Structured data came third in a 1,100 URL citation audit

What the GEO-16 audit counted, and on which pages

The design is a cross-sectional audit rather than an experiment, and that difference decides how far the results travel. The authors, Arlen Kumar and Leanid Palkhouski, wrote 70 intent-focused prompts spread across 16 business software verticals, put each prompt to Brave Summary, Google AI Overviews and Perplexity, and kept every URL those engines cited in their answers. That produced 1,702 citations: 612 from Brave at 36.0 percent, 598 from Google AI Overviews at 35.1 percent, and 492 from Perplexity at 28.9 percent. Normalisation lowercased hosts, stripped tracking parameters and fragments, resolved redirects and applied canonicals, leaving 1,100 unique URLs. Each of those was fetched and fully rendered before it was scored.

The scoring is where the framework does its work. Every page receives a band of 0, 1, 2 or 3 on each of sixteen pillars. The overall score the paper calls G is the sum of those sixteen bands divided by 48, which places it between 0 and 1. A second figure, H, counts how many pillars reached a band of 2 or better, and the paper calls those pillar hits.

Two consequences of that arithmetic matter before any result does. The first is that G is unweighted. Structured data counts for one sixteenth of it, and so does every other pillar, including the one the paper names only in a figure caption. The second is that the corpus was harvested from citations. The paper's notation defines the audited set as the URLs the three engines cited, and its outcome variable as whether a particular engine cited a particular member of that set. Read that way, a citation rate in this study is a share of pages at least one of the three engines had already chosen, not a share of the web and not a share of any one site's pages.

Seventy prompts is also a sample, and the number of prompts an answer measurement needs before it stops moving is a question with its own arithmetic. None of this is concealed in the paper, and a citation-first corpus is close to the only design available to anyone without an engine's index. It is worth naming plainly all the same, because the distance between "cited pages score well" and "scoring well gets a page cited" is the whole distance between an association and an instruction, and generative engine optimization advice crosses it constantly. A survey of the field reached a related conclusion from a different direction when it put crawling at stage two of seven and found few studies looking there.

Does structured data predict an AI citation?

On this evidence it is associated with one, and it is not the strongest association the paper found. Table 3 reports correlations between pillar scores and citation likelihood for six pillars. Metadata Freshness leads at r = 0.68 with a 95 percent interval of [0.64, 0.72]. Semantic HTML follows at 0.65, then Structured Data at 0.63, Evidence Citations at 0.61, Authority Trust at 0.59 and Internal Linking at 0.57. Every row is reported at p below 0.001. A second column in the same table, headed citation impact, reads +47, +42, +39, +37, +35 and +33 percent down the same ordering.

Three things about that table are worth stating rather than skipping past. It reports six pillars out of sixteen, and the paper does not print the other ten. The correlations descend by 0.03 and then by exactly 0.02 at every remaining step, which is a regularity a reader should notice in a table of independently estimated coefficients. And the design is observational, which the authors say themselves under threats to validity, noting that unobserved confounding may affect internal validity and that factors such as brand reputation or backlink profile could influence both the scores and the citations.

What the paper means by its Structured Data pillar is specific and checkable. Section 3 asks for a single h1 and a logical h2 and h3 hierarchy, valid JSON-LD using types such as TechArticle, with datePublished, dateModified, author and breadcrumb where relevant, canonical URLs exposed, and social cards present. It then adds one clause that carries more weight than the rest: ensure schema matches visible content.

That clause is the one place where this site holds a measurement of its own, and it is not encouraging. Parsing the four JSON-LD blocks across five real pages captured on 15 July 2026, we found on 26 August 2026 that 70 of 146 human readable markup values were not present in the visible text of the page they sat on. We also found, on 28 August 2026, that a valid BreadcrumbList scored zero because of the syntax it was written in, since our scorer reads JSON-LD and that page published RDFa. Those are Lantad measurements of five pages, with their dates attached, and they say nothing about citation outcomes. They do say that the pillar the paper ranks third is one most pages fail on a definition stricter than "the markup validates". A separate controlled experiment posted in May 2026 is a useful counterweight here, because it found that four content factors decided the first citation and formatting was not among them.

The three engines did not agree on what a good page looks like

The single most useful number in the paper is not in the correlation table. It is in Table 2, and it is the gap between the engines. Asked the same 70 prompts and scored by the same framework, the pages Brave Summary cited averaged a GEO score of 0.727 with 11.6 pillar hits. Google AI Overviews averaged 0.687 with 11.0 hits. Perplexity averaged 0.300 with 4.8 hits. The confidence intervals for Brave and Google overlap each other and neither comes near Perplexity's [0.267, 0.333].

