framing I find most useful comes from marketer Nate B Jones. The Attention Economy rewarded shouting: eyeballs, impressions, funnel mechanics. In the Interpretation Economy, the agent reads on the buyer's behalf and quotes whoever it can parse and trust.
Humans increasingly don't search for products or talent directly. They ask an AI agent to do it for them. The web is now filtered through what an AI thinks about you.
That filter has two consequences. First, the model flattens fuzzy positioning. A value proposition that isn't structured and specific gets averaged into the “internet average” for your category, which is the same as invisible. Second, the work now splits into two surfaces you have to optimise for at the same time:
- The agent-facing internet. Structured, evidence-dense, technically clean, so an LLM can parse you and quote you.
- The human-facing internet. The “offline wedge” of real-world memory, preference, and trust. A human who remembers your name asks the AI for you by name, and the model's comparison filter never runs.
Both matter. Most of what follows is about the first surface, because it's the part you can engineer.
How LLMs decide what to mention
Cited answers come from two pathways, and most operators optimise for one.
- Parametric knowledge. What the model “memorised” during training. A Nectiv study of 8,500+ ChatGPT prompts found that only ~31% trigger a live web search. The remaining ~69% of answers come straight from training data, where Wikipedia is over-represented: ~0.6% of raw training tokens, ~3% of weighted training data for GPT-3.
- RAG retrieval. A live search dispatched at query time. The model fetches 8–10 candidate pages, ranks them, and quotes 3–6. Perplexity leans hardest on retrieval, ChatGPT mixes both, Claude routes through Brave Search.
The engines also have different taste in sources, and cross-platform overlap is low. Ahrefs found ~11% of cited domains shared between ChatGPT and Perplexity for the same query. “Optimise for AI search” is the wrong unit of work. Pick the engines that matter for your category and treat each one as its own surface.
ChatGPT
Heavy on Wikipedia, mainstream news, structured authority sites. Bing index pages carry weight: 87% of ChatGPT Search citations match Bing's top-10 organic for the same query. A page Bing hasn't indexed is a page ChatGPT Search won't quote.
Perplexity
The most community-driven of the four. Reddit held ~46.7% of top-10 citations through 2024–2025; 2026 data shows YouTube overtaking it as the #1 social citation source (16% vs ~10%). Perplexity also favours review platforms: G2, Trustpilot, Capterra, Clutch.
Google AI Overviews / Gemini
High-authority editorial domains plus YouTube and Reddit. The wrinkle worth knowing: ~48% of Overview citations come from URLs outside Google's top-10 organic results. A #1 ranking on Google no longer guarantees a quote.
Claude
Cites long-form blog content at high rates (one study attributed 43.8% of Claude citations to it) and rewards honest trade-off acknowledgments with a ~1.7x citation boost for “here are the limitations” language. Hedging earns citations on Claude where it gets penalised on most other engines.
Off-site beats on-site: the finding that changes the playbook
Ahrefs studied ~75,000 brands in May 2025 to find what correlates with AI Overview visibility:
- Brand web mentions correlated 0.664 with AI Overview visibility.
- Backlinks correlated 0.218.
- The top three correlations were all off-site factors.
Muck Rack's Generative Pulse team analysed 25 million AI-engine links and found 84% of citations came from earned media. Owned content and paid placement accounted for the remaining 16%.
Build a Truth Layer, not a marketing site
Jones calls the antidote a Truth Layer: a verifiable, structured, granular body of evidence about what you do. The data backs him. AI agents map facts to a buyer's question; adjectives have no anchor point and get discarded. The Truth Layer takes the same shape whether you're a brand, a product, or a solo operator like me:
- Specific outcomes with numbers. “Reduced onboarding time by 40%” survives retrieval. “Industry-leading” reads as filler and gets dropped.
- Named, attributable claims. Attach every claim to a person, a date, and a source. Floating claims rarely make it into a quote.
- Traversable identity. Author bio → LinkedIn → Wikidata → conference profile. LLMs verify people by walking these chains, and a thin byline is an active liability.
- Structured metadata. JSON-LD,
sameAsarrays, schema for Person and Organization. The cheapest move on the list, and one of the highest-leverage for identity resolution.
Contently's 2026 data: sites whose author identity is present on four or more platforms appear in ChatGPT responses 2.8× more often. Treat identity as a retrieval input.
What to write, ranked by measured impact
Princeton's peer-reviewed GEO: Generative Engine Optimization paper (Aggarwal et al., KDD 2024) tested nine content tactics across 10,000 queries. Six produced reliable lifts in AI visibility.
1. Add original statistics (+41%)
The biggest content lever the paper measured. A single page with one quoted, sourced number outperforms three pages of opinion. You don't need a 10,000-respondent survey. A 200-person sample, a 30-day benchmark, or a small audit of your own data is enough to make you the canonical source for that fact.
2. Publish (or get into) a “Best X for Y” listicle
Glen Allsopp's December 2025 Ahrefs study (26,283 URLs, 750 prompts) attributed 43.8% of all ChatGPT citations to blog-style “best X” lists. Evertune corroborated with an analysis of ~400 million LLM citations: 63% point to listicles, 71–86% in Top-N format. Two facts that should reshape your content calendar:
- 79.1% of cited lists were updated in 2025. Freshness compounds.
- 35% of cited lists sit on low-authority domains. You don't need Forbes.
3. Add named expert quotes (+37%)
Quotes from a real, attributable person, with a name and a role, give LLMs a clean span to lift. Source these via HARO, Qwoted, or Featured.com. One quotable sentence per article is enough.
4. Cite credible outbound sources (+31.4%)
Linking out to .gov, .edu, or peer-reviewed studies raises your own citation rate. The model reads your page as part of a credible chain. Most operators avoid outbound links because they're afraid of leaking authority. They're costing themselves citations.
