Which of two pages do AI answers cite? The 4 gatekeepers

Which of two candidate pages an AI answer engine cites is effectively decided by four gatekeepers: topic match, a stated price, a recent timestamp, and a front slot in the context. If even one of them fails, the citation probability approaches zero regardless of the content's other strengths. That is the conclusion of a controlled experiment in which Sprinklr researchers attached 18 content factors, one at a time, to 6 commercial LLMs and asked them 252,000 times ^[1]. They put only two candidates into the answer context, changed a single factor between them, and isolated, as a per-factor odds ratio, "which of two similar pages wins the citation when they go head to head."

This is a different question from earlier GEO research. The original GEO paper measured how much a single page is reflected in citations within a fixed set of search results ^[2], but in real AI answers the visibility bottleneck is less about rank and more about winning a citation slot against competitors. This paper reproduces that competition directly.

Why measure "competitive citation" against competitors separately

Traditional SEO optimizes for rank, but RAG-based AI answer engines cite only a subset of the retrieved candidates. That shifts the visibility bottleneck from "how well are you retrieved" to "do you beat competitors for the citation slot."

Prior research had two gaps. First, it measured single-source visibility scores but did not estimate relative preference when two candidates face off directly ^[2]. Second, production logs mix content effects with search rank, presentation order, and interface elements, so you cannot isolate which content factor caused a citation. This paper solves both at once with a controlled testbed that injects only two candidates, changes one factor at a time, anonymizes brands, and swaps the presentation order.

Two candidates, one factor changed

The core design is "isolate one factor." The researchers picked 100 real review blogs, two per category across 50 B2C product categories, and used GPT-4o to replace brand and publisher names with fictional aliases ^[1]. The goal was to remove the bias toward brands the models already know from pretraining. They then kept facts, prices, specs, and length the same, built two variants that differed in exactly one factor, placed both candidates side by side in the answer context, and checked which one was cited.

Figure 1. The pipeline: real product blogs are collected into a brand-anonymized synthetic corpus, matched pairs are built for each content factor, and the result is 4,320 scenario-query pairs ^[1].

The study measures the 18 factors shown in the table, grouped into six categories from content relevance through completeness, credibility, readability, competitive positioning, and recency ^[1].

Table 1. Six categories of the 18 content factors ^[1].

The run covered 2,400 trials per factor, 42,000 per model, and 252,000 across all six models ^[1]. Each factor had 80 scenarios, three query phrasings, presentation order flipped both ways, and five repetitions. These controls separate position bias and familiarity bias from content effects. In 86.4% of responses, exactly one URL was cited, and citations that matched neither variant were excluded from the regression ^[1].

Figure 2. Experiment scale: 4,320 scenario-queries across 6 LLMs with position randomization, 50.4K prompts per run repeated 5 times, 252,000 trials in total ^[1].

Effect sizes were estimated separately for each factor with a logistic generalized linear mixed model (GLMM).

$Y=1$ is the event that the variant favored by the factor (A) is cited, and $X_i$ is whether that factor's treatment is applied. $u_s$ and $v_{so}$ are random intercepts on the scenario and on the scenario-by-order combination, correcting for the fact that repetitions within the same scenario resemble each other. The final effect is read as the odds ratio $\text{OR}=\exp ({\hat{\beta }}_0)$: the larger the value, the more overwhelmingly the side with that factor wins the citation. The paper classifies an OR of 100 or more as "very strong" and 10–100 as "strong" ^[1].

Four gatekeepers: fail topic, price, date, or position and the citation disappears

The paper groups four of the 18 factors as gatekeepers that all six models agreed on: topic match, a stated price, a recent timestamp, and position 1 in the context ^[1]. Its text says all four exceeded an odds ratio of 100, but in the paper's own Table 2 only topic match and position 1 clear 100 in all six models. A stated price came in at 36.1 for Kimi-K2, 7.82 for GPT-5-Nano, 6.26 for GPT-5-Mini, and 30.4 for GPT-5.2, and a recent timestamp at 68.7 for Kimi-K2 and 14.4 for GPT-5-Nano. All are still far above 1, so the direction holds: the variant with the factor wins the citation. The paper calls these four gatekeepers, because if any one fails, the citation probability effectively vanishes regardless of other content strengths.

Table 2. Citation odds ratios (OR) of the four gatekeepers by model. "10,000+" marks cases where quasi-separation pushed the estimate to an extreme; read these as a decisive advantage rather than a precise multiplier ^[1].

