Your GEO lift assumes nobody else rewrites

A GEO lift is the change in how visible a page is inside AI-generated answers after you rewrite it, measured against the same answer engine and the same competing sources before the rewrite.

That last clause is where most published numbers quietly break. The standard experiment rewrites one document and leaves every competing document as it was. Your competitors do not work that way. They read the same studies, buy the same tools and apply the same checklists. A growing body of research now measures what happens when they do, and the answer is consistent: the lift you were promised was a single-player number.

#Where the headline numbers come from

The paper that named the field, GEO: Generative Engine Optimization by Aggarwal and colleagues, reported that optimization "can boost visibility by up to 40% in generative engine responses." Tactics such as adding statistics, adding quotations and improving fluency became the default GEO playbook on the strength of that benchmark.

The setup behind that number is a reasonable first experiment. For each query there is a pool of sources, one of them is rewritten, and the rest stay as they are. Visibility is then measured by how much of the generated answer credits the rewritten source. The result tells you what one optimizer gains against a field that is not optimizing. It does not tell you what anyone gains once optimization is common.

#What happens when the field optimizes too

C-SEO Bench (Puerto and colleagues, NeurIPS 2025 Datasets and Benchmarks track) was built to test that second case. It covers two tasks, question answering and product recommendation, across three domains each, and adds an evaluation protocol with varying adoption rates among the competing documents. The authors found that most current conversational SEO methods were "largely ineffective" and frequently had a negative effect on ranking. They also observed that as the number of adopters grew, "the overall gains decrease, depicting a congested and zero-sum nature of the problem."

A July 2026 critical survey of GEO research summarized the benchmark's scale: roughly 1,900 queries and 16,360 documents, with only three of 54 method and domain combinations significantly positive in the main experiment, and none positive in question answering. The survey's framing is blunt: when impression shares sum to one, GEO is at least locally redistributive. Visibility one source gains is visibility another source loses.

The most direct test so far is Beyond the Vacuum, submitted on August 27, 2026. The authors built a competitive version of GEO-bench in which each query has five documents: one target and four competitors. They then varied the share of competitors that also rewrite their content, from 0 to 0.8 in steps of 0.2, using a mix of competitor strategies that includes random combinations, quotation addition and an automated GEO method.

Three findings from that paper matter for anyone running GEO work.

Single tactics collapse under full competition. The authors report that single-strategy methods at the highest adoption rate "drop below the unoptimized" baseline. A page that adds quotations while every rival also adds quotations ends up worse off than a page in a field where nobody rewrote anything.

The vacuum rankings are close together. In the uncompetitive setting, measured with a position-weighted word count metric, the best single strategies scored 23.44 (conciseness), 23.18 (LLM guidance) and 23.08 (statistics addition), against 20.03 for no rewrite at all (Table 1). Tactics separated by a fraction of a point in isolation are not a stable basis for a content program once the field shifts.

Even the best adaptive method loses ground. The authors' own competitor-aware method, which picks combinations of strategies based on what rivals have done, scored 32.62 with no competitors optimizing and declined 11.4% across the full adoption range (same paper). That was the slowest degradation of any method tested, and it still degraded.

#Why the gains cancel

The mechanism is not mysterious. An AI answer has a fixed amount of room. It cites a handful of sources and gives each some share of the text. GEO tactics work by making one source look more quotable than its neighbors: a crisper sentence, a number, an attributed quote. When every neighbor gets the same treatment, the relative advantage disappears, and what remains is a pool of documents that all look alike to the model.

Researchers at ETH Zurich described the adversarial end of this dynamic in Adversarial Search Engine Optimization for Large Language Models. They showed that crafted content could push production LLM search engines (Bing and Perplexity) to promote one product and discredit rivals, and that this "leads to a prisoner's dilemma, where all parties are incentivized to launch attacks, but the collective effect degrades the LLM's outputs for everyone." Ordinary GEO rewrites are not attacks, but the incentive structure is the same shape: each party is better off rewriting, and everyone together is no better off than before.

#What the research does not say

These results come from benchmarks, not live engines. Beyond the Vacuum uses LLM-generated responses from gpt-oss-120b and Llama-3.3-70B-Instruct as a proxy for production answer engines, and its authors state that their competitor simulation uses synthetic distributions rather than observed real-world adoption. Their own method also has an information advantage, since it reads the full competing corpus before choosing a rewrite, and they report a drop in faithfulness compared with existing baselines. You should not read the exact percentages as forecasts for ChatGPT or AI Overviews.

The direction is still worth acting on, because it is the same direction every benchmark with competitors points: gains shrink as adoption grows. C-SEO Bench also found that traditional SEO strategies, the ones that improve a source's position in the model's context, were significantly more effective than the rewrite tactics. That is a hint about where durable advantage lives.

#What to do with this

Stop quoting single-player lifts in plans. If a vendor or a study promises a percentage visibility gain from a rewrite tactic, ask whether the competing documents were also optimized. If they were not, treat the number as an upper bound on a quiet market.

Measure against your actual competitors, repeatedly. A before-and-after test on your own page tells you little if your rivals rewrote in the same quarter. Track the same query set, the same engines and the same competitor domains over time, and report your share of citations relative to theirs rather than an absolute score.

Put effort into what cannot be copied by a rewrite. A tactic anyone can apply to any page is a tactic that cancels out. A first-hand dataset, a tested procedure, a product fact only you can state, or a page that ranks in the retrieval step because it is the best answer to the query are harder to match. These are also the inputs C-SEO Bench found more effective than surface rewrites.

Watch for convergence in your category. If the pages cited for your priority queries start to read alike (same structure, same stat-heavy intros, same FAQ blocks), the formatting advantage is already gone. The next citation goes to whichever source adds something the others lack.

GEO is still worth doing. The research simply moves the question from "does this tactic work" to "does it still work when everyone uses it." For most checklist tactics, the honest answer is less than the headline, and sometimes not at all.