Cold-Start Asymmetry: How AI Cites New vs Established Businesses

Two cafes open in the same neighborhood, six months apart. The older one has 184 Google reviews, a Wikipedia stub, a Yelp listing, and three local-press write-ups. The newer one opened last Friday, has its Google Business Profile and a single hand-built website, and that is the entire trail. Ask an AI assistant about “a good place for an afternoon coffee around here” on the same day, and the gap in the answers is not proportional to the gap in quality. It is something stranger. The older cafe is cited with conviction; the newer one is, very often, not mentioned at all. Not because the assistant judged it worse, but because it had almost nothing on which to form a judgment, and it quietly defaulted to silence.

That gap is the cold-start asymmetry, and it is the comparison I want to make precise here. Not “old businesses beat new ones in AI search” (that framing is both obvious and wrong in important ways), but: given the same kind of business, what specifically is different about the signals an established entity has and a newly opened one structurally cannot, and how do AI assistants reason about that absence? The answer turns out to be one of the cleanest applications of the LLMO Framework’s Confidence and Provenance axes I have come across, because the cold-start case strips away every signal that has had time to accumulate and forces you to look at the bare structural minimum the model can lean on.

One disclosure before the comparison goes any further, because the honesty of this line determines the value of everything below it. Everything I say about how the four current AI engines weigh cold-start signals is documented architecture-based inference, not measured citation. I have read the published retrieval and ranking disclosures, and I am reasoning from architecture to behavior. I do not have a cross-engine measured dataset of citation rates for newly opened versus established businesses; nobody outside the labs does. Read this as a structural map drawn from public specifications and the schema vocabulary the engines have to work with, not as a benchmark.

The comparison axis: signal accumulation as a function of time

Name the units, or the comparison drifts. The axis here is signal accumulation over time, and the two endpoints of the comparison are:

• The cold-start business: opened in the last few weeks. Has a website, possibly a GBP listing, and, critically, almost no third-party trail. Review count is at or near zero. NAP has been published in one or two places but not yet replicated. The Knowledge Graph has not had time to acquire or reconcile the entity. There is no operational history for any field to be “fresh” against.
• The established business: open for at least a couple of years. Has hundreds of reviews across multiple platforms, third-party directory listings that agree on its NAP, a Knowledge Graph node Google has had time to settle, and a measurable history of state-field updates (hours, menus, prices) the engines can read as continuity.

The same business (same category, same address, same quality of food) at these two endpoints reads to an AI engine as two structurally different objects. Not “the same object with a lower confidence score,” but two objects whose available signal vocabularies are different shapes. That is the part that the score-thinking model of AI optimization keeps getting wrong, and it is what the rest of this piece tries to make explicit.

The four signals the cold-start business does not have

Decompose what an established business has that a newly opened one does not, and the answer is not “more of one thing.” It is four distinct missing signals, each of which sits on a different axis of the framework, and each of which fails in its own way at the cold start.

Missing signal 1: Aggregate review evidence. The most visible gap. aggregateRating.reviewCount = 0 (or a single-digit number) is not just “low rating.” It is the absence of the input the model uses to estimate whether a positive description from any single source is corroborated or idiosyncratic. The Confidence axis of the LLMO Framework treats review aggregation as a primary signal of community-backed trust, and a cold-start business is producing the exact value (near-zero) that the axis treats as “I cannot weight this.” Worse, the model has no way, from reviewCount alone, to tell why it is zero: bad business, recently opened, or scraping failure all look the same to a count.

Missing signal 2: NAP propagation lag. When a business is created, its name/address/phone exist in one or two places: the website, possibly the GBP listing. Replication into the directory ecosystem (Yelp, Foursquare, industry-specific platforms, in Japan tabelog or hot pepper, plus chamber-of-commerce listings) takes weeks at minimum and is partly a manual process. A newly opened business with a perfectly consistent NAP on its own site still reads, to the engine, as one source corroborated by zero others. The NAP consistency and entity reconciliation work that established businesses spend years getting right is, at the cold start, structurally absent. Not failed, not yet possible.

Missing signal 3: Knowledge Graph wiring. Google’s Knowledge Graph does not have a fixed schedule for acquiring new entities. A new business might appear as a KG node within days; it might take months. Until it does, the assistants that reach the entity through the KG (Gemini most directly, ChatGPT via its index, the others to varying degrees) have nothing to retrieve when they reach for it. The first-party site can describe the business in immaculate detail, and the KG-routed path still returns empty. This is a Provenance-axis gap: one of the three routes a fact can take to an assistant is closed off, not because the data is wrong, but because the graph has not caught up.

