Three Strong Articles. Three Different AI Visibility Gaps.
Published on August 9, 2026 at 7:37 PM
AI visibility problems are easy to imagine when content is poorly structured. A vague headline, weak organization, missing context, or thin answers give an audit obvious places to look. But strong content … well, that presents a more interesting challenge.
I recently ran the same AI Discovery Gap Analysis across three B2B manufacturing articles from Thomas, Manufacturing Business Technology (MBT), and Dassault Systèmes' DELMIA. Each article was evaluated across the same seven dimensions: intent alignment, answer extractability and structure, query coverage, topic depth, trust and authority, semantic entity clarity, and differentiation, and each had clear strengths in many of those areas. Their scores reflected how well each article performed across those dimensions:
Thomas: 88/100 MBT: 87/100 DELMIA: 92/100
A five-point range separated all three, and if the goal had been to determine which article was “best optimized,” the analysis could have ended there with a mildly interesting ranking. But the scores weren't the interesting part. The problems were.
Each article had a different AI visibility gap, and those gaps appeared at completely different levels of the content. One emerged at the level of the article's central promise. Another appeared inside individual passages. The third required looking deeper into definitions and evidence.
The experiment revealed something more useful than a winner: strong content can contain the right information, use many of the right structural elements, and still create unnecessary work for an AI system trying to understand and reuse what the page knows.
The Gaps Appeared at Different Levels
Before running the audits, I already had some ideas about what I might find. That was partly the point of using three articles from the same broad industry. Keeping the environment relatively consistent made it easier to see whether the same diagnostic would uncover different problems.
The results didn't line up perfectly with my expectations, which ultimately made them more useful.
Thomas had excellent retrieval architecture, but its biggest answer required more synthesis than many of its smaller ones. MBT fulfilled its central thesis extremely well, while valuable secondary answers were less visible within the article's narrative structure. DELMIA had a strong thesis, question architecture, and answer placement, which forced the audit deeper into how foundational concepts were defined and quantitative evidence was supported.
In short, the articles weren't missing the same optimization feature. They were asking AI systems to do extra work in different places. Let’s take a closer look at each of them.
Thomas: When the Smaller Answers Work Better Than the Biggest One
Thomas' How Boston Dynamics Is Leading the Robotics Revolution is structurally impressive. The article uses descriptive headings, dedicated sections for Spot, Stretch, and Atlas, substantial entity context, commercial examples, authoritative sourcing, and an FAQ with concise answers. Ask a specific question and the article performs very well.
What does Spot do? The answer is easy to find. What does Stretch do? Just as easy. Who owns Boston Dynamics? Which industries use its robots? Again, the article provides clearly identifiable answers. Then consider the question sitting above all of them:
How is Boston Dynamics leading the robotics revolution?
The evidence exists throughout the article. Boston Dynamics' work in advanced mobility, commercial deployment, industrial automation, humanoid robotics, and AI research all support the headline's premise. The problem is that those pieces never fully converge into one clear answer. A reader can make the connection, and an AI system can probably make it too, but both have to assemble the answer from those pieces. That creates an unusual situation.
Thomas makes many secondary answers remarkably easy to retrieve while requiring more synthesis to answer the primary question promised by the headline. The gap therefore isn't a lack of useful information or poor structure; the individual pieces work. Their relationship to the article's central question, however, could work harder.
MBT: When the Answers Exist but Their Boundaries Don't
MBT establishes its central idea early and carries it through the article. AI is presented as a workforce multiplier that can support people rather than simply automate their work. Examples involving worker safety, accessibility, labor shortages, human-AI collaboration, embodied AI, and workforce adoption continue developing that argument.
The central thesis isn't difficult to find at all, but many of the article's secondary answers are.
MBT contains useful information for questions such as: Which manufacturing tasks are best suited to embodied AI? How can AI support manufacturing workers? What does a manufacturer need to scale embodied AI? How can manufacturers improve worker acceptance of AI?
Those questions are answered. But several answers sit inside larger narrative sections without strong structural boundaries showing where one answer begins, where it ends, or which question the information most directly addresses. The audit produced a useful distinction:
Query coverage was high. Query signaling was lower.
That's more than a formatting issue. The article contains the knowledge someone might be looking for, yet its architecture doesn't always make the relationship between that knowledge and the corresponding question equally visible.
Thomas made the smaller answers easy to identify while leaving its largest answer more distributed. MBT made its largest idea clear while leaving some of its smaller answers embedded in the narrative. Same broad industry. Similar diagnostic score. Almost the opposite visibility problem.
