acceptodds
Withdrawn from ICLR 2027

OFF-GEO: Mitigating Off-Target Visibility in Generative Engine Optimization

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Generative Engine Optimization (GEO) methods use large language models (LLMs) to rewrite source content and improve its visibility in generative engine responses. However, these gains can extend to queries that a content provider does not wish to target. Mitigating these undesired gains is challenging because desired and undesired queries may rely on overlapping source content, so reversing GEO edits can affect both. We propose OFF-GEO, a framework that learns a shared editing rule for GEO-optimized sources. We optimize the rule to maximize desired visibility subject to a minimum average reduction in undesired visibility. Candidate rules are compared using visibility changes measured from generative engine responses, while feedback from edits that fail to reduce undesired visibility guides rule refinement. The learned rule guides edits to new optimized sources without query inputs or generative engine feedback. Experiments on E-commerce and GEO-Bench show that OFF-GEO reduces undesired visibility more than desired visibility across all metrics, while keeping desired visibility above that of the original sources. Despite removing portions of GEO-optimized content, OFF-GEO also maintains competitive answer quality on desired queries

Primary area:
foundation or frontier models, including LLMs
Keywords:
generative engine optimization, generative engines, information retrieval, query relevance

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_U2EEOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper is about a side effect of generative engine optimization (GEO). When a source is rewritten to get cited more often by a generative search engine, it also tends to get cited for unrelated queries that the content owner never wanted to target. The authors call this off-target visibility. It is hard to remove because the target and non-target queries often rely on the same text, so simply deleting the optimized parts can hurt the target queries too. Also, the edits have to be made offline before any user query is known.

Their method, OFF-GEO, learns one shared, document-agnostic editing rule from engine feedback. Training data are documents paired with a desired query and a related undesired query. An LLM proposes candidate rules based on failure cases, and a constrained successive-rejects procedure picks the rule that keeps the most desired visibility while meeting a minimum reduction in undesired visibility. At inference, the learned rule is applied to new sources, with no test queries or engine feedback needed.

They evaluate on the AutoGEO E-commerce benchmark and on GEO-Bench. On E-commerce, undesired visibility drops by about 10.8 points while desired visibility drops by about 6.9, and on GEO-Bench the drops are about 7.2 and 6.3. Answer quality on desired queries stays close to the unpruned GEO baseline.

Critical strengths
  1. A clear formulation of the problem. Most GEO work only tries to push visibility up for the target queries. Treating off-target leakage as a constrained optimization (keep as much desired visibility as possible, subject to a minimum cut in undesired visibility) gives a clean way to talk about selective visibility control.

  1. No test-time queries or engine feedback. In a real publishing setting the content is edited long before anyone searches, so a rule that can be applied offline is practical. The learned rule is applied through plain text edits, and answer quality on the desired queries holds up on E-commerce. Note that it still needs an LLM editor call per source.
Critical weaknesses
  1. The baselines are weak and the method is never ablated. Table 1 only compares OFF-GEO with the original text and with AutoGEO_API, so I can't tell whether a simple alternative would do as well. I'd want to see easy baselines like zero-shot LLM pruning or length/similarity-based filtering. I'd also like to see a multi-query method such as IF-GEO adapted to penalize passages that trigger the non-target query. On top of that, the method has several moving parts: the learned rule, the selection procedure and a prompt-level "orthogonality guard" that throws away edits touching the source's core topic. The authors say the guard complements the rule, but never remove it. Without that, I can't tell whether the retention comes from the learned rule or just from the guard.

  1. The GEO-Bench margin is thin and the statistics are missing. On GEO-Bench the net gap is only about +0.9 points, and it costs more than a third of GEO's desired-visibility gain (39.8 down to 33.5). The paper doesn't give the number of train or test examples, and Table 1 has no standard deviations or significance tests, so with a stochastic LLM engine I can't tell a gap this small from noise. I'm also unsure about the filtering: sources with zero undesired visibility under AutoGEO_API are dropped. The filter doesn't depend on OFF-GEO, so it isn't unfair to it, but keeping only sources with high baseline leakage can make the measured reductions look bigger than they are because of regression to the mean.
Questions for authors
  1. How many train and test examples are there for each dataset, and what share of sources is removed by the zero-visibility filter?

  1. What happens to the results if the orthogonality guard is removed?

  1. Can you report standard errors or confidence intervals, and a paired significance test, for the GEO-Bench gap in Table 1?

  1. Can you compare against simple pruning baselines and an adapted IF-GEO?

  1. The threshold gamma is 0.02, but the observed reductions are 7 to 10 points. Does this constraint ever actually bind during selection?

  1. Pairs with cosine similarity above 0.6 are excluded, yet that is where passage conflicts should be strongest. How does OFF-GEO do in that range?

  1. Why is answer quality only evaluated on E-commerce and not on GEO-Bench?
Reviewer_e35WOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper argues that GEO can raise a source's visibility on queries the provider does not want to target, and proposes OFF-GEO to reduce this. OFF-GEO learns a shared editing rule, through LLM critiques of failure cases and a constrained Successive Rejects selection, which guides an LLM editor to revert part of the GEO edits. On E-commerce and GEO-Bench with AutoGEO-API rewrites and Gemini 2.5 Flash-Lite as the engine, it reports that undesired visibility drops more than desired visibility, by 3.88 and 0.94 points on average.

