Category Creation Versus Category Entry When AI Defines the Landscape

Category creation and category entry used to be a brand strategy question, weighed against how much capital a company could burn before payback. That calculus still matters, but a second variable now sits on top of it: whether an AI system can figure out what category a business belongs to. AI doesn't browse a shelf of options the way a shopper scans a store aisle. It resolves ambiguity by picking whoever maps most cleanly onto a category it already recognizes, and everyone else disappears from the answer. Kantar's "Winning in 2026" report states that brands now compete to be chosen inside AI-led journeys, and winning that competition requires content and signals that read as legible and credible to a model making the choice on a buyer's behalf. This piece is about how to make the creation-versus-entry decision with that legibility as a measurable input, not about picking a clever category name.

How AI systems assign categories to businesses they encounter

There's no shelf manager sitting inside a chatbot or an AI assistant deciding where a brand belongs. A paper from September 2026 (arXiv:2609.16304) describes large language models as "emergent choice architects": at the moment someone asks a question, the model builds a small set of alternatives on the spot, with no ranking to inspect and no fixed placement to lobby for. The category gets inferred, in real time, from patterns baked into training data and whatever gets pulled from the live web during retrieval.

That inference process rewards categories that already exist and punishes categories that don't. Research from xeo.marketing makes the point bluntly: AI doesn't invent new categories, it reinforces the strongest ones already sitting in its training data. A company that describes itself as "redefining X" often gets read by the model as a business with no fixed home, which in practice means "we don't fit anywhere yet."

That behavior comes from a fairly mechanical process. LLMs build internal representations of brands, concepts, and problems based on how those things relate to each other across everything the model has read. Fragmented messaging, or a generic label slapped on a genuinely different product, tends to group a brand with its nearest competitors, and the model ends up treating it as a derivative version of something that already exists.

Three signals drive that grouping: the language patterns already dominant across the web, the market archetypes the model has learned to recognize, and repeated contextual cues, meaning site structure, headings, how other sources describe the company. Humans forgive fuzzy positioning all the time; a prospect might sit through a rambling pitch and still get the idea. AI doesn't extend that same patience. Ambiguity, to a model resolving a query in milliseconds, doesn't read as nuance. It reads as absence.

The ambiguity penalty: what happens to brands AI cannot place

Research reported by almcorp.com looked at 1,000 enterprise brands and found 62% of them invisible to generative AI models, despite 94% of those same companies pouring real budget into traditional SEO. That gap is the whole story in two numbers: the work companies have been doing for two decades to rank on a search engine doesn't transfer to getting named by an AI assistant. The signals are different, and category ambiguity is one of the biggest reasons the transfer fails.

Retrieval itself is also more porous than most marketers assume. A separate paper (arXiv:2609.05059) found that across AI engines answering without live web search, the median brand repertoire per engine ran between 15 and 31 organizations, and even after 15 repeated runs of the same question, engines were still surfacing brands they hadn't named before, in 86% to 92% of question-engine combinations tested. The set of brands an AI system might name is nowhere close to fixed after one query. There's room to break in. But that room only opens up for brands the model can categorize cleanly enough to retrieve.

The stakes for missing that window keep rising. AI search traffic grew 527% year-over-year heading into 2026, which means the surface where ambiguity gets punished is expanding fast, not staying flat. And the nature of an AI citation is different from a search results page. According to almcorp.com, when an AI Overview or a chatbot names a brand, it's closer to an endorsement than a ranked listing, often naming just one, two, or three companies total. The model is staking its own credibility on each name it surfaces, which makes it more selective, not less.

Positioning that shifts every quarter doesn't help. Neither does leaning on words like "platform" or "engine" that could describe a thousand different products, or a category name that sounds sharp in a pitch deck but carries no context a model can latch onto. Those choices confuse humans a little. They confuse machines a lot, and machines are doing more of the deciding.

