AI search visibility (GEO)

AI search surfaces local businesses two ways: as entities recommended in the answer, and as pages the model cites or quotes from. Most local-SEO writing focuses on the first and ignores the second. For a local business, both matter, and the right balance shifts depending on the query type and the surface the user is on. This guide covers both layers, surface by surface.

Two ways AI search surfaces a local business

Almost every conversation about AI visibility collapses into one of two narrow framings: "schema and Knowledge Graph" (entity-only) or "great content" (page-only). For local SEO neither is sufficient on its own. Here is the actual picture.

For a local business optimising for AI visibility, all three pathways exist and all three deserve effort. They are not interchangeable. A perfectly-built GBP with rich Knowledge Graph presence still loses for informational queries where the AI is pulling from a how-to article. A beautifully-written guide on your site loses for "best plumber in Camden" if your GBP and reviews are weak. Earned media in TimeOut or a recognised trade publication can win both at once.

Local SEO has always been the messy middle between entity-shaped queries ("ABC Plumbing") and page-shaped queries ("how often should I service a boiler"). AI search inherits that mess. Optimising only the entity side or only the page side leaves half the surfaces uncovered.

Different queries trigger different surfaces

"Will I appear in AI search?" is the wrong question. The right question is "for which queries, on which surfaces, and as which type of result?". The same query can produce a Map Pack on Google, an AI Overview above the SERP, an AI Mode conversation, a Knowledge Panel on a brand search, or a synthesised paragraph in ChatGPT or Perplexity. Each surface ranks different things differently.

The four moving parts of an AI answer

Every AI assistant answer (and increasingly every AI Overview / AI Mode result) goes through four stages. Different products implement each stage differently, but the structure is the same.

1. 1

   Query interpretation

   The model parses the natural-language question into structured intent. "Recommend a Saturday dentist near Brighton who takes nervous patients" gets decomposed into entity type (dentist), location (Brighton), temporal constraint (Saturday), and an attribute filter (nervous patient care). Conversational queries in AI Mode often get fanned out into multiple sub-queries here.

2. 2

   Retrieval (entity + page indexes in parallel)

   The system queries multiple sources at once. A web search index (Bing, Google, Brave, or proprietary) returns candidate pages. A structured data source (Knowledge Graph, Maps API, Wikidata) returns candidate entities. Embedding-based vector retrieval may run against pre-indexed content. Hundreds to thousands of candidates feed the next stage.

3. 3

   Reranking, entity resolution, and source selection

   Candidates are reranked by relevance, source authority, and recency. The model resolves multiple references (a Knowledge Graph node, a GBP record, a Wikidata entry, mentions on third-party sites) to a single canonical entity. Pages are scored for coverage of the answer, recency, and trust. Ambiguous business names or fragmented identities lose visibility here.

4. 4

   Synthesis with citation selection

   The model writes the answer using the highest-scoring retrieved sources as grounding. Citation selection is partly about coverage (does this source carry the bit of the answer the model needs?) and partly about authority (is this a source the model trusts to cite?). The user reads the answer; whether they ever click through depends on the surface and the user.

Layer 1: The entity layer

An entity in the search-engine sense is a disambiguated, structured identity for a thing in the world. Google's Knowledge Graph has held billions of entity nodes since 2012. Wikidata is a public knowledge base with well over 100 million entity records. Each entity has a stable identifier (a Knowledge Graph machine-readable ID, a Wikidata Q-number), a type (LocalBusiness, Person, Organization, Product), and properties.

Your business is, or wants to be, one of those entities. The signals that let AI retrieval systems identify "you" come from multiple sources, which need to converge:

When all four sources reference the same entity, retrieval systems converge on a single canonical record with high confidence. When they disagree (different names, addresses, missing nodes), the system either picks one and gets it partially right, or fails to surface you at all. Entity convergence is the entity-layer goal.

Layer 2: The page layer

A page in this context is one URL on your domain (or someone else's) that an AI might quote, summarise, or link to in a synthesised answer. Pages get cited based on different signals than entities get recommended. Treating "AI visibility" as purely an entity problem leaves all the informational, comparison, and how-to queries on the table for whoever puts in the page work.

