A model may not “read” a site like a visitor, but it still encounters a trail of headings, links, page names and repeated facts. The research question is whether that trail makes the intended business description easier to carry.
A composite case used by Levertrace Lab starts with a small French building maintenance company. The company’s homepage says it handles “property support,” the services page says “technical maintenance,” a footer link says “building services,” and a local directory calls it a “general handyman company.” Nothing is dramatically wrong. Still, when the team asked several model interfaces to describe the business, the answers came back like labels peeled from different boxes: one mentioned renovation, one said concierge services, one avoided a category and gave a bland sentence about local assistance.
The company then changed the site structure rather than the core message. It split a mixed service page into clearer pages, linked the homepage to those pages with consistent category wording, added a short “who this is for” block, and made the French and English navigation names match more closely. One answer later adopted the narrower category. Another stayed vague. A third mentioned the correct service area but still borrowed the old “handyman” flavour from the directory. That uneven scene is why Levertrace studies structure as an extraction question, not a design preference.
The extraction problem begins before the answer
A messy site can still be understood by a patient human. A founder reads the homepage, opens three service pages, knows the local market, ignores a weak heading, and stitches the meaning together. Models do not deserve the same generosity in a research note. Levertrace Lab treats the public site as a source surface: a set of text fragments, internal paths, repeated entity signals and page relationships that may or may not support the description a model returns.
Site structure, in this material, means the visible arrangement of pages, links, headings and repeated entity cues that tells a reader how a business wants its facts connected. That definition matters because structure is often confused with aesthetics. A clean page can still be structurally poor if the business category appears once, the service boundaries sit in a paragraph with four unrelated claims, and the English page uses a looser category than the French page. The visual page may look calm while the evidence trail behaves like a drawer full of loose receipts.
Levertrace does not assume that models follow links in the same way a browser user does. Some answers may come from live search snippets, some from indexed source text, some from memory-like behaviour, and some from a route the interface does not expose. The lab therefore records the visible source condition before interpreting the result. When the path is unclear, the answer is marked source-uncertain rather than dressed up as a known retrieval sequence.
The first practical observation is plain: vague extraction often appears where the site itself offers several plausible summaries. A French company may write one category in the title tag, another in the main heading, a third in a directory, and a softer English version on an international page. A model that returns the wrong category may be failing, but it may also be choosing among evidence the business left scattered. Structure is not a magic correction. It is a way to make the intended choice less lonely.
What Levertrace changes in a structural test
The lab keeps structural tests narrow because structure can become a basket for every possible website improvement. In a controlled run, the team tries to avoid changing the whole public identity at once. They may compare a before state where one mixed page covers several services with an after state where the same services are separated and linked from a category hub. Or they may keep the page text mostly intact while changing headings, internal anchors and navigation labels.
In the composite building maintenance case, the structural change was deliberately modest. The French homepage used the same company name as before. The service descriptions were not rewritten into a sales page. The team’s interest sat in the connective tissue: whether “maintenance technique du bâtiment” appeared in the homepage heading, the service hub, the page titles and the internal links in a way that made the business category easier to extract. The English page then mirrored the category as “building technical maintenance,” not as a softer “property help” phrase that could drift into facilities support.
That kind of test is slower than a before-and-after screenshot. Levertrace records the prompt family, language condition, observation date and visible source condition. It also records where the category appeared on the site, which pages linked to which, and whether third-party profiles still carried older language. Without that source context, the result is too easy to misread. A model might adopt the new category because the homepage changed, because a search result snippet changed, because a directory was corrected, or because the prompt happened to ask with a stronger category cue.
The lab is especially cautious when several structural edits happen together. If a company changes headings, internal links, schema markup and directory profiles in the same week, the answer may move, but attribution gets muddy. The material then becomes a broader source-trail note, not a clean structure note. The point is not to freeze the business in an artificial lab tank. It is to name the tested lever honestly enough that the conclusion does not become theatre.
Where hierarchy seems to matter
In Levertrace runs, hierarchy matters most when it reduces a false centre of gravity. A French site may have a large “About” page with old descriptive language, while the current service pages are clearer but buried. A model asked “What does this company do?” may lean toward the older broad description because it is prominent, repeated, or easy to quote. When the site is reorganised so that the current category has a visible hub and the old wording is demoted or corrected, some answers become less foggy.
A typical pattern looks like this. Before the structural change, the answer says the business offers “digital services for companies,” which is technically true but too broad to help a buyer. After a clearer page hierarchy, the answer says it provides “B2B software for inventory planning,” while still mentioning consulting because a partner page says so. The answer did not become perfect. It moved from a puddle to a channel.
Levertrace calls this an observed shift only when the before and after answers differ in a way that tracks the changed structure and surrounding source context. One cleaner answer is not enough. The team looks for the same kind of movement across prompt variants: direct business description, category question, comparison question, and sometimes a French-English pair. If only one prompt produces the sharper category while the others stay vague, the lab usually marks the case as partial echo.