Critical strengths
  1. The paper raises an interesting question of whether GEO gains spill over to untargeted queries.
  2. The method is simple and transparent, and the full prompts are given.
Critical weaknesses
  1. The setup doesn't seem realistic: It is unclear why a provider would want less visibility. The undesired queries are mostly unrelated, eg Habitat for Humanity donations vs. protecting endangered species (Table 2), and the source is put directly into the engine's context without retrieval. Original sources are as visible on the unrelated query as on their own (22.57 vs 20.01 on E-commerce), so the metric barely reflects relevance in this setup.
  2. No baselines and no ablation of the learned rule: Only Original and AutoGEO are compared. Simple baselines would be helpful, eg reverting a random subset of GEO edits, or the B.4 edit prompt with an empty rule, since that prompt already tells the editor that the rewrite has been found to "leak into unrelated topics". The final rule is a single sentence, and the optimizer is essentially ProTeGi with a constraint.
  3. Gains are small: On GEO-Bench, the gap is 0.94 points, close to a uniform reduction, with no variance, significance tests or number of test sources.

Other Minor Weaknesses:

  1. Test filtering can inflate the gap: Only sources with non-zero AutoGEO undesired visibility are kept (lines 289-291). If the engine is not deterministic, this biases AutoGEO's undesired visibility upward. AutoGEO's desired visibility on GEO-Bench is also 39.84 here vs 34.92 in the AutoGEO paper. Results on the unfiltered test set would be helpful.
  2. Only one engine is used.
  3. The "has-room subset" (line 431) is not defined.
  4. The optimizer LLM and the cost of rule learning are not reported.
Questions for authors

Please see the weaknesses above. In addition, for the undesired query, which other documents are given to the engine?

Reviewer_VfgTOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes OFF-GEO to alleviate the visibility improvement of non target queries caused by GEO optimization. The author studied shared editing rules to reduce non target visibility while preserving information and visibility on target queries, and validated the effectiveness of the method on two datasets.

Critical strengths
  1. The problem has practical significance and is clearly defined. The paper focuses on the improvement of non target query visibility caused by GEO optimization, and formalizes the problem of "reducing non target visibility and preserving target visibility" as a constrained optimization problem, which has clear research value.

  2. The method has certain practicality. OFF-GEO learns editing rules that can be reused across documents, without the need to access future queries or generative search engine feedback during the inference phase.

Critical weaknesses
  1. The paper does not provide source code or an anonymous repository. Although the prompts and experimental details are included, the implementation of the editor, verifier, visibility evaluation, and constrained candidate selection cannot be independently reproduced.

  2. Undesired query is constructed based on embedding similarity rules and lacks manual annotation or real business objectives, which cannot fully prove that it represents the true off target intention.

  3. The experiment relies on a single editor, a single generative engine, and a single judge, lacking cross model validation.

  4. Lack of important ablation, such as random deletion, direct recovery of GEO new content, fixed rules, and versions that do not use Successful Rejections, makes it difficult to confirm that the benefits come from rule learning itself.

Questions for authors
  1. How to verify that the semantic and information requirements of a desired query are indeed different from those of an unsigned query? Currently, only the MiniLM cosine similarity threshold of 0.6 is used for screening. Is there any manual annotation or threshold sensitivity experiment to support this construction method?

  2. There are too few comparison methods. Since the GEO bench dataset was used, why not compare its proposed optimization strategies?

Reviewer_YdnyOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper focuses on generative search optimization (GEO). In GEO, website owners rewrite the webpage content such that a generative search engine will prefer this page. The paper makes the point that the changes could extend to non-targeted queries as well. That is, queries that the website owner did not wish to optimize for. The paper proposes a method for selective optimization. Under this method, the webpage content is iteratively optimized. The webpage admin starts from a set of queries that they wish to optimize for, and a set that they do not wish to optimize for. The method then iteratively updates the webpage to optimize for the positive and negative sets.

Critical strengths
  1. With increasing adoption of LLM-based search, generative engine optimization is a very timely problem.
  2. The paper is generally easy to follow.
Critical weaknesses
  1. The paper assumes that the positive and negative sets are provided. This creates two problems: First, the users can pose the same query but with different phrasing. It is unclear how the method proposed here would cover this rephrasing. Second, the webpage admins might want to optimize for whole topics instead of queries. For instance, BBC Sports might want to optimize for any queries related to football. They do the optimization for current games and tournaments. A new tournament might come (e.g., FIFA World Cup 2030). They will have to optimize all over again for the tournament. How does the paper handle these cases?
  2. In general, the paper says very little about the exhaustiveness of the queries. How well can the positive and negative set cover the distribution of the queries in the wild?
  3. The paper does not discuss how the optimization for a generative model A (Gemini) carries over to generative model B (Claude). Different models work in different ways so may not prioritize sources in a similar manner.
  4. In a similar vein, different models might use different harnesses, that is, they might call the web search tools in different configurations. It is unclear how the results would generalize over harnesses.
  5. The paper impoliticly assumes that a large number of calls can be made to the generative search engine while performing the rule optimization. For closed source search engines like Google, this may be an unrealistic assumption due to rate limits. How does the paper get around this issue?
  6. A key parameter of Algorithm 1 can be explained better. How is L selected? Is it the same for all types of queries and webpages?
  7. It would be great if at the very beginning (e.g., Introduction), the paper provided a feeling to the readers about what the positive and negative sets could look like. Without knowing concrete examples, the readers would find it difficult to get a sense of what the method is trying to achieve.
Questions for authors

Please see weaknesses 1-6.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.