AI citations are graded on a curve

AI citation is a relative judgment: an answer engine picks its sources by comparing the candidates already sitting in its context window against each other, not by scoring your page against a fixed quality bar.

That sounds like a distinction for researchers. It is actually the difference between a GEO checklist that works and one that produces polished pages nobody cites.

#What the experiment showed

A March 2026 paper by Roman Smirnov trained a small language model with reinforcement learning to rewrite search snippets, then measured whether those rewrites changed which sources an LLM Overview system selected (arXiv 2605.00012). The setup restricted the policy to snippet text only and limited reward hacking, which keeps it closer to the constraints a real publisher operates under. It drew roughly 3,000 samples from the Amazon Shopping Queries dataset, with seven to ten search results per query (arXiv 2605.00012).

The paper states its central result without hedging: LLM Overview preferences are not absolute but relative, which it describes as generating the biggest value inside a subset of allowed values (arXiv 2605.00012).

Read that as a scoping statement. The unit of optimization is not your page. It is your page sitting beside the other results the engine happened to retrieve for that specific query.

#Some slots never open

The same paper reports a persistence pattern worth sitting with. Across 90 test samples run through seven experimental variations, 48 cases had the same snippet selected every single time, and 30 cases had two or more snippets that were never selected at all (arXiv 2605.00012).

Roughly half the candidate sets had a locked-in winner that no amount of snippet rewriting displaced. A third had candidates that were structurally out of contention. The contested middle, where a rewrite actually changes the outcome, is smaller than the GEO advice market implies.

This is why per-query measurement beats per-page scoring. A page that is invisible for one query can be the fixed winner for a neighboring one, and the page did not change between them. The candidate set did.

#Your competitors are one of your ranking factors

A July 2026 critical survey by Olivier Martinez reviewed 45 studies published between November 2023 and July 2026 and reached a compatible conclusion from a completely different direction (arXiv 2607.14035). It argues that GEO is not a single ranking task but a stochastic, partially observable pipeline running through search activation, crawling and indexing, retrieval, reranking and context allocation, citation, prominence, factual absorption, fidelity, and user behavior (arXiv 2607.14035).

Two findings in that survey deserve to be on the wall of every content team. First, competition can erode individual gains (arXiv 2607.14035). If a tactic works because it makes you comparatively better, it stops working when the field adopts it. Second, the widely quoted gains from the foundational GEO paper are valid inside their experimental setting but conditional on a source already being present in a fixed context, and establish neither organic discoverability nor durable traffic effects (arXiv 2607.14035).

That caveat is the whole ballgame. The best-known numbers in GEO describe what happens after retrieval has already put you in the room. They say nothing about getting into the room, and most vendor decks quote them as though they do.

#Generic rewrites are the wrong unit of work

If citation is comparative, then a uniform rewriting rule applied across a whole site is close to useless. Everyone gets the same lift, so nobody's relative position moves.

A March 2026 paper from Zhihua Tian, Yuhan Chen, Yao Tang, Jian Liu, and Ruoxi Jia makes that argument directly and builds the alternative (arXiv 2603.09296). Their criticism of existing methods is that they measure contribution rather than citation, and apply generic rewriting rules uniformly without diagnosing why an individual document failed to be cited (arXiv 2603.09296). Their system, AgentGEO, instead diagnoses the specific failure mode, selects a targeted repair, and iterates. They report over 40% relative improvement in citation rates while modifying about 5% of content, against 25% for the uniform baselines (arXiv 2603.09296).

Five times less editing for a better result. The lever is diagnosis, not volume.

#A bias with a practical read

Smirnov's analysis surfaced a source-type preference: the tested systems favored earned media, meaning information aggregators, over brand-owned or social media content (arXiv 2605.00012). Model behavior also varied. The trained policy shifted citations for gpt-4.1-nano, gpt-4.1-mini, and gpt-5-nano, but not for gpt-5-mini, where the analysis found a strong bias toward the URL itself, which the author gives as a likely reason snippet rewriting had limited impact there (arXiv 2605.00012).

The operational reading: in candidate sets where an aggregator or a high-authority domain is present, your own product page may be structurally disadvantaged regardless of how well written it is. Getting represented accurately on the aggregator is sometimes the higher-yield move, and it is a different workstream from editing your site.

#The safety half

The paper also tested a reference attack, poisoning the context snippets, and it succeeded: poisoned search results were cited by the LLM Overview system, producing inaccurate results (arXiv 2605.00012).

Two things follow. Any technique that works by exploiting a selection bias is sitting on the same surface engines are actively hardening, so treat it as a decaying asset rather than a strategy. If third-party pages about your brand can be manipulated into an answer engine's context, monitoring what gets said about you off-site is a defensive requirement, not a vanity exercise.

#What to change

Score against the candidate set, not a checklist. For each priority query, capture who else the engine retrieved, then ask what would have to be true for you to be the comparatively better source in that specific set.

Segment queries by contestability. Some have a locked winner and some have you structurally out of the running. Spend your editing budget on the contested middle.

Diagnose per page. Before you rewrite anything, name the specific stage that failed for that page and that query. Not retrieved is a different problem from retrieved and passed over, and the fixes have nothing in common.

Expect gains to decay. Anything that helps because it is comparative will erode as competitors catch up. Budget for maintenance, not for a one-time win.