What Really Matters for Getting Cited by AI Search Engines

What changed in how AI engines pick their citations?

Two studies landed within two weeks of each other in May 2026 and together they settle an argument that has run since generative search became a traffic channel: what actually makes a large language model cite one page over another.

The first is a SIGIR '26 paper, What Gets Cited: Competitive GEO in AI Answer Engines, built on 252,000 trials across six language models. Researchers fed each model exactly two candidate sources per query and measured which one the model cited first, testing 18 content factors along the way. Four of those factors turned out to be decisive gatekeepers. Fail one and the odds of citation collapse regardless of how good the rest of the page is.

The second is an Ahrefs data study published May 11, 2026, tracking 1,885 pages that added JSON-LD schema between August 2025 and March 2026 against 4,000 matched control pages. It found schema markup produced no meaningful citation lift on ChatGPT or Google AI Mode, and a measurable decline on Google AI Overviews. That result corrected months of SEO chatter claiming schema was either the whole game or worthless. Neither claim was true, and neither study actually tested the same thing as the other.

Who does this actually affect?

B2B SaaS and tech growth teams feel this first. Their real fear is not a ranking drop, it is not knowing whether ChatGPT recommends them or a named competitor when someone asks for the 'best [category] tool'. The gatekeeper factors explain why that answer can flip overnight: a competitor updates a comparison page with a current date and a visible price, and the model starts citing them instead.

Local, multi-location and service businesses face a quieter version of the same problem. Nobody sees a lost citation the way they see a lost rank. A business simply stops showing up when someone asks an AI engine for the best plumber, dentist or agency in a given city, and there is no notification for that. List position and recency matter even more here, because local answer sets are small and get refreshed often.

Both groups share one exposure: pages built once and left alone. A page with no update date, no price and thin topical coverage fails three of the four gatekeeper factors before an LLM even gets to judging the writing.

What are the four gatekeeper factors, exactly?

The SIGIR researchers call them gatekeepers because each one behaves less like a ranking signal and more like a pass or fail switch. Across all six models tested, the effect sizes were large enough that the authors reported odds ratios in the thousands, not the low single digits typical of most SEO factors.

1. Topic match: the page has to directly answer the query. Off-topic content lost almost every matchup regardless of authority or length.
2. Price stated: a visible, specific price beat a gated or missing one, with odds ratios ranging from 6.26 up past 10,000 depending on the model.
3. Recent timestamp: a current, visible published or reviewed date beat old or absent dates, with odds ratios from 14.4 into the thousands.
4. List position: ranking first among competitors carried odds ratios from 1,795 up past 10,000, confirming that where a page sits in retrieval order still shapes what gets quoted back.

Does schema markup still matter?

Not the way the last two years of advice suggested. The Ahrefs test found adding JSON-LD schema moved Google AI Mode citations by 2.4 percent and ChatGPT citations by 2.2 percent, both statistically indistinguishable from random noise, and Google AI Overviews citations down 4.6 percent, a small but real decline against the control group.

It is worth being precise about what the SIGIR paper did and did not test, because a version of 'schema does nothing for AI citations' has been circulating as if it came from that study. It did not. The SIGIR trials measured topic match, price, recency and list position. Schema markup was never one of the 18 factors in that testbed. The claim that schema is irrelevant to AI citation rests on the Ahrefs test alone, and that test says schema is not a lever worth pulling on its own, not that structured data is worthless for every purpose.

The practical read: schema is hygiene, not strategy. Keep it clean because it helps other systems parse your pages, but stop treating a script tag as a substitute for a current date, a visible price and content that actually answers the question being asked.

How should you respond, starting now?

None of the four gatekeeper factors require a trick. They require doing ordinary things consistently, at a scale most in-house teams cannot sustain.

1. Put a real, visible date on every page that could plausibly answer a buying question, and actually refresh it when the underlying facts change instead of only editing the timestamp.
2. State pricing wherever you can. Gated pricing pages are functionally invisible to a model comparing you against a competitor who lists a number.
3. Cover the full question universe for your category, not just your flagship page, since topic match is evaluated per query, not per domain.
4. Track Citation Share, the percentage of relevant AI answers in your category that cite you, alongside rank position. A page can rank first in search and still lose the citation to a fresher, more specific competitor page.

Where does this leave citation engineering?

This is the part the two studies confirm without either one setting out to. There is no page two in an AI answer, and the factors that decide who gets page one are coverage, freshness and specificity applied consistently across an entire question set, not a one-time optimization pass.

That is the mechanism behind what Omnicite calls Citation Engineering: engineering authoritative content at the scale and quality AI engines trust, then tracking citation share across ChatGPT, Perplexity, Gemini and Google AI Overviews. It is not a claim to game or manipulate any model. The four gatekeeper factors are exactly the kind of signals that quality, breadth and timeliness produce naturally when content is published at real volume, something Omnicite does at 10 to 12 articles per day per client.

LeadHaste, the group behind Omnicite, ran this playbook on its own site and went from zero to 1 million impressions and over 200 AI citations in four months, with Domain Rating moving from 1 to 24 over the same period. That is one dataset, from one company, over one stretch of time, and it lines up with what the SIGIR gatekeeper factors predict: consistent, current, on-topic coverage wins the citation, not a single clever tag.

Sources

Source: SIGIR '26 (arXiv preprint)

252,000 trials across six LLMs identified topic match, price, recency and list position as the four gatekeeper factors for AI citation, each with large effect sizes. SIGIR '26 (arXiv preprint), 2026-05-25

Source: Ahrefs

Adding JSON-LD schema to 1,885 pages produced statistically insignificant citation changes on ChatGPT and Google AI Mode, and a 4.6 percent decline on Google AI Overviews, against 4,000 matched control pages. Ahrefs, 2026-05-11