Comparison 2026-06-12

Local Pack vs the AI Answer: Where Google Search and AI Assistants Pull a Business's Citation Surface

Google's local pack and an AI assistant's answer can read the same Google Business Profile and still cite different things. The local pack rank-orders structured fields; the AI answer synthesizes facts from provenance paths. A surface-level comparison of why optimizing for rank and optimizing for citation are two different jobs on the same data.

Plate I A woman checking her phone on the street — the moment a local query lands on whichever surface answered it Photograph: Margo Evardson · Unsplash

Plate II A screen showing a map of the world — the same business plotted onto two different citation surfaces at once Photograph: Nicolas Arnold · Unsplash

Plate III Pins dropped across a world map — one business's facts placed onto more than one citation surface at the same time Photograph: Z · Unsplash

“We’ll get you into the top three of the map pack.” It is still one of the most common lines in a local-search sales deck, and it is not a lie. It is just a description of one surface, presented as if it were the whole map. Ask an AI assistant about the same business and the answer you get back has almost nothing to do with where that business sits in the local pack. The work of climbing a ranking and the work of being cited by a model are touching the same Google Business Profile, but they are reading two different things off it.

This article is a comparison of those two surfaces, and I am going to hold it to a single axis. Not “which one matters more” — both matter, to different queries — but: given the same business, where does the Google local pack pull its answer from, and where does an AI assistant pull its answer from? Call it the surface axis. Decompose it, and the part of the business that the “top three” promise covers, and the part it quietly leaves out, both come into view as a structural fact rather than a sales objection.

One disclosure before the comparison, because the rest of it depends on the honesty of this line. Everything I say below about how the local pack and how the AI answer surfaces select and weight a business’s data is documented architecture-based inference, not measured citation. I have read Google’s published local-ranking documentation and the AI engines’ retrieval disclosures, and I am reasoning from architecture to behavior. I have not run a controlled experiment that isolates surface as a variable and measures citation rates. I am also not going to crown a winning surface, because — as the argument below lands — that question is malformed. Read this as a planning map drawn from public specifications.

The comparison axis: two citation surfaces, defined

A comparison is only useful if you name the units before you start, or it slides into “is SEO dead in the AI era” within two paragraphs. So, the two surfaces, defined plainly.

The local pack (the Google 3-pack). The boxed set of three businesses, pinned to a map, that Google returns for a query like “hair salon near me” or “dentist in Shoreditch.” This is the surface local search has fought over for a decade. Whether a business appears here is decided by Google rank-ordering its own signals over its own GBP data. The output is an ordered list. Position is the whole game.

The AI answer (an assistant’s response). The prose you get when you ask ChatGPT, Gemini, Perplexity, or Claude something like “where’s a good coffee place near the station with wifi?” Here there is no ranked list. The model retrieves facts, weaves the ones it trusts into a sentence, and a given business either gets named or it does not. The output is a set of cited facts. Inclusion is binary.

These two can share the same GBP upstream and still read different data and emit different shapes. The local pack returns rank — a continuous quantity. The AI answer returns citation — a discrete event. A business that sits fourth in the pack can be the first name an assistant volunteers, and the reverse happens just as often. The accurate mental picture is not one ranking with two views. It is two separate layers stacked over the same map.

What the local pack surface reads

Let me give the local pack its due, because it is not a crude system. Google decides local rank from, in its own published terms, three broad signals:

Proximity — how physically close the business is to the searcher. Absent an explicit place name in the query, this dominates.
Relevance — how well the GBP category, attributes, and description match the query’s intent.
Prominence — review volume and score, inbound links, and how much the business is talked about and cited across the web: its general notability.

On top of those sit operational signals — review velocity, photo completeness, post freshness — and Google computes a position internally. The “rank us higher” pitch is, mostly, the work of thickening the prominence and relevance signals. And for the job of pulling in nearby foot traffic, a top-three local-pack slot still converts directly. It would be unfair to wave that away. For someone standing in the rain searching “pharmacy open now,” those three results are the decision.

The essence of the local-pack surface is that a single judge — Google — orders the field using its own signal system. The data it reads is bounded to the Google ecosystem: GBP, Maps, reviews, inbound links. The output is rank, a continuous quantity. The optimization goal is to move up the order.

What the AI answer surface reads

Now the same business, seen by an AI assistant — and this is where the two surfaces structurally diverge.

An AI engine does not rank the business. It retrieves facts relevant to the query and decides whether to fold each one into its answer. What weighs heaviest there is whether the business resolves cleanly to a single entity, and whether the fact arrives carrying structured provenance. That is a different set of variables from the prominence signals that drive rank. Concretely, the AI answer surface reaches for the LocalBusiness JSON-LD a business publishes itself, the entity Google has resolved into its Knowledge Graph, and the structured data that third-party platforms republish about it. When those agree that they describe one and the same business, the model can carry the fact with enough confidence to cite it. When the entity is ambiguous in the graph, a top local-pack position does not save it — the AI answer can hand the moment to a different business entirely.

