AI search for local business: AI Overviews, AI Mode and ChatGPT

When someone asks ChatGPT for a plumber, or Google puts an AI answer above the map for "emergency electrician", a handful of businesses get named and the rest do not. This guide is about being one of the named ones. It starts with what each of these systems is actually documented to run on, because they are not one thing and the differences change what is worth doing, then works through the levers that make a business easy for all of them to recommend.

The systems doing the naming

"AI search" is a useful umbrella, but underneath it sit products built quite differently, by different companies, on different indexes. For a local business, seven are worth knowing about: Google AI Overviews, the AI answer above ordinary results; Google AI Mode, the conversational tab inside Search; Google Ask Maps, the Gemini answer built into Google Maps; ChatGPT search; the Gemini app; Microsoft Copilot; and the citing assistants, Claude and Perplexity.

They fall into three families, and the family matters more than the brand. AI Overviews and AI Mode are search surfaces sitting on top of a web index you already work on. Ask Maps is a map surface, grounded almost entirely in business records and reviews rather than web pages. The rest are assistants that call out to a search provider and summarise what comes back. Where a product sits tells you which of your assets it can even see.

How each system finds a local business

What follows is what each company documents or has publicly stated, kept separate from what the industry infers. The gap between the two is wide, and most advice in this area does not mark it.

Google: AI Overviews and AI Mode

Google is the only one of these publishing guidance aimed at site owners, and much of it is about what does not matter. To be eligible, a page needs to be indexed and eligible to be shown with a snippet. There are no additional technical requirements, no new machine-readable files, AI text files or markup to create, and no special schema.org structured data to add.¹

On mechanism, both surfaces use a query fan-out technique, issuing multiple related searches across subtopics and data sources and bringing the results together into one response.¹² The launch announcement described AI Mode as running on a custom version of Gemini, drawing on real-time sources including the Knowledge Graph, and working by making a plan, running searches, then adjusting that plan based on what it finds.²

For a local business the consequence is specific: the sentence your customer types is not the query that gets run. A conversational question is decomposed into narrower searches, and it is those you need to be retrievable for.

The controls Google offers are the ordinary ones: nosnippet, data-nosnippet, max-snippet and noindex, plus robots.txt for Googlebot.¹ Worth knowing that these are not AI-specific dials. Using them to keep your text out of an AI answer also removes it from ordinary search snippets, which is almost never a trade a local business wants.

For the local half of the answer, Google's published local ranking factors still apply: relevance, distance and prominence.⁴ That document predates AI Mode and Google has not restated it for the AI surfaces, so read it as the best public account of how Google ranks local businesses, not as a description of what the AI layer then does with them.

Google: Ask Maps

Announced in March 2026, Ask Maps puts a conversational Gemini answer inside Google Maps, handling questions a map could not previously answer. The announcement describes it as bringing together the map with its most capable Gemini models, grounded in information from over 300 million places including reviews from a community of more than 500 million contributors, and personalised using places you have searched for or saved.⁵ It began rolling out in the US and India on Android and iOS.

This is the surface where the lever is clearest. There is no page to write and no markup to add, because the grounding data is business records and reviews. If your Google Business Profile is thin, stale or miscategorised, there is very little here to work with. It is also the clearest illustration of why a screenshot is not a measurement, since results are personalised to the individual.

ChatGPT search

OpenAI describes ChatGPT search as using third-party search providers alongside content provided directly by partners, typically rewriting the user's question into one or more targeted queries that it sends to those providers. It says results are determined by automated systems weighing factors including intent, relevance and recency, and that the ordering of the sources shown alongside an answer is influenced by the provider's own ranking systems.⁶

That last point is the one to hold on to. For a meaningful part of what ChatGPT shows, you are being ranked by a conventional search engine first and summarised second. Ordinary search visibility is not separate from this work, it is an input to it.

The partner content matters locally too. In July 2026 Yelp licensed its reviews, ratings, photos and business information to OpenAI, reported as around 330 million cumulative reviews and more than 8 million listings.⁷ In markets where Yelp has real coverage, an accurate and well-reviewed Yelp listing became a direct input to ChatGPT's local answers rather than a nice-to-have directory entry.

What is actually known about the local stack. ChatGPT shows a local business panel: a map with business cards, website buttons and directions. The map layer is on the record, because Mapbox announced that its maps power ChatGPT search.⁸ The listings layer is not. OpenAI has never published where the business data behind that panel comes from, and the provider market-share figures circulating in this space are inference presented as fact. The one local data source actually on the record is the Yelp licence,⁷ which is also the one you can do something about.