Why category creators face a different AI legibility problem than category entrants

A company entering an established category starts with a real advantage: the model already has a strong pattern to match it against, because the category anchor is sitting in the training data before the company even launches. Retrieval becomes a matter of signal strength, not invention. The risk runs the other way, though, since the model's default answer tends to favor the incumbent. A new entrant has to supply differentiation cues sharp enough to get retrieved even when it isn't the obvious first name. Diagnostic probes in the arXiv:2609.16304 paper found that brands left out of a model's default recommendation list could still surface when the query supplied distinctive enough cues, meaning conditional retrieval is possible, just not automatic.

Category creators face something closer to a blank canvas, and a blank canvas is not a gift in this context. There's no prior pattern for the model to reach for. "We're redefining X" isn't a retrievable position, because a model can't retrieve a category it has no representation of. The creator's job is to anchor to something the model already understands, then narrow that anchor down to the specific thing being built. xeo.marketing lays out a three-step version of this: anchor first to a known category, since that's the retrieval entry point; narrow the context to the specific industry, buyer role, or use case, rather than narrowing the vision itself; and reinforce that narrower frame everywhere, consistently, because if the About page, the product page, and the press coverage each describe the company differently, the model won't reconcile the differences on its own. It defaults to silence.

Category creators also carry a clock that entrants don't. Foundation model providers and platform incumbents are moving fast to reframe emerging AI categories as mere features of their own existing products. A category creator that takes too long to lock in outside recognition, analyst coverage, press language, community discussion, risks getting absorbed into someone else's platform story before the category ever becomes legible to AI.

The structural difference is this. Category entrants need to out-signal incumbents inside a frame the model already has. Category creators need to build the frame itself, in language the model can already parse, which is a slower and riskier project by design.

The decision rule: which path AI legibility favors

One heuristic cuts through most of the noise: a new way to do a known job favors entry, and a job that wasn't possible before favors creation. That's not just a branding preference. It maps directly onto AI legibility, because a known job already has a retrievable category anchor sitting in the training data, while a genuinely new job doesn't yet.

Entry, when the capability gap against incumbents is wide enough, can move fast. Cursor reportedly crossed a substantial annual recurring revenue milestone faster than any category creator has managed at a comparable scale, and it did so without inventing a new category name. It entered a space AI models already understood well, and the capability gap against existing tools was wide enough that buyers self-selected toward it without needing new vocabulary explained to them.

Creation costs more, and takes longer, almost by definition. Drift, Gong, and Datadog reportedly each burned through a hefty sum before reaching payback, and that was before AI legibility was even part of the equation. Building a category's presence in a model's training data from nothing adds work, and the timeline stretches further still.

A simple test helps sort which path a business is actually on. Ask how an AI system would describe the company in one sentence today. If answering that question requires inventing vocabulary the model doesn't already have, the category doesn't yet exist in its training data, and creation requires anchoring work before any other positioning move makes sense. If the answer instead defaults to naming a competitor, the category already exists, but the brand's own signals are too weak to get named inside it, so the fix is sharper differentiation, not a new frame.

None of this is a permanent fork, either. Research on needs-based queries (arXiv:2609.16304) found that framing a query around a user's specific goal changes which brands surface, which means a brand can position around a precise buyer problem and get retrieved even without owning a category label yet. In practice, a company can enter an existing category to secure AI legibility now, while quietly building the signal base that eventually lets it carve out and own a narrower sub-category later.

The signals that make a category position legible to AI systems

A paper (arXiv:2606.25787) analyzed 167,551 URL-grounded citations across 128 brands and found that 85% of brand mentions AI systems draw on come from third-party pages, not from the brand's own website. Encyclopedic sources, market-specific outlets, independent coverage: that's where models go to figure out what a brand actually is. A homepage can say whatever it wants. If the surrounding web doesn't echo the same category language, the model has little reason to trust it.

Stability compounds that effect. Research from AirOps' "The 2026 State of AI Search" found that brands earning both mentions and citations show a 40% higher likelihood of reappearing across different AI answers, yet only 28% of answers include brands with that dual visibility. Most brands get one or the other, rarely both, and that gap appears directly in how often they get named again.