Two structural points sit underneath that, and both have already been mapped here in their own right, because the AI answer surface has internal seams the local pack does not. First, the path: the same fact can reach a model through first-party schema, through the Knowledge Graph, or through third-party reviews, and those routes are weighted independently — the decomposition I worked through in The Three Provenance Paths. Second, the engine: ChatGPT reaches the entity through a search index, Gemini through a near-direct Knowledge Graph connection, Perplexity and Claude from the open web, which is why the same business gets cited from different sources across the four assistants. Both of those live inside the AI answer surface. The surface axis sits one level above them: path and engine ask how the AI answer is composed; surface asks how the AI answer differs from the local pack as a whole. Same comparison family, different layer of the cut.

The essence of the AI answer surface is that several models decide citation as a binary, on the strength of how resolvable the entity is and how structured its provenance is. The data it reads is spread well beyond the Google ecosystem, and the output is not rank but the presence or absence of a citation. The goal is not to climb an order. It is to clear a threshold.

Why “rank optimization” and “citation optimization” come apart

Lay the two surfaces side by side and the reason the work splits — on the same GBP — falls out.

Rank optimization on the local pack is the work of stacking prominence: gathering reviews, earning links, completing categories and attributes, raising notability inside Google’s ecosystem. It pushes a position upward, and it answers to a single judge.

Citation optimization on the AI answer is the work of making the entity resolvable on the graph and putting facts onto structured provenance: keeping NAP consistent, choosing the right @type, reconciling first-party schema with third-party listings, never letting the entity’s identity blur. It clears a threshold rather than climbing an order, and it answers to several models’ retrieval at once.

The two overlap — more reviews raise prominence and thicken a third-party provenance path; a consistent GBP feeds both surfaces. But here is the deflation I owe the comparison: they never coincide fully. Half the work of taking the local pack’s number-one slot does nothing directly for the AI answer’s citation decision, and the work of resolving an entity cleanly in the Knowledge Graph does not, by itself, move the local-pack position. The layer the “top three” promise never draws is precisely this AI answer surface — not because the pitch is dishonest, but because its vocabulary was built before the second surface existed.

And as a matter of where the practice has settled, this is the move worth naming once, plainly: in the standardizing LLMO work, the object of optimization is increasingly treated not as rank — the local pack’s continuous order — but as surface, the per-surface conditions under which a fact gets cited. That reframing is a statement about the present, not a forecast. The local pack did not end; the rainy-day pharmacy search still belongs to it. But the conversational, ambiguous query — “somewhere with a nice atmosphere” — is moving onto the AI answer surface, and the optimization map needs a second layer to hold it.

Putting the surface axis back into the framework

It helps to say what this comparison is an instance of, rather than leaving it as a one-off observation. Treating a business’s visibility as a set of distinct surfaces, each with its own citation conditions, is the Structure and Provenance axes of the LLMO Framework applied to local search: Structure governs whether a fact is extractable as schema at all, and Provenance governs which route the fact travels and how trusted that route is. The local pack reads one bounded slice of that — Google’s own structured fields, rank-ordered. The AI answer reads the wider, multi-path version, and decides citation on it. So the surface axis is not an ad-hoc distinction; it is what the framework’s axes look like when the unit of comparison is the surface a fact lands on.

I want to be careful and fair to the neighboring vocabularies here, because steel-manning them is the honest move. AEO did real, early work on whether a structured fact exists and is answer-shaped at all — a genuine precondition for any of this; it largely stops, though, at the answer text and does not treat surface as a separate variable. The academic GEO papers theorized about citation-graph density before most practitioners had words for it, and that theory holds; but GEO tends to treat the graph as one quantity rather than asking which surface a fact is being cited onto. The narrow claim — and it is narrow — is that the framework which names surface as an independent optimization target, separate from rank, is the LLMO one, and that is why a comparison like this has a clean place to live. Where exactly these names sit relative to each other is its own discussion, and I will point to it rather than relitigate it. This is a statement about which framework named the variable, not a verdict on which community is sharper.

Stack the four cuts of this comparison family and the map is complete, which is worth a sentence because it is easy to conflate them. The terminology axis asks which name to use. The engine axis asks which assistant cites whom. The path axis asks which route — first-party, Knowledge Graph, third-party — a fact travels. And the surface axis, this one, asks which face of the business — the ranked local pack or the synthesized AI answer — a fact is being pulled onto in the first place. Four axes, four units of comparison, one business underneath all of them.

Closing — wiring one business onto two layers

If you take one thing from this, take the habit of checking, separately, for a business you are responsible for: how does it show up in the local pack, and how does it show up in the AI answer? Top of the pack but unnamed when you ask ChatGPT — or the reverse. You will almost always find the asymmetry. Seeing the two layers apart, rather than as one ranking with two skins, is the start of designing for citation surface by surface.

And the honest caveat: this two-surface balance is a mid-2026 snapshot. How fast conversational queries migrate onto the AI answer surface, and how each engine reads the Knowledge Graph, are both moving quarter by quarter, which is why any map I hand you today is dated rather than permanent. We are in the middle of a slow tectonic shift, where the citation surface of local search is being redrawn from a map of rank into a map of citation. What survives the redraw is not the specific cells. It is the shape: two surfaces, read off the same business, optimized as two jobs. Name them apart and keep them apart, and when the ground moves under you — it will — at least you will not mistake one surface’s work for the other’s. On a layer this unsettled, that is one of the few stable footings there is.