How Content, Entities and Category Framing Shape AI Visibility

You have useful content, a clean About page and valid organization markup. Yet your brand still disappears when someone asks an AI assistant for options in your market. The missing piece may not be authority. The system may know who you are without considering you eligible for the category named in the prompt.

You can diagnose that problem by separating three jobs: establish the category in which you belong, make the relevant entities and relationships unambiguous, and publish evidence that supports recommending you for the user’s task. That distinction turns AI visibility from a vague branding exercise into work you can assign, test and improve.

Key takeaways

• Brand recognition and recommendation eligibility are different. An AI system can identify your company accurately and still exclude it from an unbranded category answer.
• Choose category language before planning content or schema. Your primary category should describe what you sell now; adjacent categories should reflect real customer language and a defensible part of your offer.
• Build an entity map before building more pages. It should connect your organization, offers, audiences, problems, methods, people, proof and category claims.
• Use JSON-LD to declare facts that visible content already supports. Schema can reduce ambiguity, but it cannot manufacture relevance or compensate for missing evidence.
• Category association is also built away from your website. Relevant reviews, editorial coverage, comparisons and co-mentions help establish the contexts in which your brand is considered.
• Measure recognition, category eligibility, recommendation and supporting evidence separately. A single visibility score hides the reason you are being omitted.

First, determine whether you have a recognition or category problem

Start with two prompts that look similar but test different things:

• Recognition prompt: What is [Brand], and what does it offer?
• Category prompt: Which [category] providers should [audience] consider for [task]?

If the first answer is accurate and the second omits you, rewriting your About page again is unlikely to address the main constraint. Your entity is recognized, but it is not being retrieved or selected in that category context.

The effect of category wording can be substantial. A controlled test covering 14,140 API runs across ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews evaluated 12 athletic apparel brands in the U.K. over seven days. Changing the category from athleisure to athletic footwear moved New Balance from a 1% appearance rate to 90%, while lululemon moved from 90% to 0%.

That is strong evidence that framing controlled recommendation behavior in that test. It is not a universal performance benchmark: one market, one prompt design and one testing period cannot establish how every model will treat every category. The practical lesson is narrower and more useful. Test the category noun instead of assuming that general brand strength transfers across every way a customer might describe your market.

Define one primary category and a small set of adjacent frames

Your primary category should be the plainest accurate answer to: What kind of provider, product or organization is this? An adjacent frame is a different but truthful way a buyer may classify the same offer. For example, a platform may belong firmly to one software category while also serving a narrower workflow, audience or outcome category.

Do not collect every loosely related label. For each candidate category, record:

• Customer language: Do real buyers use this term when expressing the need you solve?
• Offer fit: Can you point to a current product, service or capability that makes the label true?
• On-site evidence: Is the category explained on a crawlable page, or does it appear only in a slogan?
• Independent evidence: Do credible third parties discuss you in that context or alongside established members of the category?
• Decision value: Would visibility for this category attract the audience and use case you actually want?

Then write a control sentence: [Brand] is a [primary category] for [audience], helping them complete [task] through [offer or method]. Treat this as an editorial constraint, not a slogan and not a Schema.org type. Every element must be demonstrably true, and the same relationship should be understandable from your core pages.

Build an entity map that gives every page a job

Once the category is chosen, map the things a search system must connect to decide that you belong. Entities are not limited to your company and founder. They include products, services, people, audiences, locations, problems, methods, features and other identifiable concepts. The useful unit is not an isolated noun; it is a relationship that helps explain the brand.

Create an entity ledger with one row for each important relationship:

• Subject: the organization, person, offer, category, audience or problem being described.
• Relationship: offers, serves, solves, teaches, authored, includes, supports or another accurate connection.
• Object: the entity on the other side of that relationship.
• Visible evidence: the page and passage where a reader can verify the claim.
• Structured declaration: the standards-supported markup, if any, that can express it accurately.
• Independent corroboration: a review, profile, comparison, citation or other external evidence.
• Gap: missing, vague, contradictory or fully supported.

Use those relationship words as planning labels. They are not automatically valid Schema.org properties. Your conceptual model can and often should be richer than the standardized vocabulary you publish.

This distinction matters in specialized markets. One higher-education framework found that 23 existing Schema.org entities were insufficient and added more than 60 domain-specific concepts to represent a prospective student’s journey. You can use a custom ontology internally to expose content gaps without pretending that proprietary terms are recognized Schema.org vocabulary.

Turn the map into a content system, not one oversized page

Assign each important relationship to a canonical page. Your About page should establish organization identity and positioning. An offer page should explain what the offer does, whom it serves and how it differs. A method page should explain the process. A use-case page should connect a specific audience and task to the offer. An author page should establish the person behind relevant expertise. Supporting resources should answer the questions that arise before and after the main decision.

This division helps with the way AI search may expand a request. A query can trigger related searches across subtopics and data sources so that the system can assemble an answer to the broader task. A buyer asking for a category recommendation may also need selection criteria, implementation details, limitations, alternatives, audience fit and next steps. One page does not have to answer everything, but your site should make the connections explicit.

Use this brief for every page you keep or create:

• Page job: State the single decision or question this page resolves.
• Primary entities: Name the organization, offer, audience, problem and category involved.
• Direct answer: Put the answer near the beginning in visible text. Do not make a reader infer it from a slogan, image or schema block.
• Boundary: Explain who or what the answer is for, where it applies and what it does not cover.
• Evidence: Support claims with concrete capabilities, examples, authorship or other facts you can substantiate.
• Related questions: Link to the next useful pages with anchor text that describes the relationship, rather than generic text such as learn more.
• Duplication check: Merge or differentiate pages that make the same claim about the same entities without serving different intents.