The others, briefly

These carry less local volume today, but the same two questions apply: what does it retrieve from, and what can it see about you?

What moves it

No company publishes how its AI layer chooses which businesses to name, and anyone claiming otherwise is inferring. What can be said with confidence is which assets these systems demonstrably read, and there are three. They are listed here roughly in order of leverage, and each strengthens the others.

Lever 1: A clear business record, positioned to what people ask

These systems treat your business as a thing in the world, not a string of words on a page. That thing has records, usually several, carrying your name, location, category and details. The cleaner and more consistent those records, the less ambiguity there is to resolve before you can be named.

What ties these together is consistency and consensus. Consistency is every record describing your business the same way. Consensus is the wider web agreeing about what you are good at, so your reviews, listings and mentions describe you in the same terms your own profile does. When records conflict, or what others say does not match what you claim, there is less to be confident about, and a competitor with a cleaner record is the safer thing to name.

Having the records is half of it. The other half is positioning: making sure what you list matches how customers ask. A plumber whose primary category is "Plumber", with services like "Emergency plumber" and "Boiler repair", lines up with real questions. One filed vaguely under "Contractor" does not. A dentist who spells out "nervous patients" and "same-day emergencies" can be matched to exactly those. Say plainly what you do, in the words people use to ask for it, across your categories, services, description and service pages.

Wikidata, if you qualify

Wikidata is the open, structured sibling of Wikipedia. Google's Knowledge Graph draws on public sources such as Wikipedia alongside licensed data,¹³ and Wikidata is widely used as structured input across the industry. Its bar is far lower than Wikipedia's, but not zero:¹⁴ an entry must be describable using serious, publicly available references rather than self-published claims, and a company-register entry alone is thin evidence. Many established local businesses clear that bar; a brand-new one with no outside coverage will not. If you qualify, create or claim the entry, fill in the core properties (P31 instance-of, P17 country, P159 headquarters location, P856 official website), and add the Q-number to your sameAs list.

Lever 2: Your reputation, what people say about you

Reviews are the largest single part of your reputation, and they do two jobs at once. Google names review count and rating as part of prominence,⁴ which affects whether you are a candidate at all. Separately, the review text is prose about your business sitting on a page these systems retrieve, which is where the specific phrases in an answer tend to come from.

The operational detail on the biggest piece is in our guide to getting more Google reviews.

Lever 3: Earned media, getting named on pages that get retrieved

For competitive "best of" questions, the page an assistant quotes is usually not yours. It is a round-up, a directory, a local news piece, or a trade-body register. Getting onto those pages puts your name in the text that gets retrieved.

A point traditional SEO understates: these mentions can count even when they do not link to you. An answer is assembled from text, so what matters is whether your name appears next to the thing being asked about, on a page that gets retrieved, not only whether a link passes value.

Your own content: a supporting role

This is where most "AI content" advice over-promises. There are no additional requirements to appear in AI Overviews or AI Mode, no new machine-readable files or markup to create, and no special schema.org structured data to add.¹ That is the documented bar in full, and nothing sold as an AI file or markup layer sits on top of it.

You will still be shown a correlation, and the correlation is real: pages carrying schema do get cited more often. It means something other than what it is sold as. Schema tends to live on sites that are better built and better maintained, and those are the same sites that publish stronger content and earn more mentions. The markup is a marker of that care rather than the cause of the citation, which is why controlled tests that add schema to existing pages keep failing to find a lift.

Where your own content does earn its place:

• Service pages that clearly say what you do, who for, and where. These support both your record and the commercial questions directly, and they are worth more than any blog post.
• Practical FAQ content answering real pre-booking questions ("do you charge a call-out fee", "how soon can you come"). It helps customers and gives a clean, quotable answer tied to you.
• Genuine depth on your subject, which is what gets pulled into the research-stage questions where AI answers are most common. A long game, not the core of the plan.

Write it for a human.¹⁶ How you structure it so a machine can lift a clean answer out of it is a technical question, and it is covered properly in the next section.

The technical layer: being readable in the first place

Everything above assumes these systems can reach your content and parse it. Often they cannot, and the failure is silent: nothing breaks, no warning appears in Search Console, you simply are not in the answer. Three things decide it, and only one of them is robots.txt.

Rendering: whether they can see your content at all

This is the single biggest technical gap, and it splits along exactly the lines this guide has been drawing. Vercel's analysis with MERJ, covering over 500 million GPTBot fetches, found that none of the major AI crawlers render JavaScript: OpenAI's three agents, ClaudeBot, PerplexityBot, and the Meta and ByteDance agents. They download JavaScript files (ChatGPT on 11.50% of requests, Claude on 23.84%) without executing them.¹⁷ Independent testing through 2026 has found the same.

