GEO-bench, from the 2023 paper that coined the term [12], pairs each user query with source documents [1]. That construction is the whole reading of the headline number. The document sits in the input set before the generator writes anything, so what gets scored is selection among candidates. The dev.to author says as much: it is an improvement in how likely an already-plausible page is to be selected and quoted, a shift among candidates already in contention rather than a promotion from invisible to authoritative [2].
The gate upstream of that is binary, and Google's own documentation says where it sits. AI Overviews and AI Mode draw on the same index and the same quality systems as classic Search, and there is no separate AI submission process [3]. A page has to be indexed and eligible to be shown with a snippet before formatting work can reach it at all [4]. The cases the author reports investigating were mostly that shape: a page that renders nothing without JavaScript, or a bot rule someone switched on in a CDN eighteen months ago and nobody remembers [5]. FAQ markup leaves both problems untouched.
The budget arithmetic follows from Pew's panel. Clicks on a link inside an AI summary run at 1% of visits [6]. Grant the benchmark's 40% in full and apply it there: 1% becomes 1.4%, a gain of four tenths of a percentage point [2]. The traditional-result click rate in the same panel falls by seven percentage points when a summary is present [1], which is about seventeen times the best case gain from the citation lift [3].
For those numbers to belong in one sentence, GEO-bench's visibility metric would have to map onto citation clicks in Google's summaries, on a document mix resembling yours, with the same query types. The paper's second finding, that which technique works varies significantly by domain [7], is the paper itself confirming that the transfer is uneven across domains.
The piece is headlined as a licensing decision with a formatting problem attached [8], and the argument underneath is about admission rather than contract terms [3][4][5]. Those are the same layer viewed from two ends. A robots rule and a content deal both answer whether you are a candidate; one you configure, the other you negotiate. Formatting only orders the set you were already admitted to.
So the sequence I would fund, in my context, starts with an eligibility check, then structured data, because FAQ, HowTo, Organization and Article markup give a retrieval system a fact to extract rather than infer [9]. Then narrow pages with the answer near the top of each section, because these systems lift a passage, and a page that spends three paragraphs warming up leaves the model choosing between the preamble and nobody [10]. Then a dull afternoon making the description of the business identical in every directory, because answer engines cross-reference sources rather than trusting one, and three different descriptions get attributed less confidently than three matching ones [11].
