The stage in the middle of AI search cannot read your entity work at all — no schema markup, no knowledge graph, no sameAs trail. Structured data still earns AI search visibility, but at the two stages on either side of that one. That finding comes from Peec AI's August 2026 guide to rerankers, the models that decide which passages survive into an AI answer, and it is one of the clearest technical explainers AI-search marketing has produced. If you build entities for a living, the blind spot is not bad news. It is a map of where each kind of work pays.
Why does my page rank but never get cited in AI answers?
On August 6, 2026, the AI-search analytics firm Peec published a deep guide to rerankers, demonstrated across twelve open-weight AI models. The pattern it explains is one every marketer has felt: your page is indexed, it ranks, it even gets retrieved, but the answer quotes someone else.
Their explanation starts with a picture worth keeping in mind. AI search is not one decision. It runs in stages. Retrieval casts a wide net and pulls a shortlist of candidate passages. A reranker re-reads each candidate against the question and scores it. The strongest survive. Then the model writes the answer and chooses what to cite. Three stages, three different judges.
Do AI rerankers read structured data?
No — and the guide shows it without quite saying so. The middle judge reads exactly two things: the question, and the passage text. That is the whole exam. No knowledge graph. No JSON-LD. No sameAs trail, no Wikidata item, no review stars. Your structured data is not demoted at the reranking stage — it is simply not in the room. The model reads text and only text — and it grades meaning, not keyword overlap: a passage that answers plainly in different words can outscore one that parrots the question.
If you have invested in entity engineering, that sentence might sting. It shouldn't. It should focus you, because the stages on either side of the blind spot are where your entity work was earning its keep all along.
Does entity engineering still matter for AI search?
Before the reranker, entity signals — structured data, corroborated identity, a name that resolves — decide whether you are in the candidate pool at all. Retrieval leans on authority, freshness, and recognition — and Peec's own guide notes that "explicit brand terms make retrieval easier." That ease is not luck. It is a resolvable entity doing retrieval's work in advance. After the reranker, the system decides what to cite, and source identity returns to the exam. "Accuracy and source identity still matter" — their words. A system choosing which source to put its name behind consults the record: who this is, whether independent sources corroborate it, whether the identity resolves cleanly.
So the division of labor is sharper than most advice admits. Entity engineering gets you into the room and gets you trusted on the way out. Clear, answer-shaped prose is the only tool that works in the middle. Better: a page whose entity resolves cleanly and whose sections open with a plain answer — it enters the pool on identity, survives the middle on clarity, and earns the citation on trust. Worse: flawless JSON-LD wrapped around sections that never answer the question. That page gets retrieved on brand strength, loses the middle exam, and watches a clearer paragraph take the citation. The mirror image fails too: crystal-clear prose attached to an entity no system can resolve wins an exam it rarely gets invited to sit.
This is entity engineering's central claim wearing a lab coat: a recognized entity earns relevance credit more efficiently than an unrecognized one publishing the same content. Entity work and content work are both necessary. Neither is sufficient alone. What the reranker literature adds is the mechanism — a named stage where one discipline goes blind and the other carries the weight.
Does structured data help AI search visibility?
Yes — at the two stages that can read it. Structured data earns AI search visibility where a system is deciding who you are. That happens twice: when retrieval builds the candidate pool, and when generation picks which source to credit. At the reranker it does nothing, because the reranker never sees it.
That split answers the question most marketers are really asking. Is schema still worth the effort now that answers come from a model? It is, for a narrower job than the vendors sell. Schema is a resolution tool, not a relevance tool. A JSON-LD block names your organization. It points to the same record everywhere through sameAs. It matches what the page visibly says. Together, those let a system file your page under the right entity before any passage is read. That is retrieval's problem, and schema helps solve it. It is generation's problem too, when the model decides whose name goes next to a claim. Schema helps there as well.
What schema cannot do is make a passage answer the question better. No type on schema.org does that — not FAQPage, not Article, not Product. The passage has to do it in words, because words are all the middle stage reads. Three habits follow. Keep the markup; it is cheap, and it works where it works. Stop expecting it to carry a section that never states its answer. And never let the markup claim more than the page shows, because Bing now says so in writing.
What do Bing's webmaster guidelines say about structured data and AI search?
Bing says structured data "may support clearer grounding but does not guarantee visibility or grounding traffic" — and it names no schema type for AI search at all. That sentence is section 14 of the Bing Webmaster Guidelines. Microsoft rewrote the guidelines in late February 2026 to cover "Bing search experiences, Copilot, and grounding API results." (The rewrite was reported on February 26, 2026. The page carries no revision date. Every quote here was checked against the live page on October 21, 2026.) The section is two sentences long. The second one is the rule: "Markup must accurately reflect visible content."
The guidelines spend their words elsewhere. Where they spend them reads like the three-stage map above. Section 15 wants "facts and definitions" that "are explicit." Key statements must "not rely on implied content." Important information has to be "visible on the URL itself." Section 16 asks for "clear and consistent naming for people, organizations, products, and locations." Its reason: "clear entity definition improves grounding visibility and citation accuracy." Section 18 says to "place essential information near the top of the URL" and to "avoid long introductions before addressing the main topic." Section 17 wants one topic per URL. Stated facts, named entities, early answers. That is the reranker's exam and the entity layer's job, written as a search engine's house rules.
Bing's abuse list turns the markup rule into a penalty. Under "Misleading Structured Data," markup that is "irrelevant, inaccurate, or misleading may be ignored." It "can affect trust and eligibility for enhanced features." The same page defines the new discipline without overselling it. "Generative Engine Optimization (GEO) focuses on content eligibility for grounding and reference in AI responses." And: "GEO does not guarantee grounding or citations in AI experiences."
The schema types Bing does name live on an older help page about rich search results. The list: businesses and organizations, people, products and offers, events, reviews, recipes, breadcrumbs, and fact checks, in JSON-LD, Microdata, RDFa, or Microformats. That page says Bing uses annotations "to supplement and validate our data sources." Read the two pages together. Bing's answer to "which schema types boost AI search visibility" matches the reranker guide's: none in particular. Accurate markup tells the system who you are. Clear prose gets you cited.
One caveat. Microsoft's guidelines describe Microsoft's systems. Its grounding API does serve other assistants — "nearly every major AI assistant in the market," said Jordi Ribas, Microsoft's corporate vice president for search and AI, in February 2026. But that is a vendor describing its own reach, not a measurement.
What is the highest-leverage edit for AI citations?
What to do with this today is short. Give every question you want to win one direct-answer sentence at the top of its section — Peec's demonstrations and established entity engineering practice converge on exactly that edit. Name your subject in every section, because a passage gets judged alone, and "it" and "this tool" mean nothing to a judge who just walked in. And keep the entity layer current — schema, profiles, corroborated identity — because the two judges who can see it are the ones who control the door and the credit.
The reranker grades the sentence. Everything around it grades you.
The reranker guide discussed here is Peec AI's "Rerankers for GEO/AEO", by Metehan Yesilyurt (August 6, 2026) — worth your time in full. Bing's rules quoted here are the Bing Webmaster Guidelines (rewrite reported February 26, 2026; quotes verified against the live page October 21, 2026) and Bing's Marking up your site help page.
Quantum Quill Digital helps personal brands, academics, SMBs and marketing teams build entity-first visibility across search and AI. Authority, Engineered.