Do GEO Tactics Stop Working When Everyone Uses Them?

Every ranking of GEO tactics you have read shares an assumption: your competitors' pages hold still while you optimize yours. A preprint posted on 27 August 2026 by Vaibhav Sourirajan and colleagues at Capital One AI Foundations drops that assumption and re-runs the question. When the share of competing documents that are also optimized rises to 80%, every fixed single tactic they tested scores below the baseline of never having optimized at all.

Key facts

Key takeaways

Updated 11 September 2026: added key facts and the CHASE results table. The CHASE simulation itself was added on 1 September 2026.

On this page: The missing question · What the paper did · The adoption experiment · Where is your category? · What repeated rounds do · The accuracy cost · Limits · What to do

Since the KDD 2024 GEO paper reported that optimization "can boost visibility by up to 40%," the field has treated its tactic list — add statistics, cite sources, add quotations, improve fluency — as a ranked menu. The ranking came from a specific experimental shape: one document in a retrieval pool gets rewritten, the others stay as they were, and the change in the engine's answer is measured.

That was a reasonable design in 2023, when almost nobody was optimizing. It describes a land grab. The Capital One paper asks what the same tactics are worth once the land is taken, and its answer is that the ranking itself is unstable: the value of any edit depends on how many competing documents already carry it.

The intuition is familiar from every prior optimization discipline. If one page in five has extractable statistics and yours is it, the engine's preference for statistics is your advantage. If four in five do, the same edit buys parity at best. What the paper adds is machinery for measuring this and a demonstration of how sharp the falloff gets.

The method has two phases. First, for each query and document set, it runs Bayesian Optimization of Combinatorial Structures — BOCS — over the 2^15 = 32,768 possible on/off combinations of 15 rewriting strategies drawn from the published GEO literature, using 50 evaluations per search. The 15 include the KDD 2024 tactics (statistics addition, citing sources, quotation addition, fluency, authoritative tone, keyword stuffing) alongside later additions like structured formatting and LLM guidance from C-SEO Bench.

Second, because running BOCS per page is impractical, the observations from those searches become training data: preference pairs and reasoning traces that fine-tune a language model to look at a document corpus and propose a strategy combination directly.

Visibility is scored with PAWC — Position-Adjusted Word Count — the standard GEO-research metric that credits a document for the words the answer draws from it, discounted by position. On GEO-Bench with gpt-oss-120b as the engine:

The gap worth noticing is not the headline one. It is that the best single tactic — the one at the top of most practitioner checklists — sits five points below an agentic rewriter and nearly ten below combination search. The checklist's unit of action, the individual tactic, is the wrong granularity even before competition enters.

Then competition enters. The authors build geo-bench_comp, a variant in which a fraction α — the share of competing documents that are themselves GEO-optimized — is rewritten by simulated optimizers, and test α from 0 to 0.8 — from an unoptimized field to one where four of five competitors are running GEO.

Two results carry the paper. Averaged across adoption rates, the competitor-aware method scores 29.93 PAWC against 25.20 for the strongest baseline — an 18% edge. And at the high end, the fixed approaches do not merely lag: "single-strategy methods at full competitive optimization (α=0.8) drop below the unoptimized (α=0) baseline." Running the standard playbook in a saturated category simulated as worse than never having run it.

The competitor-aware selector avoids this by doing the thing its name says: it infers the adoption level from the documents it can see and shifts "toward less saturated alternatives as the competitive landscape is increasingly optimized." Its total degradation across the full adoption range is 11.4% — it bends where the fixed tactics break. The method also transferred out of distribution, scoring 37.78 on an e-commerce set and 33.48 on Researchy-GEO, more than 8% above the strongest baseline on each.

For anyone doing answer engine optimization, the mechanism matters more than the scores. Nothing about a statistics block or an FAQ became less legible to the engine at high α. What changed is that the edit stopped being a difference. Engines cite what stands out from the pool; an optimization everyone applies is, by definition, no longer what stands out.

The paper does not say, because it cannot — its competitors are synthetic. But the measurement exists elsewhere, and reconciling the two is where this becomes usable.

The CISPA-led GEO-Flag study, posted ten days earlier, put the first number on real-world adoption: 8.90% of pages retrieved for 1,000 real queries were GEO-optimized, rising to 16.36% among pages modified in 2026 — we covered it in our analysis of GEO prevalence. Against the simulation's axis, the measured web sits near α = 0.1, one page in six for recent content. The regime where fixed tactics go negative starts around α = 0.8. On the web average, the 2024 playbook is still operating in something like the vacuum it was measured in.

But the average is doing heavy lifting. The same study found 0% detected optimization on Wikipedia and major medical sites and 20% on sampled Amazon pages, and its year-over-year trend — 7% of 2024-modified pages, 12.8% for 2025, 16.4% for 2026 — points one direction. A competitive commercial category is plausibly years ahead of the mean, and it is precisely the categories where AI answers drive purchases that fill fastest.

That yields a decision rule the paper stops short of stating. The question "do GEO tactics work" has no fixed answer; the question "how optimized are the pages cited for my prompts" does, and it determines the first one. Pull the citations for your tracked prompts, assess how many are visibly running the playbook, and read your position off the curve: a quiet pool means the standard tactics still buy differentiation, a crowded one means they buy parity, and the trend line tells you how long the first state lasts.