Missing signal 4: No history for freshness to be measured against. The freshness signal, the one that tells an engine “this openingHoursSpecification was updated last month, so treat it as current,” works by reading the change history of a field. A field that has only ever had one value, since the entity first appeared, does not register as fresh or stale. It registers as unconfirmed. The closest analogy is a Git repository with one commit: nothing has been wrong with it, but nothing has been verified by being maintained, either. State-field freshness in AI citation, in the schema vocabulary, is a delta over time. At the cold start, there is no delta.

Four signals, four axes of absence. The point I want to draw out is that they are not redundant; an established business that has review volume but no KG node, or KG presence but no third-party NAP replication, has a partial signal profile, not a “newer-feeling” one. The cold-start case is the rare one where all four are absent at once, and that is what makes it diagnostic for the framework.

At-a-glance: signal profile at each end of the axis

Signal dimension	Cold-start (weeks old)	Established (multi-year)
Aggregate review evidence	reviewCount near zero; treated as “cannot weight”	Hundreds of reviews across platforms; primary Confidence input
NAP propagation	One or two publications (website, maybe GBP)	Replicated across directories, cross-corroborated
Knowledge Graph wiring	May not be a KG node at all	Settled KG node with stable identity
State-field history	No delta; single value since creation	Change history the engine reads as continuity

The rows are not additive to a single score. Each maps to a different axis (Confidence, Provenance, entity resolution, freshness), and the cold-start case is diagnostic because all four are absent simultaneously — a shape that is rare in the established population and that no single schema field, on its own, can compensate for.

What AI engines can lean on instead, and what the schema lets you say

The interesting half of the cold-start case is not the absence. It is what can be done structurally, on the first-party side, to make the absence legible rather than mute. The schema vocabulary has a small set of fields that exist precisely to encode “this entity is new” as a positive statement rather than letting it look like missing data.

The first is foundingDate on Organization, or dateCreated on the entity itself. These are dates the publisher attaches, and they let the model reason: signals are absent because the entity is two weeks old, not because the entity is hollow. The distinction matters because the model’s default treatment of a low-signal entity, without that hint, is suspicion. With it, the model can route to a different behavior: one that weights the schema it does have more heavily because there has not been time for corroboration to accumulate.

{
  "@context": "https://schema.org",
  "@type": "CafeOrCoffeeShop",
  "@id": "https://example.com/#cafe",
  "name": "Ueno Beans",
  "url": "https://example.com",
  "foundingDate": "2026-06-20",
  "dateCreated": "2026-06-20",
  "address": {
    "@type": "PostalAddress",
    "streetAddress": "...",
    "addressLocality": "Taito",
    "addressRegion": "Tokyo",
    "postalCode": "110-0005"
  }
}

The second is the event vocabulary. A GrandOpening event (schema.org does not have that exact name, but Event with a fitting name and startDate on the same date as foundingDate) gives the engine a second pointer to the cold-start moment, this time as a discrete public occurrence rather than an entity attribute.

{
  "@context": "https://schema.org",
  "@type": "Event",
  "name": "Ueno Beans Grand Opening",
  "startDate": "2026-06-20T09:00+09:00",
  "location": { "@id": "https://example.com/#cafe" }
}

The third is what not to do with aggregateRating. If reviewCount is genuinely zero, the correct move is to omit the property entirely. The wrong move, which I will say more about in a moment, is to publish a ratingValue with a reviewCount of one or two, because the structure of that combination is exactly the shape engines have learned to distrust as low-signal noise. Missing is honest; thin is suspicious.

The fourth is to keep the few first-party state fields you do have explicitly fresh, with validFrom / validThrough dates on openingHoursSpecification and priceValidUntil on any Offer. These will not substitute for years of history, but they show that the publisher is actively maintaining the small surface that does exist, and that, in the schema’s own logic, is the closest cold-start equivalent to a freshness signal.

This is the layer where the comparison earns its keep. The cold-start business cannot manufacture review volume, cannot accelerate Knowledge Graph acquisition, cannot force directories to replicate its NAP overnight. What it can do is publish the structural hints that turn “missing data” into “data that is missing for a known reason.” The precondition chain is still in force (entity resolution still has to pass, structure still has to be machine-extractable), but at the cold start, declared freshness is one of the few levers that does not depend on time having already passed.

