GEO has a real evidence base and a large grift economy sitting on top of it. Separating them is easy once you look at what has actually been tested.

Start with the foundational research. The Princeton GEO paper, published at KDD 2024, tested nine content modifications across 10,000 queries and found visibility gains of up to 40 percent in generative answers. The winners were specific: citing sources, adding statistics, and adding quotations lifted relative visibility by roughly 40, 37 and 22 percent.

The same study buried two findings the industry likes to skip. Keyword stuffing did nothing, and even performed worse than baseline on Perplexity. And lower-ranked pages gained the most, with position-five sources seeing visibility gains over 100 percent, which makes GEO a leveling mechanism rather than a rich-get-richer game.

Follow-up research adds a corrective. C-SEO Bench, the first systematic benchmark of conversational-SEO tactics, found most of them do not help and several hurt, while plain source relevance keeps working. Other 2026 experiments show token-level tricks failing to beat unmodified pages. Translation: there is no magic phrasing. There is being the best source.

So here is the working list, each item with evidence behind it.

Extractable, direct answers with real data in them: that is the Princeton result, and it matches how content depth and readability affect chatbot citations. Entity clarity, meaning consistent names, identifiers and descriptions of who you are and what you sell, because engines retrieve entities, not keywords.

Freshness, since about 65 percent of AI citations come from content under a year old. Third-party presence, because UGC and earned media dominate what models cite and reviews measurably correlate with citations. And it pays: brands cited in an AI Overview earn 120 percent more clicks per impression on the same query.

Now the snake oil, in descending order of popularity.

llms.txt sold as a ranking lever: Google says it is not used for any AI Search feature, and crawler logs show AI bots rarely fetch it. It has narrow legitimate uses, covered honestly here, but it is not visibility.

Mass-generated "GEO pages" targeting thousands of AI prompts: that is scaled content abuse with a new name. Secret AI keyword densities and per-engine phrasing hacks: contradicted by the benchmarks above. Hidden text and prompt-injection tricks aimed at models: adversarial, detectable, and the fastest way to get filtered. Guaranteed-citation packages: nobody controls a probabilistic answer engine, so a guarantee is the tell.

One meta-trap: the terminology itself. GEO, AEO, AIO and SEO get sold as separate disciplines needing separate retainers, when they are mostly one discipline with different surfaces.

One thing that also works: testing instead of believing. Pick one topic cluster, apply the evidence-backed changes (sources, statistics, direct answers, fresh dates), leave a comparable cluster untouched, and re-run your target prompts across ChatGPT, Perplexity and AI Mode a month later. Citations either moved or they did not. Every claim in this space should survive that experiment, and most of the paid ones do not.

The straight summary: what works is being a fresh, specific, verifiable source that the rest of the internet corroborates. What is snake oil is anything promising visibility without doing that.