Structured data is tempting because it looks precise. A field says the business type, address, language or service. The harder question is whether an LLM answer visibly follows that field, or simply keeps reading the surrounding page.
A French training provider adds structured data to its site. The page already says it offers safety certification courses for small construction firms. The new markup identifies the organisation, location, course category and contact details. In a later live-search answer, the model describes the company more cleanly. In a non-browsing answer, nothing changes. In one French prompt, the answer gets the category right but still pulls an old address from a directory.
Levertrace Lab treats this as a composite scenario, not a claim about one named company. It appears often enough to deserve a careful test. Structured data has the feel of a tidy drawer: labels, fields, properties, nested objects. LLM answers, by contrast, often behave like someone searching a workshop bench with one hand. They may find the labelled drawer. They may grab the nearest tool.
Markup changes the evidence trail, but not always the answer
Structured data is machine-readable page information because it states entity facts in a format systems can parse more directly than ordinary prose. That definition is deliberately modest. It does not say structured data forces an LLM to use the fact. It says the public source now contains a more explicit signal.
The lab begins with that modesty because schema markup attracts overconfident claims. A business adds organisation markup, local business fields, service information or course details, then expects model answers to align. Sometimes the visible answer improves. Sometimes the answer already had the right fact because the prose was clear. Sometimes the markup is correct, while old third-party sources continue to shape the description.
For this work-item, Levertrace Lab keeps structured data separate from page copy. If a company adds schema and rewrites the visible page at the same time, later movement cannot be cleanly attributed. The test becomes a bundle. That may be sensible for implementation, but it is poor for interpretation. A structured data test needs the page text, directory state and language pages held as steady as possible.
The team also inspects whether the markup repeats facts visible to a human reader or introduces facts that the page prose does not state clearly. Those are different cases. If schema says “software company” while the page says “digital partner,” and the answer later says “software company,” markup becomes a plausible source influence. If both markup and page say the same thing, the answer may be following either. The lab can record movement, but attribution stays cautious.
The visible page still matters
A recurring finding in Levertrace observations is that structured data rarely rescues a page that hides its own category. Markup may state the organisation type, but if the human-visible text remains vague, model answers often keep a vague shape too. That does not mean markup is useless. It means the structured signal may be only one part of a larger source condition.
A business page is read in layers. The title, headings, introductory sentences, repeated claims, internal links, directory traces and structured fields may all sit near the answer. The lab cannot assume which layer a given system used unless the interface exposes enough evidence. When it does not, the answer is marked source-uncertain.
The composite Object B is useful here. A French B2B software company has English landing pages, French documentation, trade listings and partner summaries. Suppose its schema says it is a software application for maintenance planning, while the English hero still says “operations intelligence for industrial teams.” A later answer that says “operations intelligence platform” has probably not adopted the schema wording in any visible way. An answer that says “maintenance planning software” may be showing adoption, but the lab still checks whether that phrase also appears in page text or partner summaries.
For local service companies, the same problem appears with location and category. A LocalBusiness markup field may contain the correct city. An older directory may list a previous suburb. A model answer may adopt the service category but keep the old address. That is not a clean structured data win. It is partial echo with a source conflict around location.
Structured data is best read as a clarifying layer, not a substitute for visible truth. If the prose is muddy and the markup is precise, the public evidence trail becomes split. If both say the same clear thing, the lab has a stronger source environment, though still not a guaranteed answer change.
What the lab tests after schema is added
Levertrace Lab records the before state in ordinary language. What did the model say before the markup? Which facts were wrong, vague or missing? What did the page text already state? Which third-party sources repeated or contradicted those facts? Only then does the markup change become a lever.
The after-runs use the same prompt family where possible. A general description prompt tests whether the business category changes. A service prompt tests whether the service boundary becomes clearer. A location or language prompt tests whether the structured fields appear in the answer. The lab compares French and English conditions when the site carries both languages, because structured data may sit on one locale while answers are requested in another.
The anchor classification again prevents the result from becoming mush. A structured data lever may produce adoption, partial echo, source conflict or no visible movement. Adoption would mean the answer visibly carries the structured fact in a stable way under the tested condition. Partial echo might mean the answer adopts the category but misses the service boundary. Source conflict appears when markup-aligned facts collide with older page or directory facts. No visible movement means the answer does not reflect the markup in any meaningful observed way.