5. Coin a named framework
A short, ownable label like “Truth Layer,” “Prove-It Economy,” or “Value Equation” gives the model a specific string to recall and attribute. Naming is also how a concept crosses the “parametric” threshold: once enough articles repeat your label, it gets baked into the next round of training.
6. Use the answer capsule pattern
A Search Engine Land audit found 72.4% of pages cited by ChatGPT contain a 20–25 word self-contained answer placed under a question-format H2. No jargon, no links inside the answer. One extractable sentence.
Same audit: LLMs pull 44.2% of all citations from the first 30% of the document. Lead with your most citable claim instead of burying it under preamble.
The silent killers (engineer's checklist)
Before any of the content work matters, the bots have to be able to read your site at all. The #1 reason brands get zero AI citations is mechanical: blocked or unreadable pages. Three checks before you write another word.
1. Your robots.txt
Most default robots.txt files predate these crawlers. Allow:
OAI-SearchBot: ChatGPT Search citationsChatGPT-User: ChatGPT live fetchingClaudeBot: Anthropic / ClaudePerplexityBot: Perplexity real-time retrievalGoogle-Extended: Gemini (block this one only if you want to opt out of training)Bingbot: ChatGPT Search's underlying index
2. JavaScript-only rendering
OAI-SearchBot does not execute JavaScript. A client-mounted SPA with no SSR or static export is invisible to ChatGPT Search. SSR, SSG, or a pre-rendered HTML fallback is now a hard requirement for any page you want quoted.
3. Slow servers
Crawler fetch timeouts run around 2 seconds. A slow first-byte page never makes it into the candidate pool. The same Core Web Vitals work you'd do for Google is now a prerequisite for AI citation too.
Schema and llms.txt: where the hype outruns the data
Two tactics get a lot of hype right now. Both are worth doing. Neither is the lever some posts make them out to be.
- llms.txt. A Markdown file at your site root that lists your key pages, proposed by Jeremy Howard in September 2024. Anthropic has adopted it; OpenAI, Google, and Perplexity have not confirmed they use it. An OtterlyAI study of 62,100 AI bot visits found no measurable citation lift from llms.txt alone. Ship it anyway. 30 minutes of work, zero downside.
- JSON-LD schema. Ahrefs ran a controlled study of 1,885 pages adding schema against 4,000 controls and found no significant citation lift. AI Overview citations dropped 4.6%, the only statistically significant result they observed. Treat schema as identity hygiene. It earns its keep by making your Person and Organization entities resolvable, which feeds the Truth Layer.
GEO levels the playing field more than SEO does
The Princeton paper buries the most useful single statistic I've seen for anyone starting from zero. Pages ranking around Google position #5 saw up to ~115% AI visibility gain from GEO tactics. Position #1 pages barely moved. The mid-ranked page has the most to gain from being structured, evidence-dense, and citable. The front-runner has the most to lose by ignoring it.
Combined with the Allsopp finding that 35% of cited listicles sit on low-authority domains, one implication holds for almost any category: out-structure the incumbents, and the engines start quoting you whether you outrank them or not.
Six moves to start, in priority order
Ordered by impact per hour.
1. Unblock the bots and add an answer capsule (~2 hours)
Audit your robots.txt, allow the six crawlers above, and add a question-format H2 with a 20–25 word answer capsule under it on your three most important pages. This single structural change matches what 72.4% of ChatGPT-cited pages already do.
2. Publish one piece of original data this quarter (~1 day)
A 100–200 person survey, a 30-day comparison, an audit of your own usage data. One specific number you own, like “67% of freelance designers raised prices in 2025,” makes you citable in a way no opinion post can match. This is the +41% Princeton lever.
3. Get featured in (or publish) a “Best X for Y” listicle (~1 week)
Search for best [your category] for [your audience] 2026. Identify the top 10 listicles AI engines cite (ask Perplexity, it will tell you). Pitch the authors with a one-paragraph value prop and one verifiable customer result. In parallel, publish your own honest comparison that names competitors fairly. Neutrality wins citations.
4. Build your Truth Layer (~2–4 hours)
Three concrete moves: (a) a clear About page with a one-sentence definition of what you do and who you serve; (b) an author bio with photo, real credentials, LinkedIn link, and a Person JSON-LD block; (c) a claimed and completed profile on the top review site for your category. Make your identity verifiable and traversable.
5. Seed authentic answers in 3–5 communities (1–2 hours/week, ongoing)
Reddit appears in ~92.8% of AI search opportunities. YouTube is now the #1 social citation source overall. Find the threads where your buyers ask the questions your product answers. Contribute useful, detailed replies. No self-promotion for the first 4–6 weeks. One well-upvoted, specific Reddit answer can generate multiple AI citations for months.
6. Earn 1–2 third-party mentions per month (~2 hours/week, ongoing)
HARO, Qwoted, Featured.com. Respond to two relevant queries a week with a specific, data-backed, quotable sentence. Muck Rack: 84% of AI citations come from earned media. A trade-press quote with a number compounds in retrieval indexes for years; a blog post on your own domain rarely lasts a quarter.
Three caveats
- GEO doesn't replace SEO. The top Princeton tactics (cite sources, add statistics, add quotes) are also classical SEO best practices. Pages that earn AI citations tend to earn more featured snippets too. Keep doing SEO.
- The numbers drift. Citation-share percentages (Reddit, YouTube, listicles) shift month to month as engines retrain and re-rank. Treat any specific percentage in this article as directional.
- Accuracy is imperfect. A Tow Center and Columbia Journalism Review audit found 60%+ incorrect answers across AI search platforms. Monitor what AI says about you. Check what gets cited and whether the citation is right.