Why Do AI Search Agents Keep Following a Bad Query? Path Inertia in GEO

When an AI deep-research run fails to find a brand, the website is not always the cause. The agent may have produced an unhelpful first query and then kept defending that path because it was the agent's own choice.

This failure develops during search, not only in the final response. A brand-mention score alone cannot distinguish missing content, query drift, and self-reinforcing tool use.

Inertia caused by action ownership

The August 24, 2026 paper *From Inertia to Objectivity* calls this inertia bias: a model is less willing to change course when judging a query, plan, or intermediate conclusion that it previously generated than when judging identical information as a neutral observer.

The paper's IBIS benchmark keeps the task and search observations constant and varies only whether the model “owns” the preceding query. On items where re-searching was correct, switching to observer mode improved success by roughly 15% to 30% for almost every tested model. The final benchmark included 245 Should Re-search and 209 Should Visit Page items. Of decisions that flipped between modes, 91.9% moved in the inertia direction: the observer re-searched while the acting model remained on its path.

These findings come from controlled search agents and research benchmarks, not brand-exposure tests of ChatGPT, Gemini, or Perplexity. The paper also treats the underlying mechanism as a black box, so one percentage cannot describe every AI search product.

Diagnose the path, not only the answer

Retain the first query, search snippets, visited pages, and rewrite points as well as the final answer. For a brand absence, compare three runs: preserve the original path; ask an independent reviewer, shown only the task and results, whether to visit or re-search; and restart from a different but equivalent initial query.

If an independent path consistently reaches official evidence while the original run keeps opening irrelevant pages, the failure looks more like a trajectory problem. If several reasonable paths fail to find a page or the page cannot support the claim, then content, accessibility, or authority is a stronger diagnosis.

Classify the breakdown: an overly narrow first query; relevant-looking snippets without an answer; an early conclusion polluting later decisions; or final validation re-approving its own reasoning. Each failure requires a different intervention and should not automatically become a request to write more content.

GEO Radar at https://www.georadar.top can retain fixed questions, cross-platform answers, and historical differences, helping teams find anomalies that deserve a path-level retest. It cannot expose every internal query or browsing step in closed products, so inertia should remain a diagnostic hypothesis unless the trajectory is observable.

Sources for this article

• arXiv, August 24, 2026, *From Inertia to Objectivity: Improving Deep Research Agents with Noise Isolation*: https://arxiv.org/abs/2608.23045
• arXiv HTML full text, including IBIS construction, actor/observer controls, results, and limitations: https://arxiv.org/html/2608.23045v1