The 252,000-Trial GEO Factorial: What Gets Cited, and What It Can't Tell You

"What Gets Cited: Competitive GEO in AI Answer Engines" (arXiv 2605.25517) asks a simple question. When a model has several pages in front of it, which content features make it cite one first?

The paper is a Sprinklr-authored submission to the SIGIR industry track, and it ships a GEO checklist. That matters for how you read it. Sprinklr has a commercial interest in GEO techniques looking useful. So does every vendor in this space, including the publisher of this page. An interested author does not make the data wrong, but it raises the bar for replication.

The study sits in a small group of GEO experiments with measured effect sizes. Others include the original GEO paper and C-SEO Bench. All of them share one design trait, covered below.

The authors ran a factorial design: they varied page features in combination and measured the effect of each across 252,000 trials.

The setup

• Corpus: a synthetic set of 100 ecommerce blogs, built for the experiment.
• Candidates: two documents per trial. The model chose between a pair.
• Retrieval: none. No search engine was called. The documents were placed directly in the model's context.
• Brands: anonymised by design, so brand recognition could not drive the result. The authors warned that real systems may favour established domains.
• Analysis: mixed-effects models, which separate the effect of each feature from the noise of individual pages and trials.

What that design measures

It helps to split AI answers into four stages: parametric knowledge, crawler access, retrieval and generation. This study only tests the last one. Every page was already retrieved before the model saw it. That is a clean way to isolate generation-stage preferences. It is not a test of whether any page would be found in the first place.

Three findings stand out. The quotes are the paper's own wording as recorded in our evidence base.

1. Relevance and position dominate

"Mixed-effects models show that topical relevance and list position are the biggest drivers of being cited first." Being on topic for the question and appearing first in the candidate list mattered more than any edit to the page.

2. Formatting barely moves the needle

The paper found "formatting-only edits have little impact." Rearranging headings, lists and layout without changing what the page says did not reliably change citation.

3. Prices and dates helped

"Including explicit price information and a recent timestamp also helps consistently." This is one of the few positive on-page signals in any controlled GEO experiment.

How it lines up with other work

The relevance finding agrees with the critical survey of 45 GEO studies (arXiv 2607.14035), which names topical relevance and context position as "the most reproducible levers." C-SEO Bench also found position in the context mattered more than rewriting tactics. Keep the caveat in view: the survey is single-author and non-peer-reviewed, and two interested sources agreeing is corroboration, not replication.

The limits are as important as the findings.

• Two candidates likely overstate position. With only two pages per trial, being first is a coin flip with large consequences. Production answers draw from larger pools, so the position effect is probably smaller in the wild.
• Position is not something you control. The study placed pages at the top of the context directly. In a real engine, the retrieval operator decides order. No publisher can move a page to slot one.
• No search engine, no live web. The results say nothing about whether a page gets crawled, indexed or retrieved. Those are the stages where most brands actually fall out.
• Anonymised brands remove the biggest real-world factor. Brand prominence is strongly associated with citation in other research. This design deliberately excluded it.
• A synthetic ecommerce corpus. One category, built for the test. Results may not transfer to B2B software, local services or news.
• The timestamp finding is not a freshness lever. A timestamp on a page is self-declared. The study shows a model reading a date in context, not an engine rewarding updated pages. Freshness as a ranking lever remains unresolved.

In short: this is good evidence about how a model chooses between pages it already has. It is no evidence about getting into that set.

The study supports two small, honest moves and rules out a larger set of busywork.

Do this

1. State specific facts plainly. Put your price, your plan limits and your key numbers in a sentence on the page. Our claims policy treats this as the only positive experimental signal for on-page content: it gives the engine a sentence it can quote.
2. Make sure those facts are in the raw HTML. Vercel's crawler study found ChatGPT and Claude crawlers fetch JavaScript but don't execute it. A price that only appears after scripts run may not exist for them.
3. Stay on topic. Each page should answer the question it is about. This is topical fit, not keyword coverage. Google says you don't need to worry about capturing "every variation" of a query. Our explainer on whether staying tightly on topic helps goes further.
4. Keep dates accurate. Show a real publication or update date. Never change a date without changing the content.

Skip this

• Reformatting pages for AI. See whether headings and bullet lists help AI quote you for the wider evidence.
• Rewriting copy in a supposed AI-friendly style. Google's AI optimization guide says you don't need to write in a specific way for generative AI search.

Check the gate first

None of this matters if engines never retrieve your page. Before editing content, confirm crawlers can reach you and that engines name you at all. The free checker on this site shows whether AI platforms name you for up to three buyer questions, as a one-off snapshot.