Perplexity, on this framework, cited pages that scored less than half what Brave's cited pages scored, and cleared fewer than five of sixteen pillars on average. Whatever it is selecting on, the sixteen signals in GEO-16 are not a good description of it. That is a finding about the framework as much as about the engine, and both readings are useful: either the pillars miss what Perplexity weights, or Perplexity weights page quality far less than the other two do.

One methodological detail belongs with those numbers rather than in a footnote. Section 4.3 records that Perplexity results were collected through sonar-pro, its API model, rather than through the consumer interface a buyer is thinking of. An API model and a product surface are not obliged to retrieve the same way, which is a caveat the paper does not draw out and a reader should. Perplexity separately documents two named crawlers with different robots.txt behaviour, so the plural is real on the fetching side too.

The practical consequence is the one this blog keeps arriving at from different directions. "Optimise for AI citation" describes at least three targets that do not agree, which is why the eight engines in our own registry split on whether they can search the web at all and why a page that Google AI Overviews will quote is a different artefact from one Perplexity will. If you only have the budget to aim at one, aim at the one your buyers use, and read what Google publishes about its own generative surfaces rather than a cross-engine average. An average across engines that disagree by 0.427 of a score is a number with no addressee, and counting citations without knowing which engine produced them is a habit this blog has argued against before.

Two of the sixteen pillars are never named

A framework is only as reusable as its definition, and this one is not fully written down in the paper that introduces it. Table 1 maps six core principles onto their key pillars and names thirteen of them: UX Readability, Claims Accuracy, Microcontent, Semantic HTML, Structured Data, Metadata Freshness, Authority Trust, Evidence Citations, Transparency Ethics, Content Depth, Internal Linking, External Linking and Engagement Interaction. A fourteenth, Visuals Media, appears once, in the caption of Figure 5. Reading the v1 text on 4 September 2026, those fourteen are every pillar name it contains. The framework is called GEO-16.

The gap goes deeper than the names. Section 4.1 defines a pillar score as a weighted sum of sub-signals whose weights sum to 1, and then maps that score to a band through fixed thresholds. It gives no values for the sub-signal weights and none for the thresholds. So a reader holding the paper cannot score a page: not because the method is vague, but because the two numbers that turn a signal into a band are not printed. The data availability note says an anonymised artefact of per-URL pillar bands will be released upon acceptance to a peer-reviewed venue, and that until then data are available from the corresponding author on reasonable request.

None of that makes the correlations wrong. It makes them unverifiable by a reader, which is a different and more ordinary problem, and it is the reason the figures in this post are attributed to the paper in every sentence that carries one rather than absorbed into our own voice. It is also why this site publishes the whole of its scoring method including the parts that make the product look worse, and why it withholds a grade rather than emit a confident wrong one when a scan cannot see enough to grade. A score you cannot reproduce is a brand, not a measurement. Our own aggregate findings live at the research page with their sample sizes attached for the same reason.

What Lantad weights, and why that is a decision rather than a finding

The honest comparison to draw here is not between our score and theirs. It is between two sets of chosen numbers, neither of which is a measurement of what an answer engine does. GEO-16 divides by 48, which gives all sixteen pillars equal influence. That is a decision. Lantad's AI Visibility Score weights prose parity at 50 points of 100, access at 25, structure at 15 and schema at 10, read from SCORE_WEIGHTS in core/src/config.ts. That is also a decision, made by this team, and this blog has said before that a configured weight is not a finding anybody measured.

So the paper ranks structured data third of six and we weight schema last of four, and neither of those facts is evidence about the other. What the two do share is a preference for signals a machine can check without asking a vendor, and that is where a reader can actually act. The two pillars ranked above structured data in the paper's ordering are both cheap to inspect. Metadata Freshness asks for a visible date and a machine readable one that agree; on the five pages we captured on 15 July 2026 we found three of five carried JSON-LD at all, which is the prior question. Semantic HTML asks for one h1 and a hierarchy under it, which is inspectable in a browser view source in under a minute.

The order to work in follows from what a missing signal costs. A page an AI crawler cannot fetch scores nothing on any framework, which is why access sits second in our weighting and first in practice. A page whose text arrives only after JavaScript runs is invisible to a client that does not execute it, whatever its markup says, and what an AI crawler receives for one URL is worth looking at before tuning any pillar. Structured data is the layer above both of those, and a third place finish in an observational audit of pages that were already cited is a reason to get it right, not a reason to expect a citation for it.