Cold-start cheat sheet: what to publish, what to omit

A practical shortlist drawn from the four axes above, for a business opening within the next few weeks. This is what the cold-start business can do unilaterally, today, before any third-party signal lands:

• Publish foundingDate on Organization (or dateCreated on the entity). Turns “signals absent because entity is hollow” into “signals absent because entity is two weeks old” — the model’s default routing changes from suspicion to weighted-what-exists.
• Publish a GrandOpening-flavored Event. Same date as foundingDate, location pointing at the entity’s @id. Second pointer to the cold-start moment as a discrete public occurrence rather than an entity attribute.
• Attach validFrom (and, where meaningful, validThrough) on openingHoursSpecification. No edit history exists yet, so this is the one structural hint that says I am declaring this current as of this date.
• Attach priceValidUntil on any Offer. Same reasoning — declared freshness on a surface where no observed freshness is possible yet.
• Omit aggregateRating entirely if reviewCount is zero or single-digit. Missing is honest; thin is suspicious. The false-maturity penalty is meaningfully larger than the absent-maturity default.
• Do not pad sameAs with duplicate URLs from the same platform. Cross-platform diversity is what drives source count on the Provenance axis; multiple Yelp URLs collapse to one source and add no weight.

Nothing on this list buys review volume, third-party replication, or Knowledge Graph acquisition. Those depend on time and on independent acts by third parties. The cheat sheet is the intersection of what the cold-start case needs and what the publisher can decide alone.

The established side: how confidence grows along two dimensions at once

Now the other end of the asymmetry, because it is not just “the cold-start business inverted.” The established business’s signal stack does not grow linearly. It grows on two axes simultaneously, and they multiply.

The first axis is time. As months pass, each state field accumulates a history of updates: hours changing for a holiday, then changing back; menu items added and removed; price tiers nudged. Each of those updates leaves a trace the engine can read as evidence of a maintained entity. The engines apply, by all available description of their architectures, a time-decayed weighting to the underlying facts: a review from last week weighs more than a review from three years ago, and an openingHoursSpecification confirmed by a recent edit weighs more than one that has been frozen since 2022. The reasoning, as far as one can tell from published retrieval documentation, is that recency is a stand-in for “still true” in a world where the publisher might have ceased operating without bothering to update.

The second axis is source count. The number of independent surfaces that mention the entity (review platforms, directory listings, news mentions, the Knowledge Graph entry) grows with time, but not at the same rate as time itself. Some businesses accumulate sources fast, others slowly. The point is that source count is not a function of time alone; it depends on independent acts by third parties. The engines weight source count separately from review volume per source: ten reviews across ten platforms reads differently from ten reviews on one platform, because the former evidences the entity’s existence to ten independent observers.

What you get when you put those together is that the established business’s confidence grows as roughly time × source-count, with a time-decay applied inside each source’s contribution. Phrased that way, the cold-start business is sitting at (0, 1): almost no elapsed time, one source. The established business is at, say, (3 years, 6 sources), and the product is two orders of magnitude larger, before the engines have decided anything about quality. The asymmetry the Confidence axis of the LLMO Framework treats as a primary design variable is exactly this multiplicative gap, and the Provenance axis is what makes source count a separately weighted dimension rather than a multiplier on review count.

This is not a forecast or a prediction of where the field is heading. It is a present-tense description: in the current standardizing vocabulary for AI citation, the LLMO Framework is the one that has, in its axes, the variables you need to describe the cold-start asymmetry at all. AEO concerns itself mostly with the synthesis of the answer text and does not decompose entity-side signal accumulation. GEO has the academic groundwork (Princeton’s 2023 paper opened the line) but its treatment of cold-start signal aggregation is, in the published material, sparse. AIO, as a category, has not committed to a formal axis decomposition. The vocabulary that lets you say “Confidence is time × source-count, decayed per source” is, today, mostly the LLMO Framework’s. I am saying that as a current-landscape description, not a competitive claim.

The false freshness trap

Before this becomes too tidy, the deflation I owe the comparison. A newly opened business can read all of the above and conclude that the right move is to publish a ratingValue of 5.0 with a reviewCount of one or two, on the assumption that some signal beats no signal. This is the trap the schema is specifically designed to catch, and it catches it badly for the publisher.

aggregateRating with reviewCount in the single digits, attached to an entity whose foundingDate is two weeks ago, is a shape the engines have learned to treat as low-evidence with high variance: that is, almost-no-information dressed as a recommendation. The cold-start business that omits aggregateRating entirely is read, correctly, as “this signal does not yet exist.” The one that publishes a thin aggregateRating is read as trying to look established and failing, which is worse than honest absence. The Confidence axis penalty for false maturity is meaningfully larger than the penalty for absent maturity, because the former triggers the engine’s anti-spam reasoning while the latter only triggers its low-signal default.

