Formatting won't win you AI citations

Formatting is not a citation lever: when two comparable pages compete inside an AI answer engine, restructuring one of them into tidier headings and bullets does almost nothing to its odds of being cited, while topical fit, concrete specifics, a recent date, and retrieval position decide the outcome.

That finding is uncomfortable, because reformatting is the cheapest thing a GEO checklist can ask for, and it is the item most teams do first.

#The experiment that isolates the variable

The result comes from What Gets Cited: Competitive GEO in AI Answer Engines, presented at SIGIR '26 in Melbourne. The design is the useful part. Rather than publishing pages and waiting to see what happens, the authors built a two-document retrieval-augmented generation testbed: exactly two candidate sources are injected into the model's context, and the trial records which one the first citation marker points to.

Every pair differs in exactly one factor. Brands are anonymized so name recognition cannot leak in, and source order is counterbalanced so position bias can be measured rather than absorbed. Across six models the authors ran 252,000 trials over 18 content factors, covering topical match, completeness, trust cues, readability, competitive standing, and freshness.

This is a clean answer to a question observational GEO studies cannot answer. Correlational work tells you cited pages tend to have property X. A paired comparison tells you whether adding X to a page changes the decision.

#Four factors act as gatekeepers

Four factors were unanimous across all six models, with effects large enough that failing on any one of them effectively removes a page from contention regardless of its other strengths:

Gatekeeper	What it means in practice
Topic match	The page answers the actual question asked, not an adjacent one
Explicit price	The concrete figure a commercial query needs is stated on the page
Recent timestamp	The page carries a current date rather than an old one or none
List position	The page arrives high in the retrieved set, not buried in it

Reported in What Gets Cited, where each of these four crossed an odds ratio above 10,000.

Secondary factors mattered but did not gatekeep: keyword coverage, missing specifications, absent comparisons, confident rather than hedged language, claims backed by evidence, internal consistency, and depth of coverage. These shift the odds. They do not decide the round on their own.

#Where formatting landed

Content structure, the factor that captures how neatly a page is organized, produced an odds ratio between 0.79 and 1.25 depending on the model. An odds ratio of 1.0 is no effect. The range straddles it. Scattering the same information across the page did somewhat more damage, but stayed far below the gatekeepers.

The authors' reading is that models parse content regardless of its visual organization. A language model does not skim. It has the whole passage in context, and it is choosing between two candidates on the merits of what they say.

#This does not contradict "answer first"

There is an obvious objection. Kevin Indig's analysis of 1.2 million AI answers and 18,012 verified citations found a strong positional gradient inside the page: 44.2% of citations came from the first 30% of content, 31.1% from the middle band, and 24.7% from the final third. If where a fact sits matters that much, how can structure not matter?

The two results describe different things, and holding them apart is the whole practical lesson.

Indig measured which sentence gets quoted from a page that was already retrieved and already going to be cited. Front-loading changes which of your claims travels. The SIGIR study measured whether your page beats another page at all. Moving your existing paragraphs around does not change what your page knows.

Put plainly: front-loading decides what gets quoted, substance decides whether you get quoted. Both are worth doing. Only one of them is a formatting change.

The same analysis points at the deeper variable. Heavily cited passages averaged 20.6% proper nouns against the 5% to 8% typical of English prose. That is not a layout property. Entity-dense writing is writing that names specific products, versions, companies, and figures, which is another way of describing a page that carries information a competitor's page does not.

#The "40% lift" is narrower than it sounds

The number most often quoted in GEO decks comes from the original GEO: Generative Engine Optimization paper at KDD 2024, which reported visibility gains in the 22 to 41 percent range from tactics like adding statistics, quotations, and cited sources.

A 2026 critical survey of the GEO literature makes the scope explicit: that gain applies to a source already handed to the generator inside a fixed multi-document context, not to organic discoverability. The survey also finds that general heuristics transfer poorly across domains, and that keyword stuffing shows null or negative results across benchmarks.

Read together with the SIGIR result, the picture is consistent rather than contradictory. The tactics that worked in the original paper were the substance-bearing ones: adding real statistics, real quotations, real sources. Those change what the page contains. The cosmetic subset does not.

#What to do differently

Four changes to the usual GEO checklist follow from this.

Stop treating a restructure as an intervention. If a page already contains the answer, reorganizing it is maintenance, not optimization. Log it as such and do not expect a citation-share move.

Put the number on the page. Price is a gatekeeper in commercial contexts, and its generalization is that a page missing the specific fact the query needs loses to one that has it. Pricing behind a "contact us" is a citation decision, not only a sales one.

Treat dates as load-bearing. Recency crossed the gatekeeper threshold in the paired trials, which is why our own audit engine ships dedicated checks for a stale dateModified and for missing authorship and date signals, rather than folding them into a generic metadata score.

Work on retrieval separately. Position was a gatekeeper, and position is set before the model ever reads your prose. Crawlability, indexation, and conventional relevance still decide whether you are in the candidate set at all. Nothing on the page helps if the page is not retrieved.

One caveat worth keeping. The testbed injects two documents directly, so it isolates the selection stage cleanly and says nothing about the retrieval stage that precedes it. Structure may well matter upstream, where pages are chunked and embedded. What the evidence does rule out is the idea that tidying a page is what earns the citation.

The uncomfortable version of this finding is that GEO has fewer cheap wins than the checklists imply. The useful version is that the expensive work is the work that pays: know something specific, say it plainly, date it, and make sure the page can be found.