Two points matter in practice. The first is price. The odds that a page with a stated price beats a page without one ranged from 6.26x to over 10,000x ^[1]. A page that hides its price behind "contact us" is not just missing information; it can fail a gatekeeper and drop out of citations. The second is position, which here means not the position within the page body but the slot order after the page has been retrieved into the context ^[3]. In other words, securing a front slot through search visibility and structural signals is a precondition for being cited.

Seven factors that decide the rest, and format that did not matter

Once past the gatekeepers, seven secondary differentiators decided the outcome. They were significant in four or more models, with odds ratios mostly between 2.1 and 243 ^[1].

Table 3. Odds ratio ranges by model for the seven secondary differentiators ^[1].

Format, by contrast, was nearly meaningless. Dense paragraphs vs. organized sections had odds ratios of 0.78–1.68, and scattered vs. consolidated information 1.13–3.87, with no consistent effect across the six models ^[1]. The paper reads this as evidence that "LLMs parse content regardless of visual organization." This result does not directly conflict with the GEO-SFE study, which optimized structure alone ^[4]. GEO-SFE dealt with the absolute citation probability of a single page, while this paper looked only at relative preference after two candidates are already in contention, so they operate at different levels. Format is best treated as a lower-priority lever to pull after the gatekeepers are passed.

Recency produced an interesting ordering too. A recent date was most favorable, undated came next, and an old date ranked last ^[1]. The implication: if you cannot update a date, removing it is better than leaving an old one in place.

Gemini, Claude, and GPT engines use different citation criteria

All six models agreed on the four gatekeepers, but their sensitivity below that diverged ^[1].

Table 4. Content-factor sensitivity by model ^[1].

Claude and Gemini look at only a few factors but respond almost decisively once a factor crosses a threshold, while Kimi and the GPT family respond gradually to more factors. Because GPT-5-Nano, Mini, and 5.2 behaved alike, the paper argues that architecture, more than model size, shapes citation tendencies ^[1]. If your target engine is Gemini or Claude, the four gatekeepers act like a binary switch, so focus there. If it is the GPT or Kimi family, building up the secondary differentiators as well pays off more.

What this experiment could not see

The paper names three limitations itself ^[1]. First, it fixed the candidates at exactly two. Real RAG retrieves 5–10 or more, but the authors reduced it to two to isolate a single factor, so these results describe pairwise preference on a controlled slate, not full multi-document competition. Second, it anonymized all brands and publishers, so any residual preference that production systems have for well-known brands or trusted domains is not reflected in this study at all. Third, GPT-4o was involved in seed curation and anonymization, so these results should not be stretched into "the models' overall judgment of writing quality."

The second limitation matters most for practical interpretation. The paper precisely shows "the conditions under which content wins a citation," but by design it cannot answer "which domains are trusted." Brand and domain authority effects need separate evidence, such as the GEO at Scale study that measured 102 brands.

How to apply this to AEO/GEO: check content against the gatekeepers first

The paper gives a checklist with clear priorities. Following the order is the key.

Figure 3. Diagnosis flow when a brand is not recommended first. If it is in the cited set, find and fix weak factors across four categories (credibility, relevance, completeness, context); if it is absent entirely, improve search visibility (SEO) first ^[1].

1. Pass the four gatekeepers first. Topic match, a stated price, a recent timestamp, and a front slot secured through search visibility. However well you do everything else, if one of these four fails, the citation disappears.
2. Do not hide prices. A missing price is not just an omission; it is a gatekeeper-level loss. Stating a price, or at least a price range, on product and service pages works in your favor.
3. Manage dates. If an old date cannot be updated, removing it is better; if it can, show a recent timestamp.
4. Then build completeness and confidence. Adding spec tables and comparisons with competing alternatives, and replacing hedges like "probably" or "it may be" with assertive, evidence-backed statements, raises the odds ratio substantially in secondary differentiation.
5. Push format fine-tuning to later. Splitting paragraphs and adding emphasis is finishing work to do after the gatekeepers are passed.

From TRAIL's perspective, this paper exposes a gap in diagnosis frameworks. A stated price and timestamp management are gatekeeper-level signals, yet most GEO checklists so far have not weighted them as heavily as structure or authority signals. It gives grounds to move "is the price visible on the page" and "is the date recent" up to top-priority audit items when diagnosing AI search optimization. Because it isolates citation factors through controlled experiments rather than correlation, reading it alongside GEO-16's correlation-based diagnosis and C-SEO Bench's effectiveness tests gives a much sharper picture of citation optimization.

You can continue with A reward ceiling for RL alignment, computed before training, How often do AI answers name people, and what rosters miss, and How AI product recommendations vary across tools and APIs.