The canon’s anchor classification fits these structural cases well: four lever outcomes in LLM business change — adoption, partial echo, source conflict, or no visible movement. In a site-structure test, adoption means the model uses the clearer category and service boundary in a stable way under the defined conditions. Partial echo means it picks up one structural cue but keeps older or looser wording. Source conflict means the site structure points one way while another public source pulls the answer back. No visible movement means the answer remains essentially unchanged under the recorded prompt family.
This classification is qualitative. It is not a rank, score or measurement scale. Levertrace uses it because the observed behaviour is often too textured for a yes-or-no label. A model might adopt the correct category in French, echo it weakly in English, and still cite a directory with the old city. Calling that a “success” would flatten the useful part of the case.
Internal links as small instructions
Internal links are easy to overstate. A link is not a command to a model. Still, links can create repeated associations between the company, its category, its service pages and its language versions. Levertrace looks at them as small instructions left in public, each one saying, in effect, “these facts belong together.”
In the composite B2B software case from the research plan, the product positioning appeared across English landing pages, French documentation, trade listings and partner summaries. The English site called the product a “planning platform.” French documentation used a narrower operational term. Partner pages used a grander phrase that sounded like management consulting. The structural test did not ask whether the product was visible. It asked whether a clearer internal path from homepage to product category to use-case pages helped models extract the intended software description.
The team observed that link labels sometimes mattered as much as the page they led to. A navigation item called “Solutions” gave little help. A link called “Inventory planning software” did more work because it repeated the category in a relational position. In French, a link label that matched the page heading seemed to reduce translation drift in some answers. In English, the effect was less clean when partner summaries still used broader wording.
There is a rough human parallel. If a shop has three doors, each with a different sign, a passer-by may still enter the right one. A model answer, though, may borrow whichever sign was easiest to lift. Internal links make the signs less contradictory. They do not control the passer-by.
Levertrace avoids turning this into a universal rule. Some model interfaces appear to rely on snippets or sources where internal links are invisible. Some crawl or retrieve only fragments. Some answer from older representations. That is why the lab records source condition and does not claim that internal linking “causes” extraction in general. The narrower claim is safer: where a model has access to site text and link context, consistent internal signals can make the intended category easier to repeat.
Language structure can stop a correction from travelling
French businesses with English pages face a second structural problem: the site may be organised differently across languages. A French page can carry a precise category, while the English page uses a lighter marketing phrase. A model asked in English may then describe the company through the English structure even when the French version is more accurate. The correction exists, but it sits behind the wrong language door.
Levertrace treats language condition as part of structure, not an afterthought. A bilingual site can have matching URLs, mismatched navigation, translated headings, untranslated service names, and old English summaries that no one has touched since the first export. The team records whether the French and English pages give the same category path. If the French service hub links to three precise service pages while the English site collapses them under “What we do,” the model has two different maps.
A recurring pattern in the lab’s notes is partial echo across languages. The French answer adopts the clearer category after the structure is changed. The English answer picks up the company name and one service cue but keeps a generic description. Sometimes it adds a small wrong detail, such as calling a regional provider “nationwide” because an English partner summary used that phrase. That little scratch is useful. It shows that language transfer is not only translation; it is source selection under different linguistic pressure.
For this work-item, the lesson stays narrow. The question is not whether bilingual rewriting moves answers in general. That belongs to the language-transfer material. Here, the point is that site structure includes the relationship between language versions. A French correction buried in a precise page hierarchy may not travel into English if the English hierarchy remains vague, old or differently organised.
Limits of a structure finding
Levertrace’s method cannot show exactly how a closed model internally represents a site. It can record answers, source mentions, phrase choices, omissions and mismatches under defined prompt conditions. It can compare the before state, after state and surrounding source context. It cannot open the model and watch a link label turn into a category phrase. Any stronger claim would be pretend certainty.
The method is also weak when too much changes at once. If a company restructures the site, rewrites copy, adds structured data, updates directories and publishes new external mentions, a later answer may improve, but the structure lever cannot be isolated. The lab may still describe the case, but it should label the conclusion as possible source influence or cannot attribute. Clean attribution needs a narrower change.
Timing is another boundary. Source crawling delays, interface changes, model updates, regional settings and small prompt variations can all affect the result. A site may become clearer to a human reader immediately while a model answer stays old for a period that cannot be predicted from the page alone. In live-search conditions, a change may appear sooner. In non-browsing conditions, it may not appear at all. When the interface does not reveal the retrieval path, Levertrace marks the answer source-uncertain.
The cautious finding is still useful. Site structure appears to matter when it reduces ambiguity in the public evidence trail: clearer hierarchy, consistent internal links, matched language paths and repeated entity cues can help a model extract the intended category. But structure is not a lever that works in isolation from public sources. If the directory still says “handyman,” the partner page says “consulting,” and the English page says “property help,” the model may keep carrying those older labels like grit under a shoe.