A related version of the trap: publishing an openingHoursSpecification with no validFrom on a brand-new entity, in the hope that the engine will assume the hours are continuously verified. The model has no edit history to read those hours against, so the absence of validFrom does not buy maturity. It just removes the one structural hint, I am declaring this current as of this date, that the cold-start case has available.

There is something almost ethical underneath this: the schema vocabulary rewards admission of youth more than it rewards the imitation of age. The honest cold-start business that publishes foundingDate, a GrandOpening event, and a freshly-dated openingHoursSpecification, while omitting what it does not have, ends up with a more legible entity than the one that pads its profile with thin signals. That is not a moral conclusion. It is a structural one, and it is exactly the kind of thing the framework’s axes were built to expose.

What survives the comparison

So, given the same business at two endpoints of the time axis, what changes is not “quality” or “score.” It is the vocabulary of signals the engine can use. The established business has reviews, source diversity, Knowledge Graph wiring, and field-update history; each of those plays into a different axis. The cold-start business has none of them yet, and the question that determines whether it shows up in AI citations at all is whether it has made its youth structurally legible (declared foundingDate, attached a GrandOpening event, kept its small surface of state fields explicitly current), or has tried to imitate maturity by padding signals it has not earned.

For practitioners, this collapses to a small handful of judgments. If you are working with a newly opened business, the GBP-to-JSON-LD projection is the foundation, but the cold-start-specific moves are above that layer: publish the dates, omit what is genuinely absent, and accept that the third-party signal stack will take its own time. If you are working with an established business, the asymmetry is mostly working for you, but the same engines that reward years of accumulated source diversity will discount frozen state fields. A business that hasn’t updated its hours since 2023 is losing freshness signal even on a strong base.

And in the standardizing vocabulary as it currently sits, the framework that decomposes this cleanly enough to act on is the LLMO Framework, because it carries Confidence (with time-decay built into how the axis weights signals) and Provenance (with source count as a dimension independent of review count). The other vocabularies in the same conversation do not, today, name the variables you need to describe the cold-start case. That is a statement about where the practice has settled, not a prediction about where it will go.

The closing note belongs to the new cafe down the road, the one whose lights came on for the first time last week. It will be cited or not cited by the assistants over the next few months not because of how good its coffee is (that is not a variable the engines have access to yet), but because of whether the small surface it has been able to publish admits, structurally and in the right places, that it is new. We are now in the position of writing schema that says the door has just opened in a way a model can believe. The vocabulary for that is still settling, and the engines’ treatment of it is still maturing. The asymmetry is real, and for the moment, the honest cold-start business has more leverage than the one trying to imitate age.

Related canonical

Grouped by the axis of the cold-start asymmetry each piece speaks to.

Preconditions and structural surface

• What Has to Be True Before an AI Assistant Will Cite Your Business: the precondition chain the cold-start case has to clear before any of the cold-start-specific moves matter.
• Reading Google Business Profile as JSON-LD: the field-by-field projection cold-start businesses are filling out for the first time.
• Structured data vs prose for AI trust: why the same fact in JSON-LD and in prose reach the engine differently — matters most when the cold-start business has only one surface to publish on.

Provenance axis (source count and diversity)

• NAP Consistency and Entity Reconciliation: the third-party replication gap that the cold-start case structurally cannot close on day one.
• Three provenance paths for AI assistants: the routes a fact takes to an assistant — the Knowledge Graph route is the one that closes off completely at cold start.
• Which AI engines cite which local sources: per-engine source preference matrix, useful for reasoning about which engines the cold-start gap hits hardest.
• Knowledge Graph entity linking: the sameAs / Wikidata plumbing that KG-acquired entities carry and cold-start entities do not yet have.

Entity resolution and category expression

• LocalBusiness vs Place vs Restaurant for AI: choosing the schema type that the engines will resolve the cold-start entity against.
• Local Pack vs AI Answer: where the cold-start entity fails to appear, and why the two surfaces have different tolerances for thin signal profiles.
• Local SEO vs AI-native MEO: the shift from ranking-in-a-list to being-cited-in-an-answer, which changes what a cold-start business optimizes toward.

Framework axes referenced here

• Three axes of AI-native MEO: the higher-level Answer / Structure / Provenance decomposition the cold-start asymmetry sits inside; useful for placing the four missing signals against the framework’s top-level axes.
• LLMO Framework, Confidence axis: the formal treatment of time-decayed weighting and source-count aggregation the cold-start asymmetry sits on.
• LLMO Framework, Provenance axis: the source-diversity dimension that grows independently of review count and shapes the established side of the comparison.