The exception is the one that matters most. Google's AI surfaces inherit Googlebot's rendering, and Copilot inherits Bing's. So a client-rendered page can be perfectly visible in AI Overviews and AI Mode and completely invisible to ChatGPT, Claude and Perplexity. If you test one assistant and conclude you are fine, this is the gap you will miss.

Access: robots.txt, and the layer above it

Indexing crawlers generally obey robots.txt, while user-triggered fetchers often do not. OpenAI notes robots.txt rules may not apply to ChatGPT-User,⁹ and Perplexity states that because a user requested the fetch, Perplexity-User generally ignores it.¹⁸ Plenty of sites block the indexing crawlers by accident¹⁹, through wildcard rules, agency defaults, or templates written before these agents existed. Check that each of these is intentionally allowed, or intentionally blocked:

You may also hear about llms.txt. None of the companies above document fetching it when answering a query, or list it as a requirement. It is cheap and harmless to add, but treat it as speculative rather than as a visibility tactic.

How your content gets broken up

Retrieval does not work on whole pages. It works on passages, and what gets quoted back is a fragment, not your article. This is not new and it is not AI-specific: Google documents passage ranking as "an AI system we use to identify individual sections or 'passages' of a web page to better understand how relevant a page is to a search".²² Retrieval-based assistants work the same way, pulling the section that answers the question rather than the document that contains it.

The practical consequence is about self-containment, and it is easy to get wrong without noticing:

• Each section should answer one question and survive being lifted alone. A passage that opens "This also applies to our emergency callouts" is useless extracted, because "this" is gone. Name the thing again.
• Put the answer in the first sentence under the heading, not after three sentences of preamble. A retrieved chunk is often short.
• Facts belong in text. A price that only exists inside an image, a service area only drawn on a map, or hours only rendered by a widget are not retrievable. Say them in words as well.
• Descriptive headings beat clever ones, because the heading is part of how a passage gets matched to a question.
• Do not hide the substance behind interaction. Content inside tabs or accordions is usually fine if it is in the raw HTML, and invisible if the tab fetches it on click.

The entity plumbing

The technical side of your business record sits under the first lever above: consistent NAP, valid LocalBusiness and Organization markup, and a sameAs list that ties your records together. Worth repeating one boundary here, because it is where technical effort most often goes to waste. That markup helps search engines resolve who you are, which is upstream of the answer. It is not a citation mechanism, and no structured data is needed for Google's AI features.¹ Get it correct, then stop. If you want the full treatment of the entity layer, our entity optimisation guide goes considerably deeper than this one needs to.

Personalisation: why there is no single answer to check

Local search has always been personalised, because distance is one of the three factors Google names.⁴ Two people asking the same question from opposite sides of a city have always got different results. What has changed is that personalisation now reaches past location into the individual, and it is documented rather than inferred.

When you have more than one location

Everything above applies per branch, and the odds get considerably longer. The reason is arithmetic rather than anything exotic. A local pack shows three businesses with a page of results beneath it, and an ordinary results page shows ten. An answer names one or two and then stops. Every surface in this guide compresses a shortlist into a sentence, and compression is brutal at scale: the same estate that has most of its branches visible somewhere on a results page can have almost none of them named.

The independent evidence here is thinner than anyone would like, because the organisations with the data to measure it mostly sell software to fix it. Treat any headline recommendation rate you are quoted with that in mind, including the ones in your own reporting, and measure your own estate rather than adopting someone else's benchmark.

Three things change once you are managing more than one:

The arithmetic is also why this stops being a manual job quickly. A covered set of thirty questions, asked three times each across four systems, is 360 answers a month for one location. At two hundred locations it is 72,000. Our multi-location local SEO guide covers the record and reporting structure underneath all of this.

Measuring it: what the answers tell you

There is no single rank to check. The same question asked twice can come back with a different shortlist, each system answers from different sources, and Ask Maps personalises its results to the individual.⁵ So measurement works more like reading a market than checking a position.

Build a covered set, not a single query

You cannot see what people type into an assistant, and one question on its own tells you very little. What you can do is build a set of questions covering the moments that lead to an enquiry, then read how the answers treat you across the whole set. Start where someone is ready to choose, and build outwards.