DELMIA: When Strong Architecture Pushes the Audit Deeper
The third audit was the one that challenged my expectations.
The article uses natural-language questions as major headings and generally answers those questions immediately. It establishes why general generative AI can struggle in critical manufacturing environments, explains how virtual companions connect users with governed industrial context, addresses the loss of tribal knowledge, and discusses KPIs and ROI.
The article's central progression is also unusually disciplined. The trust problem leads into governed industrial context, virtual companions, virtual twins, workflows, knowledge preservation, measurable outcomes, and eventually industrial action.
Once those larger relationships checked out, the audit had to move deeper.
One gap appeared around the foundational definition of a virtual companion. DELMIA explains what virtual companions can do, how they interact with other technologies, and why they matter in sophisticated industrial environments. Yet the strongest definition of the concept isn't necessarily positioned where the article explicitly asks what an industrial virtual companion is.
The other gap appeared around evidence. The article emphasizes traceability as an important requirement for trustworthy industrial AI, then presents several highly specific performance and ROI figures whose visible provenance could be clearer.
Neither problem requires rebuilding the article, but both require strengthening particular relationships:
entity ↔ definition
and
claim ↔ evidence
That was the point where the three audits began to look less like separate content evaluations and more like different examples of the same underlying issue.
The Problem Isn't the FAQ
AI visibility advice often becomes a checklist. Add question headings. Provide direct answers. Include an FAQ. Clarify entities. Strengthen authority signals. Break complicated subjects into smaller sections. Those practices can absolutely help. All three articles demonstrate some of them successfully. But their presence alone doesn't tell us whether the information is doing the job we expect it to do.
Thomas already has an FAQ. That doesn't solve the distributed answer to its central question. MBT already contains direct answers. That doesn't automatically make their boundaries visible. DELMIA already uses question headings exceptionally well. That doesn't guarantee that the strongest definition sits beneath the most relevant question.
The structural feature isn't the end of the analysis.
An FAQ doesn't help merely because it's an FAQ. A heading doesn't help merely because it's a heading. A direct answer isn't useful merely because it exists. What matters is whether the right information is connected to the right question at the right structural level.
That changes the way an audit has to look at content, so the question is no longer simply whether the pieces exist. The audit has to examine what those pieces are connected to.
AI Visibility Is Also About Relationships
Once I looked across all three articles, the pattern became easier to see.
A heading creates a relationship with the passage underneath it. A question creates an expectation for an answer. A definition establishes the meaning of an entity before other ideas depend on it. Evidence supports a claim. Individual sections contribute to the larger promise established by the article. Those relationships can be represented simply:
question ↔ answer
heading ↔ passage
entity ↔ definition
claim ↔ evidence
section ↔ central intent
AI systems don't encounter information as a collection of optimization features. They have to interpret how the information fits together. When those relationships are clear, less reconstruction is required. When they aren't, useful information may still be present, but another layer of interpretation sits between the information and the question it could answer. That's why a strong article can still have an AI visibility gap without having an obvious content-quality problem.
What an Audit Sees That a Checklist Doesn't
A checklist is useful for confirming presence. Does the page contain descriptive headings? Are important entities defined? Does the article answer likely questions? Is evidence included? Are there passages that can stand on their own? Those checks have value, but they don't tell us whether the pieces connect.
An audit has to go further.
Does the most important answer actually resolve the article's primary question? Can a useful passage be recognized as an answer without reconstructing the surrounding narrative? Does a definition appear before more complicated ideas depend on it? Can a quantitative claim be traced to the evidence supporting it?
Thomas, MBT, and DELMIA would all perform well against a simple list of desirable content features. Their scores suggest as much. Yet they don't need the same fixes.
Thomas needs stronger synthesis around its central promise. MBT needs clearer boundaries around some of the useful answers already inside its narrative. DELMIA needs greater clarity at specific points where definitions and evidence connect to the ideas they support. Applying the same optimization template to all three would miss what the audits actually found.
Same Scores. Different Problems.
Three strong manufacturing articles landed within five points of one another, which sounds like the conclusion until you look beneath the scores. The more useful finding was where each article created unnecessary interpretive work.
Thomas did so at the article level. MBT did so at the passage level. DELMIA pushed the analysis down to the level of definitions and evidence. None of those findings suggests that the articles are weak. Quite the opposite. Their overall strength is what made the differences visible.
The experiment also changed the question I think is worth asking when auditing content for AI discovery. Finding an AI visibility gap isn't always about identifying what's missing. Sometimes the better question is:
What's already here, and how well does it connect?