GEO (generative engine optimization): a practical primer

Most marketing acronyms are born in a pitch deck. GEO — generative engine optimization — was born in a research paper: a late-2023 study by researchers at Princeton, Georgia Tech, the Allen Institute for AI, and IIT Delhi, who asked a measurable question. If AI assistants answer queries by reading web pages and citing a few, what can a page do to get cited more? Then they tested it, across a benchmark of 10,000 queries.

That makes GEO unusual: the core advice comes with experimental results attached, not vibes. This article is the practical version of those results — what worked, what backfired, and what it looks like applied to an actual small-business page. One naming note before we start: you'll also see the term AEO, answer engine optimization. For practical purposes the two mean the same thing — getting your business cited when AI tools answer people's questions. GEO is the researchers' name for it, AEO the marketers'. If you're new to the whole topic and want to know whether it even matters for your business yet, start with our plain-English explainer and come back here for the tactics.

What the study actually did

The researchers built generative search engines in the style of Bing Chat and Perplexity, ran thousands of real-world queries through them, and measured how prominently each source appeared in the AI-written answers. Then they modified source pages using nine different tactics — one at a time — and measured whether the modified pages earned more presence in the answers. Visibility here isn't rankings; it's how much of the final answer is built from, and attributed to, your page.

What worked: evidence density

The winning tactics cluster around one idea — make the page read like a source, not a sales pitch:

• Adding statistics — concrete numbers in the text. The strongest single tactic tested, lifting a page's answer visibility by as much as 40%.
• Adding quotations — direct quotes from named people or organizations, similar-sized effect.
• Citing sources — pages that name where their claims come from get quoted more themselves.
• Fluent, clear writing — plain readable prose measurably outperformed clunky text. The machines share your readers' taste.

And the tactic that backfired: keyword stuffing, the one move everyone knows from old SEO folklore, scored below doing nothing at all. Generative engines are summarizers — repetition gives them nothing to quote.

Every winning GEO tactic is a form of the same instruction: give the machine a sentence worth repeating and a reason to trust it.

The finding everyone skips: it depends on the domain

The study's least-quoted result may be its most useful: no tactic won everywhere. Citations helped most for factual and legal-style topics, statistics for debate-shaped questions, quotations for people-and-culture topics. The practical reading for a business: match the evidence type to the question your page answers. A 'how much does X cost' page wants numbers. A 'is X safe / legal / required' page wants cited authorities. A 'what's it like working with X' page wants quoted humans.

Applying it: one paragraph, before and after

Here's what the tactics look like on the kind of sentence every service business has. Before: 'We offer fast, professional water heater installation at competitive prices.' Nothing there is quotable — no fact, no number, no attribution. An AI assembling an answer about water heater replacement can do nothing with it.

After: 'A standard tank water heater replacement takes our two-person crew 2 to 3 hours and costs $1,400 to $2,100 installed, including permit and disposal of the old unit. Tankless conversions run higher — $3,500 to $5,500 — because they usually require upsizing the gas line.' Same page, same service. But now there are four concrete figures, a stated timeframe, and a causal explanation an answer engine can lift whole — with your business attached to it.

That's the entire craft, repeated across your important pages: replace adjectives with measurements, claims with sources, and generalities with the specific numbers only you know. It's the same move that separates AI-drafted content that ranks from AI-drafted content that doesn't — evidence density is the common currency of human readers, Google, and answer engines.

A page-level GEO pass, in order

• Pick pages that answer questions, not pages that describe you. Answer engines respond to queries; your homepage rarely answers one.
• Put a self-contained answer directly under a question-shaped heading — summarizers lift passages, not pages.
• Add the numbers only you have: your prices, timelines, project counts, results. Generic industry stats are what every competitor's AI draft already says.
• Attribute something: a named customer quote, a cited standard or regulation, a linked source for any claim you didn't originate.
• Read it aloud once for fluency — clunky prose measurably underperforms, and the fix is free.

What GEO can't do

Keep the ceiling in view. The study modified pages that were already in the answer engine's source pool — GEO tactics amplify a page that's findable, they don't make an invisible page visible. Getting into the pool is still ordinary SEO. Citation patterns also shift when models update, so treat any specific 'AI cited us' win as weather, not climate. And none of it replaces the question that decides whether any of this matters: whether the visitors you do get become customers.

But as new-channel bets go, this one is cheap. Every GEO tactic above — concrete numbers, real quotes, cited claims, clear prose — makes the page better for human buyers even if no AI ever reads it. Optimizing for answer engines mostly means writing the way you should have been writing anyway, which is the most reassuring thing the research could have found.