AI search has no appeals process

Contestability is the ability to see why you were selected and to challenge that decision, and answer engines do not provide it: there is no ranking to inspect, no query attached to your impression, and no route to appeal an answer that gets your product wrong.

That is not a complaint about customer support. It is a structural property of the channel, and a position paper accepted to the ICML 2026 Position Track argues it is the risk the GEO industry has been least honest about (arXiv:2606.12439).

#The mechanism search had, and answers removed

The ranked list was itself a disclosure. You could see who outranked you, for which query, and form a testable theory about why. Search Console tied an impression back to the query that produced it. That closed a loop: diagnose, change the page, watch the position move. When the loop failed, there was still a documented reconsideration path for manual actions.

The ICML paper defines contestability as the capacity of affected parties to understand and challenge how recommendations are selected, and argues that synthesized answers break it. A user cannot observe why certain options appeared or which alternatives were dropped. The authors make a further point that matters for anyone who has tried to interrogate a chatbot about its own sourcing: asking the model to explain itself does not restore contestability, because the answer and its justification are produced by the same opaque pipeline (arXiv:2606.12439).

Google has partly acknowledged the gap. In June 2026 it added Search Generative AI performance reports to Search Console, showing how often URLs from your site appeared in generative AI features such as AI Overviews and AI Mode (Google Search Central). That is presence data. It tells you that you appeared. It does not tell you why, against whom, or what to change, and it is not a lever you can pull.

#Nothing was ruled against you

The absence of an appeals process is easier to accept once you see how unstable the selection actually is.

The same paper runs a sensitivity test across 30 query pairs and seven OpenAI and Gemini models. Minor wording changes produced different citation sets. For the Gemini models, every query pair changed its cited domains after paraphrasing, and one model frequently cited an almost entirely different set of domains (arXiv:2606.12439).

The mechanism is mundane. Retrieved context is assembled with a hard top-k cutoff. A small change in retrieval score can push a source across that boundary, which changes what evidence enters the context window, which changes the answer. Visibility moves in steps, not gradients.

This should reframe how you read a bad result. A missing citation is usually not a penalty and not a judgment. It is a boundary condition on a scoring function you cannot see, triggered by phrasing you did not control. There is no appeal available because there was no ruling.

#The channel is manipulable from the outside, too

The flip side is worse, and it has now been measured. A separate study, SafeGEO, built 22 attack variants and ran them across 600 product recommendation scenarios to test whether seller-controlled rewrites of source pages can corrupt a recommendation agent's decisions.

They can. GEO-style rewrites raised the placement of flawed products in top-three recommendations by up to 83.2%, and pushed constraint-violating items into the top slot in up to 59.3% of cases (arXiv:2606.28356). Defenses helped, but not enough: the strongest developer-side mitigation tested reduced harmful promotion by up to 39.2% and still did not restore baseline behavior (arXiv:2606.28356).

Read that as a competitive fact rather than a security headline. Someone else rewriting their own pages can change the answer your buyers see about your category, using tactics that are indistinguishable from ordinary content work. You have no notification, no diff, and no counterparty to contact.

#Why the published research will not warn you in time

The paper's third risk is the one that explains a lot of the confusion in this field. It calls it an academic-industry blind spot, built from two asymmetries.

The first is visibility. Academic work evaluates static benchmarks and synthetic prompts; industry optimizes against live systems that adapt continuously. Effects that only appear in deployment, such as feedback-driven dominance and market steering, are invisible to the offline setup.

The second is evaluation. Academic papers report offline ranking metrics like Recall@k and nDCG@k, while practitioners optimize citation frequency and answer visibility. The authors argue that modest benchmark improvements can still meaningfully raise the probability of being mentioned or cited in a real response, producing outsized commercial effects (arXiv:2606.12439).

That gap is why a careful paper reporting a small offline gain and a vendor reporting a large visibility gain can both be telling the truth. They are measuring different objects. Neither is measuring the thing you are accountable for.

#What recourse actually looks like

Since you cannot appeal an answer, treat the source layer and the measurement layer as your only two controls.

Correct sources, not answers. In-product thumbs-down feedback goes to the vendor's evaluation pipeline, not to a record about your brand. The durable fix is publishing a dated, plainly worded, retrievable page that states the correct fact, and then making sure the third-party pages the engine already trusts carry the same fact. You are changing the evidence pool, because that is the only part of the pipeline you own.

Measure persistence, not presence. The ICML authors propose deployment-aligned metrics: citation rate, domain citation share, citation persistence over time, and query-level bootstrap confidence intervals. Their suggested protocol is concrete. Sample a stratified query set, run it against deployed engines on a fixed schedule, for example daily for two weeks with weekly follow-ups, and log every answer and citation. They estimate the cost of that black-box auditing at $50 to $300 using public APIs (arXiv:2606.12439).

Paraphrase on purpose. If a reworded query returns a different domain set, then a prompt set with one phrasing per intent is measuring phrasing, not visibility. Carry several natural phrasings of the same buyer question and track them together.

The uncomfortable conclusion is that GEO is not a discipline with a feedback loop yet. It is a discipline with an evidence pool you can influence and an outcome you can only observe. Companies that internalize that will spend their effort on sources and on honest longitudinal measurement. Companies that do not will keep looking for a support form that nobody has built.