Read across the set and you can see where you are shortlisted, where you are mentioned in passing, where you are missing, and who is named in your place at each stage. A practice might be named for "best dentist in Leeds" and absent every time a nervous patient asks, while a rival that talks about sedation is recommended instead. That is not a ranking problem, it is a positioning gap, and it points directly at what to say on your profile and site. Our post on which AI prompts a local business should track builds a set sector by sector.

What a single answer contains

Every answer holds far more than a yes or a no. Here is what one of them contains once you break it apart.

What the answers add up to

One answer is an anecdote. Across a covered set, asked repeatedly, the same fields turn into something you can act on:

Doing this by hand across a covered set, several times over, is not practical, which is why every answer we collect is stored as the structured record above rather than as a screenshot: who was named, how strongly, which rivals appeared, what the answer said about you, and which sources it leaned on. Our ChatGPT Visibility and Gemini Visibility tools ask your questions on a schedule and build that record, and geo-grid scans record whether Google's AI Overviews cite you and whether AI Mode names you, from points across your area.

When the answer is wrong about you

Sooner or later one of these will state your old opening hours, put you in the wrong category, attach a competitor's phone number to your name, or say you closed. The instinct is to look for somewhere to correct it. There isn't one. You cannot edit a model, there is no record to log into, and a thumbs-down does not change what it tells the next person who asks.

What you can change is what it reads. Before you start, work out which of three problems you actually have, because they need different work:

1. 1

   Fix what you control first

   Your Google Business Profile, your Bing and Apple records, your own site. State the correct fact plainly in visible text, not only in markup, and put it somewhere prominent rather than on a buried page. If the wrong fact is still live anywhere you own, nothing downstream will settle.

2. 2

   Find the third party still carrying it

   This is usually where the answer is actually coming from: one stale directory entry, an old press mention, a data aggregator that never got the update. Tracking which sources an answer leans on is exactly what makes this findable, rather than guessing. Get that page corrected, because until it is, you are arguing with it.

3. 3

   Use the in-product feedback, without relying on it

   Google's AI answers carry a feedback control, and assistants let you flag a response. Use them: it costs a minute and it is the only direct channel that exists. Do not treat it as the fix, because none of these companies commit to acting on an individual report, and the correction that lasts is the one to the underlying source.

4. 4

   Re-test on a cadence, not on the day

   Indexes refresh on their own schedules. Expect weeks rather than days, in line with the lag on any other change, and check monthly. Re-asking the same question an hour later tells you nothing except that answers vary.

Common mistakes

Reference numbers

Where to start, in order

This guide is long because the subject genuinely is. The work is not, and most of it is unglamorous. If you do nothing else, do these five, in this order, because each one is wasted effort without the one above it.

1. 1

   Confirm they can read you

   View source on your key pages and search the raw HTML for your phone number, hours, service areas and prices. Then check your CDN's AI crawler setting, not just robots.txt. Everything below is wasted if this fails, and it fails silently. Half an hour.

2. 2

   Finish the business record

   Complete every field on your Google Business Profile, get the primary category right, and write your services in the words customers use. Then claim Bing Places and Apple Business Connect, which most of your competitors have not. A day, once.

3. 3

   Restart the reviews

   Not a campaign, a habit: ask every customer, soon after the job. Recency is a stated factor and a steady trickle beats a burst. This is the slowest lever to move and the one that compounds, so start it before you need it.

4. 4

   Measure ten questions before you change anything else

   Pick ten questions a customer would really ask, run each more than once, and write down who gets named and what the answers lean on. Without a baseline you cannot tell a real improvement from the run-to-run variation, and you will credit whatever you happened to do last.

5. 5

   Then go after the gap the measurement shows

   Only now is earned media worth your time, because the baseline tells you which pages name your rivals and not you. That list is the pitch list. Work it in that order rather than guessing which round-up matters.

Where this is going

These systems are not replacing Google overnight, and as the prevalence data shows, they are not yet answering most ready-to-buy local queries. What is changing is the research that happens before the choice, and what an answer does to the click. Pew found that when an AI summary appeared, people clicked through to a website on 8% of visits against 15% when it did not, and clicked a link inside the summary on just 1%.²⁵

That resolves a tension worth naming, because it is the one people get stuck on. AI referral traffic looks small in analytics, and it is small, and that is not the point. If a growing share of decisions is being made inside an answer that nobody clicks out of, then being named in it is increasingly the whole transaction rather than the start of one. The visit you used to get is the thing that went missing, not the customer.

The businesses that do well are the ones with a clear record, a genuine reputation, and a presence on the pages these systems retrieve. That is the same list that has always worked, now being read by more things at once.

Where to go next