Every GEO checklist opens with the same line: add statistics, cite sources, quote an expert. It is the founding claim of the entire field — and on the engines people actually use today, two independent 2026 studies could not find the effect.
In November 2023, a team from Princeton and Georgia Tech published the paper that gave the industry its name. "GEO: Generative Engine Optimization" tested nine content tactics against early generative search engines and found that the right combination — adding statistics, citing sources, quoting an expert — could lift a page's visibility by up to 40% (arXiv). That single number became the founding claim of an industry. Three years later, two independently built studies went back and re-ran the experiment on the engines people actually use in 2026. Neither one found the effect.
Read that number literally and it is not wrong — it is dated. The 2024 KDD paper tested how a language model rewrites its answer once it already has a source's text in hand. It did not test whether the model chooses to retrieve that page in the first place, because the retrieval systems behind ChatGPT, Perplexity and Google AI in 2026 work differently, and largely did not exist in their current form when the original experiments ran.
A September 2026 preprint by Elisha Bajemon and André-Louis Rochet took the original paper's published 2023 effect sizes — its "causal anchors" for adding statistics, citations and quotes — and re-measured them on ten modern engine families: six open-weight, four commercial, including three GPT-5.x variants. None of the levers moved citation (arXiv). Across the same ten engines, a deterministic 11-part content score modeling information density, citability and statistic density correlated with actual citation at a Spearman coefficient of 0.11 — close enough to zero that it explains roughly 1% of the variance. Even a deliberate attempt to game the score, tested adversarially against 500 sources, gained an attacker at most 6 points out of 100, with diminishing returns per added statistic.
A second, separately conducted study — SAGEO Arena, accepted to KDD 2026 — went further and built something the 2023 benchmark never had: a full search pipeline, retrieval and reranking included, running over 171,003 real web documents across nine domains. Eight of its ten tested tactics were taken directly from the original 2023 paper, now run under conditions closer to how ChatGPT or Perplexity actually retrieve a page before writing an answer (arXiv). Editing body text with injected statistics and authority phrasing did not just fail to help. In the worst case tested, it cut a page's retrieval visibility by 36% and dropped it 22 average ranking positions.
What worked, in the same study, was the boring stuff: titles that name the actual entity instead of a generic label (a supplement guidance page's "Panel" becoming "IOM Panel Vitamin D Guidance"), meta descriptions that describe the page precisely, and schema markup — plain technical SEO, not GEO. Those edits lifted retrieval visibility by 22% on average.
If a GEO consultant's pitch is "we add statistics and expert quotes to your pages to get you cited," ask which study they are working from. The 2023 number is still the one circulating in sales decks, and it was not fabricated — it measured something real, just not the thing that decides whether your page gets pulled up today. For a small business, the practical order of operations flips: before touching the body copy of a page, check whether its title tag names the exact thing a customer is asking about, whether the meta description states it precisely, and whether basic schema markup is in place. That is an afternoon of work, not a content rewrite — and per this data, it points in the opposite direction from where most GEO checklists send you.
None of this means statistics and expert quotes are worthless on a page — readers still want them, and a page that answers a question badly will not get cited no matter how clean its schema is. What the data says is narrower and more useful: stuffing them in specifically to chase an AI citation is not a lever worth pulling, and overdone, current evidence says it can work against you.
We test the structural signals this data says actually move visibility — titles, meta descriptions, schema — against the questions your customers ask. No account, no card.
run_check →