What makes a page citation-worthy

Page-level schema worth implementing

Beyond the LocalBusiness and Service schema covered in the entity layer, page-level schema is what tells AI extraction systems what a specific URL is for. The high-value types:

• Article (or BlogPosting, NewsArticle): for editorial pages, with author referencing a Person node, datePublished, dateModified, publisher referencing your Organization.
• HowTo: for step-by-step instructional pages, with step entries, tool and supply if applicable. Maps directly to AI Overviews for how-to queries.
• FAQPage: for FAQ sections or pages, with Question and Answer entities. Maps directly to chunk-level extraction.
• BreadcrumbList: navigational context. Helps AI systems understand where the page sits in your site architecture.
• Person: for any named author, with jobTitle, worksFor, hasCredential, and sameAs linking to LinkedIn, regulator listings, academic profiles. Important for E-E-A-T in YMYL sectors.
• Service on service pages, referencing the parent LocalBusiness via provider, with areaServed and optionally offers.

Chunk-friendly content structure

Retrieval systems break pages into chunks, embed each chunk into a vector, and retrieve the most relevant chunks per query. Chunks that read as standalone units win. Chunks that depend on context several scrolls earlier lose. The structural patterns that work:

• Clear H2/H3 heading hierarchy with descriptive headings (not 'Why us' but 'Why choose a Brighton-based dentist for nervous patients')
• Self-contained paragraphs that quote cleanly without context
• Definition-style opening lines for each section (the question being answered, then the answer)
• FAQ blocks with FAQPage schema, since these map directly to chunk-level extraction
• Tables for comparison data (machine-readable, easily extracted)
• Lists with parallel structure (each item answers the same implicit question)
• One topic per page where possible; long mixed-topic pages chunk badly
• Original data, statistics, or primary research (citation-worthy chunks)
• Internal links to related pages on your domain (signals topical cluster membership)

Layer 3: Earned media (third-party citations)

For competitive "best of" queries and local recommendation queries, the page the AI cites is often not yours. It is a third-party authoritative page listing several businesses, your business included. Local press, sector publications, professional body directories, regulator listings, "best of" round-ups: each of these can be the citation source for an AI answer that names your business.

Practical patterns:

How each major assistant retrieves

Different assistants are wired into different stacks. Knowing which stack each uses tells you which signals you need to invest in for that surface.

Two further surfaces worth knowing about. Microsoft Copilot (which absorbed Bing Chat in 2024) is also Bing-grounded, so the same Bing Webmaster Tools and Bing Places work that helps ChatGPT search visibility helps Copilot. Apple Intelligence in iOS surfaces businesses through Apple's own ecosystem (Apple Maps, Apple Business Connect, Spotlight) and routes some queries to ChatGPT when the user opts in, which means Apple Business Connect completeness is increasingly worth the same effort as GBP for iOS-heavy audiences.

What this means for local-SEO investment

AI Mode and query fan-out

Google's AI Mode (a more conversational AI search experience launched in 2025 and now widely available) is structurally different from AI Overviews. The key mechanic is query fan-out: a single user question gets decomposed into multiple sub-queries, each retrieved against its own most relevant sources, with the synthesis layer assembling a multi-source answer.

A query like "I run a small Italian restaurant in Hackney and I want more weekday lunch covers, what should I do?" might fan out into sub-queries on local marketing tactics, lunch-trade benchmarks for hospitality, restaurant SEO basics, common mistakes operators make, and what specific tools or services exist. Each sub-query has different best sources. The user sees one synthesised answer that draws from all of them.

Surface-by-surface signal mix

For a quick reference, here is which signals dominate which surface for local-intent queries.

Crawler access: the audit nobody runs

AI crawlers obey robots.txt. Many sites accidentally block them via wildcard rules, agency-installed defaults, or templates inherited from years ago when AI crawlers did not exist. Audit your robots.txt and verify each of these is intentionally allowed (or intentionally blocked):

An emerging proposal worth knowing about is llms.txt, a Markdown file at your domain root describing your site's content for LLM consumers. Adoption through 2025 has been gradual: a growing share of publishers have published one, and some AI products have added experimental support, but it is not yet a universally honoured standard the way robots.txt is. Adding one is cheap and the signal-cost is zero if it never gets used.

Wikidata: the open Knowledge Graph

Wikidata is the structured-data sibling of Wikipedia. It powers parts of Google's Knowledge Graph, feeds many AI systems' training data, and has a far lower notability bar than Wikipedia. Most established local businesses, especially those with a few years of trading history and any media coverage, qualify for an entry.

