What Actually Works in GEO, by Strength of Evidence

GEO advice has a credibility problem: most of it is plausible, confident and unproven. This chapter sorts the known tactics by the strength of the evidence behind them, using three labels. Measured means tested in a controlled experiment. Observed means backed by large-scale correlational data. Practice means mechanically sensible and widely adopted, but not proven. You should spend effort in roughly that order.

Measured: the founding experiment

The KDD 2024 paper that named the field ran the only widely cited controlled experiment: take a source page, rewrite it one way at a time, and measure how visible it becomes in generated answers across 10,000 realistic queries. Headline result, in the authors’ words: these methods “can boost visibility by up to 40%”, with the best methods improving on baseline by 41% on the paper’s primary visibility metric.

What won: evidence, added visibly.

• Add quotations. Relevant quotes from credible sources, woven into the text. The single best-performing change.
• Add statistics. Concrete numbers where vague claims stood, with sources.
• Cite sources. Link and attribute the claims you make.

All three are the same move in different clothes: give the engine material it can quote and verify. A generative engine is, functionally, a writer on a deadline. Content that hands it a checkable fact wins over content that hands it an adjective.

What also helped, more modestly: fluent, readable prose; plain, easy-to-understand wording; an authoritative tone; and domain-appropriate technical vocabulary. Clarity is not a nicety; it was measured as a visibility lever.

What failed: keyword stuffing. The tactic scored below baseline on the primary metric, and on Perplexity it performed about 10% worse than doing nothing. Repeating words at a system that reads for meaning is now measurably counterproductive.

The finding almost nobody quotes: GEO favors the underdog. In the paper’s ranking analysis, adding citations lifted the visibility of sources sitting fifth in the search results by 115.1%, while the top-ranked source, on average, lost 30.3% when everyone optimized. Generative answers redistribute attention downward far more than classic search ever did. If you are not the incumbent, this is the most encouraging number in the field.

The honest caveats. The experiments ran on the engines of late 2023, built around GPT-3.5-era models, against the paper’s own benchmark queries, and effects varied by domain: statistics helped most where numbers matter, quotes where people and explanations do. Treat the paper as strong evidence about the direction of these levers, not as a promise of a specific percentage for your site in 2026.

Observed: where large-scale data should sit, and why this section is short

This tier is for large-scale correlational findings: millions of real citations analyzed for patterns. Such studies exist, and here is our problem with them: nearly every one is published by a company selling a GEO or SEO product, ours included. Vendor studies choose their samples, revise their own headline numbers, and are marketing assets before they are research. So this guide applies a rule we would want any source we cite to apply: no vendor marketing studies, from anyone, including us.

Two observed patterns are still worth stating, because you can verify both yourself in an afternoon with the audit in chapter 7: pages that rank well get retrieved and cited more often than pages that do not, and the set of sources an engine cites for a topic churns visibly from month to month. Treat both as directions confirmed by your own data, not as precise industry percentages.

When we publish our own answer-stability numbers, they will appear in this tier with the full methodology attached, so you can check them the way this guide checks everyone else.

Practice: sensible, unproven, cheap enough to do anyway

These have a clear mechanism and professional consensus behind them, but no controlled experiment. Do them; just do not book the results in advance.

• Answer first, then elaborate. Put a complete, standalone answer in the first two sentences under a question-phrased heading. Engines lift passages; make yours liftable.
• One idea per paragraph. A paragraph that needs its neighbors to make sense cannot be quoted alone.
• Keep your facts consistent everywhere. Same product claims, same numbers, same phrasing across your site and profiles. Engines cross-read; contradictions read as unreliability.
• Structure for machines as well as people. Clean heading hierarchy, real tables for comparisons, FAQs for questions people actually ask, structured data (schema) describing what the page is.
• Be present where engines look things up. Directories, review sites, community discussions and reference pages feed both training data and retrieval. Off-site mentions are slow to build and hard to fake, which is precisely why they carry weight.
• Show your dates. Visible publish and update dates, honestly maintained, signal that a page is being kept true.

How to spend a finite budget

Reading the three tiers together, a priority order falls out:

1. Make your key pages quotable: claims backed by numbers, quotes and cited sources. This is the measured 40%.
2. Keep ranking for the queries that matter; retrieval still runs through search, so rankings remain your ticket into the engine’s source pool.
3. Apply the practice-tier hygiene as you touch each page, not as a separate project.
4. Measure continuously, because answers reshuffle with every model update, and because without measurement you cannot tell which of the above actually moved your number. That is chapter 5.