A small example shows why the categories help. A French course provider adds markup identifying “safety training” as the course category. The page already says “training for construction teams,” while an old directory says “general professional coaching.” A later model answer says the company offers “professional training, including safety courses for construction workers.” That is not pure adoption. It carries the new structured idea, but the old broad category still frames the answer. Partial echo is the cleaner label.
In another scenario, the markup adds the business’s current address, but the model keeps the old address from a business profile. That is no visible movement on the location fact, even if the category improves elsewhere. Levertrace Lab avoids giving the markup one overall verdict when different facts behave differently.
Why structured data can look more causal than it is
Structured data is neat. That neatness can seduce the observer. A field changes, then an answer changes; the mind draws a straight line between them. Levertrace Lab slows that line down. The answer may have changed because the visible page was re-crawled, because a live-search system retrieved the page, because a directory changed independently, or because the prompt wording nudged the model toward the new field.
The lab is especially careful when a site deployment includes several quiet changes. Developers often add schema during a broader technical cleanup. The title may change. Headings may be simplified. Internal links may be repaired. The sitemap may be refreshed. A business profile may be corrected in the same period. Each of those can alter the source environment. If the research question is schema-specific, those changes make the case harder to read.
Another problem is that structured data often restates facts already visible on the page. That is good practice for coherence, but it makes attribution difficult. If both the paragraph and the markup say the company provides “maintenance planning software,” a later answer using that phrase could be following either layer. The lab may still treat the change as an observed shift after a structured data update, but the conclusion must say “possible source influence,” not “schema caused the answer.”
There is also a difference between search-result behaviour and LLM description behaviour. Structured data may support search features, entity understanding or page parsing in systems built to use it. The lab does not collapse those systems into one effect. The question here is narrower: did the LLM answer about the business visibly change?
This distinction matters for agencies. A technical improvement can be worth doing even when it does not produce a visible LLM wording shift in the test. Clean markup may support other discovery systems, reduce ambiguity for parsers, and align facts across pages. Levertrace Lab simply refuses to count those benefits as observed answer movement unless the answer itself moves.
What structured data is most likely to clarify
In the lab’s observed patterns, structured data is most useful when it reinforces a fact that is also visible, stable and contested enough to matter. Organisation identity, category, address, service area, course type and same-as relationships can all become testable facts. The key is that the fact must be specific enough to watch.
A vague markup field does little. If a business marks itself broadly as an organisation and leaves the page copy equally broad, there may be no meaningful answer change to observe. If the markup identifies a course, local service, software product or professional category that was previously ambiguous, the test becomes sharper. The model may or may not use it, but the observation has a clear target.
For bilingual sites, structured data can also expose mismatch. A French page may include a category in markup, while the English page uses a looser phrase. If the English answer adopts the French-aligned category, the lab notes possible language transfer or source influence, depending on the visible condition. If the English answer keeps the looser phrase, the structured signal may not be carrying across language conditions.
The lab prefers consistency between markup and visible copy. Hidden precision is fragile. A field that says one thing while the page says another creates a conflict that may confuse human auditors as much as machine readers. A structured data change should make the evidence trail cleaner, not quietly contradict it.
This is where structured data differs from a simple page wording edit. A sentence can persuade a human reader and feed a model phrase. Markup mostly clarifies the skeleton. When the skeleton and the visible body disagree, the answer may limp.
Limits of schema-focused observation
Levertrace Lab cannot prove that a model used a structured data field unless the system exposes that path or the surrounding evidence makes a narrower explanation plausible. Most of the time, the team observes answer behaviour, compares source context and labels attribution carefully. “Possible source influence” is often the honest ceiling.
The method also does not show whether a search engine benefited from the markup. That is a separate question. This material studies LLM business descriptions: what the answer says, which phrase choices appear, which facts move and which conflicts remain. A technical validation tool can confirm that schema is syntactically present. It cannot confirm that an LLM description has changed.
Schema tests are strongest when the markup is the main changed lever and the target fact is clear. They are weakest when markup arrives inside a broad site refresh or when the answer shift could be explained by visible text alone. In those cases, the lab still records the observation, but the conclusion stays narrow.
The practical reading is sober. Adding structured data can make a French business fact easier for systems to parse, and it may coincide with cleaner LLM descriptions under some conditions. The lab only calls it answer movement when the before-and-after wording shows a visible shift against the page text, source condition and surrounding public trail.