Practical approach:

1. Search Wikidata for your business. If an entry exists, claim it (or have a Wikidata-active editor add the missing properties on your behalf).
2. If no entry exists and you meet notability requirements (some independent sources covering the business), create one. Editing your own entry is technically allowed but discouraged. Better to provide editors with sourced facts and let them create the entry.
3. Populate key properties: P31 instance-of (e.g. Q4830453 for a generic business, or a more specific subtype where appropriate), P17 country, P159 headquarters location, P856 official website, and any others relevant to your sector.
4. Add the resulting Q-number URL (e.g. https://www.wikidata.org/wiki/Q123456789) to your domain's sameAs array in your Organization schema.

The 9 GEO levers, in priority order

For a local business covering both layers, this is the order of impact most consistently. The top items multiply the effect of items below.

YMYL: where AI plays it safer

Your Money or Your Life queries (medical, legal, financial, safety) are handled more conservatively by every major assistant. Models default to authoritative sources, surface caveats more readily, and decline to recommend specific providers more often than in lower-risk sectors.

Sector heuristics

Patterns we have observed across sectors for what AI assistants weight most heavily, with the entity-vs-page balance broken out.

Measuring AI visibility

AI search has no single "rank" to track. You measure visibility through structured query testing across surfaces and assistants, looking at both layers: are you the entity recommended, and are your pages cited.

The query test set

Cover the realistic phrasing space across surfaces:

• Brand: '[your business name]' (Knowledge Panel)
• Local recommendation: 'best [category] in [city]' and '[category] near me [city]'
• Local recommendation: 'recommend a [category] in [city]' and 'where should I go for [service] near [location]?'
• Specific need: 'I need a [category] who can [specific need]'
• Sub-area: 'a good [category] in [neighbourhood]'
• Comparison: 'compare [category A] to [category B]'
• Filter: 'cheap or affordable [category] near [area]', 'highly-rated [category] in [city]'
• Attribute: '[category] with [specific attribute] in [city]'
• Time-sensitive: 'open now [category] near me [city]'
• Conversational AI Mode: 'I have a [specific situation] and need help, what should I do?'
• Informational: 'what does a [category] cost in [region]?', 'how often should I [task]?'
• How-to: 'how do I [common task]?'

Metrics to track

Running this manually across many queries weekly is impractical. Our AI Visibility Tracker automates this so you can scan dozens of queries weekly across all major assistants without hand-running each one.

Common mistakes

The 90-day GEO programme

A practical phased plan for a local business. Each phase covers both the entity layer and the page layer.

1. 1

   Days 1 to 30: foundations on both layers

   Audit robots.txt for all major AI crawlers. Implement comprehensive schema.org markup: LocalBusiness on the homepage, Service on each service page, FAQPage on FAQ sections, Article and Person on any editorial content. Validate every JSON-LD block. Confirm Bing Places and Apple Business Connect both match your GBP. Audit your existing page library for the obvious gaps in informational and how-to queries.

2. 2

   Days 31 to 60: entity convergence and page depth

   Create or claim your Wikidata Q-number. Add sameAs linking your domain to Wikidata, GBP, social profiles, Companies House, regulator pages where relevant. Publish or refresh 3 to 5 citation-worthy pages: original data, expert commentary, primary research, or genuinely useful how-to content. Add Person schema for the authors. Aim for 1 to 2 earned mentions in authoritative third-party sources (recognised press, professional bodies, sector publications).

3. 3

   Days 61 to 90: measure across both layers and iterate

   Run your structured query set across ChatGPT, Gemini, AI Overviews, AI Mode, Perplexity, Claude, and Microsoft Copilot. For each: note entity inclusion rate, page citation rate, earned-media citation rate, sentiment, and competitive set. Identify gaps (where competitors appear and you do not, on which surface, for which query type) and target the underlying signal. Re-test monthly.

Reference numbers

Where this is going

AI search is not replacing Google overnight, but it is eating top-of-funnel click-through and expanding the surfaces a local business has to be visible on. The businesses that win in three years are the ones whose entity identity converges cleanly across the Knowledge Graph, Wikidata, GBP, and their domain; whose page library covers the real questions in their sector with credible original content; and who maintain solid traditional SEO underneath. The ones that do not get cited slowly disappear, even if their organic rankings hold, because the user never sees the SERP. They got the answer.

Where to go next