Page structure affects citation rates. The same AirOps research found sequential headings and rich schema markup correlate with a markedly higher citation rate, while pages that go a full quarter without an update are three times more likely to lose citations they'd previously earned. AI legibility isn't a one-time setup task; it decays if the content sits still.

Community platforms carry more weight in this system than most companies assume. The AirOps data shows roughly 48% of citations trace back to sources like Reddit and YouTube, which makes forum threads and video content primary material for how a model understands a brand's category, not a secondary or nice-to-have channel.

Trust signals produce all of it. In the AirOps research, the overwhelming majority of AI Overview citations came from sources carrying strong E-E-A-T signals (experience, expertise, authoritativeness, trust), and pages referencing 15 or more recognized entities showed a substantially higher chance of getting selected. Retrieval is running on entity relationships now, not simple keyword overlap, and consistency ties the whole system together: if a brand's own site, its press coverage, and third-party mentions each use different category language, the model has no clean signal to reconcile. It defaults to silence, or to naming whichever incumbent already owns that space cleanly.

Measuring whether your current category position is AI-legible

The AirOps research found that only 30% of brands stay visible from one AI-generated answer to the next, and just 20% hold visibility across five consecutive runs of a similar query. Most companies have no idea where they land on that spectrum, because almost nobody has been checking.

Measurement needs to go further than a simple yes-or-no on whether a brand gets named. The real question is what category context the model places it in when it does. Getting named as a competitor trailing an incumbent is a fundamentally different outcome than getting named as the defining example of a narrower sub-category, even though both would technically count as "visibility" on a shallow check.

A useful way to structure that measurement runs across three lenses: how models categorize and retrieve the brand; the structured signals, entity recognition, and citation sources feeding that categorization; and reviews, sentiment, and community conversation. Checking only one of the three gives a distorted picture, because a brand can look strong on human sentiment while remaining nearly invisible to the models actually doing the recommending.

The empirical starting point is a set of diagnostic queries: category-level questions, needs-based questions framed around a specific buyer problem, and competitor-alternative questions, run across multiple AI engines. That process shows where a brand gets retrieved, in what category context, and which signals are pushing it in or keeping it out. Measurement has to come before optimization here, because there's no way to fix category clarity a company hasn't first scored. Done properly, it produces a prioritized list: which third-party sources are missing, which entity relationships are weak, which pieces of category language contradict each other across the web, so the first fixes made are the ones carrying the most leverage.

How brands with clear category positions build on that legibility over time

A category position that's legible today isn't guaranteed to stay that way. Research on incumbent advantage (arXiv:2606.17443) found that under tighter retrieval windows, with fewer chances for a brand to get surfaced, survival rates in AI answers drop sharply. A position built once and left alone tends to erode, not hold steady.

For category creators, the payoff compounds once outside sources start repeating the sub-category label back. Once analyst writeups, press coverage, and community discussion begin using the same narrower term the creator introduced, AI models start reinforcing that language on their own. The early, expensive work of anchoring and narrowing starts paying dividends as the label spreads outside the company's own channels.

For category entrants, the sequence runs the other direction. Enter on a known anchor first, build differentiation through earned media and community presence over time, and only introduce a narrower category frame once outside sources have already started echoing it. That order matters, since asking a model to recognize a frame that doesn't yet exist in its training data is asking for the same failure category creators face on day one, just self-inflicted later.

Kantar's "Winning in 2026" framing adds a cultural layer: brands that win build meaningful difference at scale while staying coherent across fragmented, co-created environments where consumers, creators, and platforms all shape the story together. That coherence across touchpoints is what AI systems use to decide a category membership is stable enough to trust.

None of this is a set-once decision. AI search answers for brand-relevant queries shift as training data updates, as competitors adjust their own signals, and as the underlying models themselves get retrained. Brands that hold their category position over time are the ones that catch signal erosion early and respond to it, treating what AI systems say about them as a discipline to monitor continuously. The creation-versus-entry choice, in that light, isn't a decision made once at launch. It gets revisited every time the competitive field, the training corpus, or the brand's own signal profile moves, and the companies that treat AI legibility as an ongoing measurement practice are the ones still getting named when it counts.