Interesting new research paper from the Capital One AI foundations team on competitor-aware GEO.
The problem#
A shitty race between you and your competitors.
As long as something becomes a best practice, you and your (worthwhile) competitors are all aware of it. At which point the tactics are zero-sum, the gains shrink to nil.
This is the definition of a race to the bottom of marginal gains. And it’s absolute crap as a practitioner. The rewrites literally make my soul shrivel. Reminds me of the early days of black hat SEO where programmatic keyword stuffing was the rage.
Relative not checklist#
The paper proposes that AEO/GEO visibility should be treated as relative in context of competition vs applying a fixed checklist for all content. This is a fairly straightforward and intuitive idea tbh.
The test: combination of 15 levers in a competitor-aware way.#
Tactically, they turn GEO into a search problem: for one query, one target page, and the other pages in the set, choose which of 15 rewrite levers to turn on so your page gets more of the AI answer (measured mainly by Position-Adjusted Word Count — PAWC — basically how much cited text is yours, with earlier sentences weighted more).
Fancy combinatorial, benchmarking, and distillation aside, I thought the 15 levers were SUPER interesting:
The 15 levers#
From earlier GEO work (Aggarwal / AgenticGEO)#
- Technical terms — more domain terminology; same facts
- Keyword stuffing — insert up to ~10 relevant SEO keywords not already in the text (i hate this personally but hey trying to communicate facts)
- Statistics addition — add a few verifiable numbers/stats inline
- Fluency — smoother sentences and transitions; same length/structure
- Unique words — more precise vocabulary; no new claims
- Authoritative — more confident/expert tone via wording only
- Quotation addition — add a few short attributable quotes (no inventing)
- Citing sources — add a few natural-language citations to credible refs
- Easy understanding — simpler language; same facts
From C-SEO Bench#
- LLM guidance — prepend an LLM-friendly markdown summary, then keep the full original text unchanged after it
Added in this paper (from AutoGEO-style heuristics)#
- Expanding facts — briefly explain how/why behind key claims (inline, not long new sections)
- Structured formatting — reorganize with tables, bullets, numbered lists where it helps (same claims)
- Balanced view — add brief counterpoints / other viewpoints inline
- Conciseness — cut filler and repetition; keep all core claims
- Pros and cons — add clear pros/cons and comparative points to support recommendations
The result#
On geo-bench and the competitive geo-bench_comp benchmark, their selector beat all 15 single-lever baselines plus AutoGEO and AgenticGEO on PAWC and related citation metrics (e.g. ~32.6 vs ~28.0 PAWC for AgenticGEO on gpt-oss-120b).
Under rising competitor GEO adoption it also lost the least visibility (~11% drop from α 0→0.8) and stayed ahead zero-shot on held-out E-Commerce and Researchy-GEO sets.
Your takeaway#
Look at 2–3 pages already cited for a particular prompt. Note which levers they already overuse. Bias your rewrite toward underused ones for that set — e.g. if everyone has quotes/stats, lean structure, concrete facts, comparison, pros/cons, or concision instead of more of the same. Not “always use tables.”