The standard is comprehension, not length. A clear page names its subject, answers the intended question and connects to the next part of the task. More copy only helps when it adds a missing entity, relationship, condition or piece of evidence.

Use JSON-LD to declare truth, not manufacture relevance

Schema is valuable because it can state entities and relationships explicitly in a vocabulary machines already recognize. It is best treated as a declaration layer over a coherent site, not a lever that forces a model to recommend you.

The evidence does not support a simple claim that adding markup produces more AI citations. Microsoft Bing’s Fabrice Canel stated in March 2025 that Copilot uses schema to understand content, while other published tests found no effect on LLM visibility or no direct reading of on-page schema. Those findings measure different things, including machine understanding, direct model access, citations and observed visibility. Treating them as one outcome creates a false yes-or-no debate.

A safer operating position is straightforward: accurate markup can reduce ambiguity for systems that consume it, but visibility remains a downstream result influenced by content, retrieval, category fit and external evidence. Do not promise a citation lift from markup alone.

Implement JSON-LD in this order:

1. Resolve identity first. Decide which organization, people, offers and other entities are canonical. Use stable identifiers so the same entity is not represented as several disconnected things.
2. Confirm the visible facts. A reader should be able to verify every material claim in the markup from the page or an appropriate linked page. Structured data should match what users can actually see.
3. Use established vocabulary where it fits. Choose the most accurate standard types and properties available. Do not force a marketing phrase into a technical type merely because the phrase is commercially important.
4. Connect entities deliberately. Markup should describe a coherent graph rather than produce unrelated blocks for the organization, author, service and page.
5. Keep custom concepts separate. Use your internal ontology to plan coverage and analyze gaps. Publish custom terms only where a consuming system understands that vocabulary; do not misrepresent them as standard Schema.org definitions.
6. Remove decorative markup. If a block exists only to qualify for a feature or repeat keywords, but adds no accurate entity relationship, it is not solving your AI visibility problem.

When markup and visible copy disagree, repair the underlying page first. Otherwise you are making two incompatible claims about the same entity and asking machines to decide which one is true.

Create off-site category evidence, then measure the whole system

Build corroboration in the category you want to earn

Your site can declare its category, but it cannot independently establish how the wider market describes you. Category coding appears to combine an entity anchor with the third-party material accumulated around a brand, including reviews, editorial comparisons, roundups and co-mentions. This helps explain why editing a description does not instantly move a brand into a different recommendation set.

Audit the external evidence for each priority category:

• Which publications, communities and comparison pages appear in AI answers for the category?
• Which brands are repeatedly mentioned together, and what language is used to explain their inclusion?
• Which attributes make a provider category-eligible: audience, use case, product form, method, price position or another verifiable characteristic?
• Where is your brand already mentioned, and which category does that coverage reinforce?
• Does the cited coverage still describe your current offer accurately?

Use the findings to shape public relations and content distribution. Give relevant publishers a truthful reason to place your brand in the target context: a category-specific capability, credible expert contribution, useful case evidence or a clear point of view. A generic mention may improve recognition while doing nothing to connect you to the category that matters.

Do not pursue an adjacent category that your product cannot support. Repetition can amplify an association, but it cannot make a misleading position useful to the customer. Establish the offer and on-site evidence before trying to earn external corroboration.

Measure recognition, eligibility, recommendation and evidence separately

Create a controlled prompt matrix for every primary and adjacent category. Keep the audience, task and wording stable, then change only the category expression you want to test. Run each prompt in a fresh conversation so earlier messages do not supply the brand or category context.

Record these fields for each model and prompt:

• Recognition: Can the system describe your brand accurately when it is named?
• Eligibility: Does the brand appear in an unbranded list for the category?
• Recommendation: Is it merely mentioned, or actively presented as suitable for the audience and task?
• Rationale: Which capabilities, use cases or associations explain its inclusion or exclusion?
• Evidence: Which URLs, publishers or page types support the answer?
• Representation: Are the description, category and sentiment accurate?
• Conditions: Which model, prompt, date and conversation state produced the response?

Do not compress these observations into one score until you have inspected them separately. A brand that is recognized everywhere but eligible nowhere has a different problem from one that is regularly recommended with the wrong description.

Use the pattern to choose the next action:

• Recognition is weak: reconcile identity, core descriptions, canonical pages, profiles and structured relationships.
• Recognition is strong but category eligibility is weak: repair category language and build relevant third-party association.
• Eligibility is strong but recommendation is weak: clarify audience fit, differentiation, limitations and supporting proof.
• Recommendation is strong but evidence is poor: strengthen pages that make the rationale attributable and easy to cite.
• Results differ sharply by category: plan content and outreach for each frame independently instead of treating visibility as a brand-wide property.
• Results differ sharply by model or prompt: preserve the raw responses and gather more controlled observations before declaring a trend.

Prioritize gaps using three questions: Does this category matter commercially? Is the missing association visible across controlled prompts? Can you support it truthfully with your present offer and evidence? A high-volume label that fails the third test is not an optimization opportunity. It is a positioning error.

Start with one primary category and one defensible adjacent frame. Run the prompt matrix, map the entities behind both, assign each important relationship to a page, align visible copy with JSON-LD, and then pursue independent coverage in the context that is still missing. That sequence gives you something more useful than a visibility score: a reason for the result and a specific next move.

References

• Search Engine Land — The new SEO rules for bloggers in 2026: Why clarity matters in AI search
• Search Engine Land — Schema for AI search: How to identify and prioritize entity gaps
• Search Engine Land — How category framing changes which brands AI recommends