Added 1 September 2026. The Capital One paper asks what saturation does to your tactic. A second simulation posted on 31 August — CHASE, by Qianwen Gao and colleagues, accepted to COLM 2026 — asks what it does to everyone: it runs 20 rounds in which creators repeatedly adapt documents toward features extracted from recent ranking winners, across six domains (retail, video games, books, web, news, debate), with roughly 29% of non-winning documents rewritten each round.

Before running it, the authors validate the thing most GEO research quietly assumes: that ranking is a fair proxy for citation. In their measurement, a document's rank predicted whether it was cited in grounded generated answers with an AUC of 0.853 ± 0.093 across the six domains. That number is useful on its own — it says rank-shaped optimization and citation-shaped optimization are close to the same game, which is why results like the adoption curve above transfer to the citation outcomes brands actually track.

The core result: quality–ranking alignment declined in every domain, with the change in Spearman's ρ ranging from −0.018 to −0.107, mean −0.068 over the 20 rounds.

The control condition is what makes it attributable: documents rewritten just as often, but toward arbitrary targets rather than ranking-derived ones, showed little or none of the decline. The drift comes from chasing the signal, not from rewriting per se.

Two honest readings the paper's title undersells. Alignment eroded but stayed positive at round 20 in every domain — after 20 rounds of exploitation the ranker still weakly favored better documents. And the homogenization in the acronym is measurable but small: mean pairwise embedding similarity within a query pool moved by a point or two (video games went from .334 to .348), and the retail pool actually got slightly more diverse. This is a direction of travel, not a collapse — the same reading discipline the adoption curve deserves.

Put the two papers together and the saturation story gains a second axis. As your category's α rises, the Capital One result says any fixed tactic buys less; CHASE says the referee's ordering drifts away from quality at the same time. The compounding conclusion for a brand is the one this post already reached from the faithfulness table: differentiation that survives is the kind no rewrite reproduces, because rewriting toward the signal is precisely the behavior that degrades the signal.

The paper logs this as a trade-off in a table. For a brand it is the whole risk, and it deserves more than a footnote.

The winning method's rewrites scored 4.90 on faithfulness against 8.29 for the unrewritten documents on the paper's LLM-judged scale. The visibility-optimal version of a page drifted furthest from what the page originally said. A rewrite that wins the citation by saying something your source material does not support becomes an engine confidently attributing that claim to you — and unlike a ranking, you do not get to see it happen.

Four independent groups now converge on this. The mechanism-design work on citation wars found unsupported claims accumulating over repeated optimization rounds. The GEO-Flag audit found decoration-grade citations on 69% of the pages it detected as optimized. This paper measures the faithfulness cost directly. And the CHASE result above is the population-level rhyme: the same pressure decouples the ranking signal from quality for everyone at once. Pushed hard enough, visibility optimization manufactures unsupported claims.

The operational consequence is the one thing on the Monday list that is not a measurement task: read every optimized rewrite against its source, and treat a claim that appears only in the optimized version as a claim you did not make until the optimizer made it for you.

The caveats here are structural, not fine print.

The competitors are simulated — three optimizer functions applied on a fixed rotation — and the engines are proxies: gpt-oss-120b and Llama-3.3-70B-Instruct, not the production systems answering buyers' questions. PAWC measures words drawn into an answer, which is an impression proxy, not a mention, a recommendation, or a sale. And the competitor-aware method holds what the authors concede is "an information advantage": it reads the entire competitive corpus before choosing a strategy. No publisher sees the retrieval pool the way this selector does — though tracking which pages engines actually cite for your prompts is the closest available approximation.

The CHASE simulation carries its own caveats in the same family: 20 queries per domain with 5–10 candidate documents each, an LLM ranker (Gemini Flash-class) standing in for a production engine, and quality judged by another LLM whose agreement with human raters was moderate (Spearman's ρ of 0.58 on shared dimensions). The authors are explicit that the framework "intentionally isolates repeated adaptation to an LLM ranking signal rather than modeling a complete generative-search system."

There is also a quieter irony worth holding: a paper about competitors adapting to each other is itself now public. A tactic selector that wins by finding "less saturated alternatives" describes an equilibrium that moves again the moment enough parties run one — and parties will: a demo system automating exactly this kind of feedback-driven tactic selection, Agent2UCB, was posted two days later.

The honest summary is that GEO tactics work the way arbitrage works: the return is real, and it is a function of how many others have found the trade. The KDD 2024 measurements described an empty field, the CISPA measurements say the field is one-sixth full and filling, the Capital One simulation shows the far end where the standard moves pay less than nothing, and CHASE shows the trade eroding the price signal itself as the field fills. The practitioner's job shifts accordingly — from running a checklist to knowing where on that curve their category sits, which is a measurement question before it is a tactics question.

Elmo is an open-source, self-hosted AI visibility platform that runs your prompt sets across ChatGPT, Claude, Gemini, Perplexity, and Google's AI surfaces and records every answer with its citations and competitor mentions. Because the cited pages for each prompt are in your own database, the adoption-rate check this research calls for is a query, not a project — and re-running it each month tells you when your category's curve starts to bend.

For the fundamentals, start with what AI citations are and where they come from. For the adjacent research, see how much of the web is GEO-optimized and what happens when platforms fight back. For the practice, the AEO guide and how to track your brand in AI search. For the vocabulary, the AI search glossary.