What a 252K-Trial Study Says About AI Citations — and the Variable It Deletes

Questions this guide answers

• What makes AI answer engines cite one source over another?
• What does the Sprinklr "What Gets Cited" GEO study find?
• What are the gatekeeper factors for AI citations?
• Is optimizing content enough to get cited by AI?
• Why does my optimized page still not get cited by AI?

Direct answer

There is now a serious, controlled experiment on what makes an AI answer engine cite one source instead of another. A team at Sprinklr ran roughly 252,000 trials across six models and published it as "What Gets Cited: Competitive GEO in AI Answer Engines" (SIGIR '26). It is the cleanest read we have on the content-level drivers of citation, and most advice quoting it will stop at the headline list of factors.

The most useful thing about the study is also the thing almost nobody quoting it will mention: to measure content cleanly, the authors deleted the variable that decides citations in the real world. Read it that way and it is genuinely valuable. Read it as "here is what gets cited" and it will quietly point you at the wrong work. The accurate takeaway is a both/and: content factors are necessary, but they are not sufficient, because the study had to hold the dominant production driver — brand and domain trust — constant in order to see them.

What the study actually did

The design is a paired, two-document test. For each trial the authors inject two candidate sources into a model's context, ask a question, and record which source the first citation marker points to. They built it from 100 product-review blogs across 50 consumer categories, then varied one content factor at a time — 18 of them — while holding everything else constant. Every brand, product model, and publisher was replaced with a fictional alias so the model could not lean on familiarity. Then they ran it about 252,000 times across Gemini 2.5 Flash, GPT-5 Nano, GPT-5 Mini, GPT-5.2, Claude 3.5 Sonnet, and Kimi K2 Thinking, swapping the document order to control for position bias.

That is a real experiment, not a correlation scrape. Varying one factor at a time and counterbalancing order is what lets you attribute an effect to the factor rather than to a confound. It is also why the result is narrow in a specific, important way, which we will get to.

The finding: a hierarchy, not a checklist

The clean result is that the 18 factors are not a flat list. They sort into tiers, and the tiers matter more than any single factor.

Four factors behaved like gatekeepers — unanimous across all six models, with effects large enough that failing any one of them effectively zeroed out a source's citation odds no matter how strong the rest of the page was. Below those sat a second tier of differentiators that broke ties once a source had cleared the gates. Formatting-only changes and promotional tone showed weak or inconsistent effects and can be deprioritized.

The variable the study deletes

Here is the part that changes how you should use it.

To isolate content, the authors anonymized every brand and domain into fictional aliases. That is the correct experimental move — it strips out familiarity bias so the measured effect is the content factor and not "the model has heard of this brand." But brand and domain trust is not noise to be controlled away. In production retrieval it is the dominant driver. The engine does not choose between two equally-trusted strangers; it retrieves and cites the domains it already trusts for the category, and the authors say exactly this in their own limitations section.

So the precise claim the study supports is: given two sources the model trusts equally, here is which content wins. It is not: here is what gets cited in the wild. Those are different questions, and the gap between them is most of the actual job. A page can satisfy every content gatekeeper in the paper and still never be retrieved, because it lives on a domain the engine does not pull for that query. The experiment holds that constant precisely so it can see everything else — which means the thing it holds constant is invisible in the results, not absent from reality.

What our own category data says about that deleted variable

We measure our own market the way we measure clients', so we can put a number on the layer the study brackets out. Across a recent seven-day window we logged 13,510 citations spanning 1,531 distinct domains, and no single domain held more than about 6.4% of them. Reddit led at roughly 6.4% (859 of 13,510), Wikipedia followed near 4%, and Semrush, Profound, TechRadar, and HubSpot clustered around 2% each. (SolCrys measurement, workspace solcysai-aeo, 7-day window ending 2026-06-18, five engines.)

A 1,531-domain consensus is the production reality the testbed compresses to two anonymous documents. Which domains get retrieved, and how consistently a claim is corroborated across independent source types, is what decides who shows up — and that is the exact axis the anonymization removes. It is also why only about 12% of AI-cited URLs rank in Google's top ten for the same query, and why ChatGPT and Perplexity overlap on only about 11% of cited sources: each engine reads a different trusted slice of the web. The content factors operate inside whichever slice gets retrieved. They do not get you into the slice.

We made the longer version of this argument in AI cites consensus, not authority.

Don't transplant B2C product reviews onto B2B

One more reason to read the paper as a hierarchy and not a checklist: the corpus is consumer product reviews. "Price not mentioned equals out" is partly an artifact of that domain — when you are buying a fitness tracker, a page with no price genuinely fails to answer the question.

For a B2B SaaS buyer comparing platforms, the equivalent gatekeeper is not a literal dollar figure; it is whether the page answers the specific decision the buyer is making — the integration, the team size, the compliance requirement — with concrete specifics instead of brochure language. So translate the shape of the finding (topic match, freshness, concrete specifics, evidence over hedging, comparisons) and do not copy the literal factor list across a domain boundary it was never measured on.

The honest synthesis: necessary, not sufficient

Put the two halves together and the takeaway is a both/and, not a winner.

The content factors are real, and most of them are cheap to fix: answer the actual question, state concrete specifics, keep the timestamp current, replace hedged claims with evidence, add the comparison. Those are editorial changes you can ship this week, and the study is good evidence they matter once you are in the running.

The source layer is the thing the study had to delete to see them — and in production it decides whether you are in the running at all. The failure mode the paper will accidentally encourage is teams treating the factor checklist as the whole game, polishing pages the engine never retrieves. The accurate reading is that content optimization is necessary but not sufficient, and the marginal dollar usually belongs on getting your claim corroborated across the trusted sources the model already pulls.

What to do with it

The loop that respects both halves is the one we run on ourselves: Measure, Diagnose, Execute, Verify.

• Measure the real source map. For your actual buyer prompts, pull the cited sources per prompt across the engines you care about, and tag them owned, competitor, editorial, or community. That tells you which trusted domains the model reads for your category — the layer the study held constant.
• Diagnose where you are losing. If you are not even retrieved, that is a source-trust gap, and no amount of on-page polish fixes it. If you are retrieved but not cited first, now the study's content factors are exactly your diagnostic checklist.
• Execute on the right layer. Fix the cheap content gatekeepers on pages that already get pulled; spend the harder budget on corroboration — comparison pages, editorial roundups, community threads — where you are invisible in the source map.
• Verify by re-running the same frozen prompt set and checking whether the answer actually moved. AI answers are non-deterministic, so measure a rate across runs, not a single yes or no.

About the author

Jia Chang is Co-Founder & CTO of SolCrys, the AEO operating system for brands and agencies competing for visibility, citations, and recommendations across the major AI engines. AI architect with 15+ years building production AI systems, most recently as an engineering leader at Microsoft. Connect on LinkedIn.

Sources

• Vishwakarma, Kumar, Jamidar (Sprinklr). "What Gets Cited: Competitive GEO in AI Answer Engines." arXiv:2605.25517, SIGIR '26 — ~252,000 trials across six LLMs, 18 content factors in a paired two-document RAG testbed.
• SolCrys AI visibility measurement, workspace solcysai-aeo, 7-day window ending 2026-06-18 (13,510 citations across 1,531 domains; Reddit ~6.4% top; solcrys.com rank 7 at ~2.0%).
• Profound: AI platform citation patterns — only ~12% of AI-cited URLs appear in Google's top ten for the same query.
• Authority Tech: ChatGPT vs Perplexity, only ~11% of cited sources overlap per engine.

FAQ

Related guides