What AI search means for a local business in 2026
“AI search” isn’t one website. It’s several places where a question gets answered in sentences instead of links:
- Google AI Overviews. The summary above Google’s normal results.
- Google AI Mode. A version of Google you talk to.
- ChatGPT. The assistant most people mean when they say AI. It searches the web when a question needs current information.
- Gemini. Google’s assistant, which can also draw on YouTube.
- Perplexity. A search product that shows its sources for every answer.
Underneath, they generally work the same way. Each gathers information from public sources, condenses it into a short answer, and names a small number of businesses. Being one of those names is what AI search visibility means.
The work goes by several names that all describe the same thing: AEO (answer engine optimization), GEO (generative engine optimization), LLM optimization, AI Overviews optimization. Local SEO is the older term for most of what sits underneath, because being findable in one geographic area has changed less than the vocabulary for it.
Why local businesses are better positioned than they think
It feels like AI search should favor whichever business is biggest. For local recommendations it usually doesn’t.
A wrong answer about a local business is easy for a customer to catch: incorrect hours, a disconnected number, a service the business doesn’t offer. So these systems lean toward information they can verify against more than one source. A national brand has volume. A plumber in one metro area has small, checkable facts nobody else can claim, and most of this article is about making those facts easy to find.
Some of what follows takes an afternoon. Some of it is ongoing local SEO work that takes months. Most of it can be done without hiring anyone.
How do AI models find information about a local business?
An AI assistant doesn’t keep a private directory of local businesses. It knows about a business in two ways.
The first is training. The system read an enormous amount of text at some point in the past and then stopped. Whatever it picked up about a business is frozen at that moment, and for most local businesses that’s very little, because a contractor in one suburb barely appears in the text these systems learn from.
The second is looking things up while answering. This is what AI Overviews, ChatGPT search, and Perplexity do. They fetch real pages during the conversation and build the answer out of what they find.
That split explains the odd things owners notice, like an assistant describing a business that closed two years ago. It’s also the good news: because these systems look things up fresh, fixing what they find changes the answer in weeks rather than years.
The six places a model reads
When a system looks something up, it pulls from six kinds of source at once:
- The Google Business Profile. The most structured of the six.
- The website. Where a business says what it does in its own words.
- Structured data. The website’s facts restated as code.
- Directory listings. Name, address, and phone number on sites the business doesn’t control.
- Reviews. What customers say, plus how the business responded.
- Indexed content that mentions the business. Its own posts, and other people’s articles and roundups.
When these disagree, there’s no way to tell which is right. The five facts that drift most often are business name, address, phone number, hours, and the list of services, and the places they most often disagree are the profile, the website footer, and the top three directory listings. Most businesses have at least one mismatch, and fixing it costs nothing.
Why Bing matters
Bing holds a small share of ordinary search, but ChatGPT’s live web results lean heavily on Bing’s index. A page Bing hasn’t indexed is a page ChatGPT is unlikely to find. OpenAI has built retrieval of its own through 2026, so Bing isn’t the only gate, but being indexed there is cheap and skipping it closes a door.
Bing Places is Bing’s equivalent of a Google Business Profile, and Bing Webmaster Tools is a free service that shows whether the site’s pages are indexed.
Being read versus being named
A system can read a page and still name someone else.
Getting read is a plumbing problem: the pages have to be reachable by the software that collects them. Getting named is a writing problem: the information has to be specific enough that an answer can repeat it without hedging. Google’s guidance on AI features confirms the first half: a page has to be indexed and eligible for a snippet to appear in AI Overviews or AI Mode, and nothing more is required.
The next few sections cover getting read. The content sections further down cover getting named.
What does AI pull from a Google Business Profile?
A Google Business Profile is the listing beside Google’s map results, the one with the hours, photos, and reviews. It used to be called Google My Business. If nobody at the business has ever logged in, it may still exist unclaimed, built by Google from other sources.
It’s also the most structured source a model has, because a profile is fill-in-the-blank. Hours go in an hours box, services in a services list, and nothing has to be worked out from a sentence. That makes it the fastest place to fix a visibility problem and the easiest place to leave one sitting.
Categories and services
Every profile has one primary category, chosen from Google’s list, and it does most of the work. It’s how a system narrows thousands of local businesses to the handful that could plausibly answer a question about tankless water heater repair in one suburb. Secondary categories add range. The services list is where being specific pays off: an itemized list, each service with its own short description, gives a system something to match against. One blanket entry gives it almost nothing.
Attributes and service area
Attributes answer the filtering questions people ask. Emergency service, financing, languages spoken, licensed and insured.
A business that travels to the customer should be set up with a service area rather than a storefront address. Listed with a storefront address instead, it lands in the wrong geographic pool from the start.
The Q&A section
People ask these assistants full questions, in the words they’d use out loud. The Q&A section on a profile stores information in exactly that form: a real question with a direct answer under it. A business can post its own questions and answer them, so the eight or ten that come up on every sales call can sit there in the format an assistant is already looking for. It’s usually empty.
Posts and photos are source material
Posts signal that a business is still operating. Photos carry information in their file names and descriptions, which is text a system can read even when the image isn’t the point. There’s a separate guide to naming and geotagging photos for local SEO, because it’s fiddlier than it sounds.
A complete profile still won’t carry a business on its own. If it says one thing and the website says another, it stops being an asset and becomes half of a contradiction.
The website is the source of truth
A Google Business Profile can say a business offers drain cleaning, but that’s the business describing itself in a box Google gave it. The website is where an assistant checks whether that holds up. When the website can’t confirm it, the profile is a claim with nothing behind it.
One page per service, one page per location
A single services page listing nine offerings gives a model one thin signal about nine things. Nine pages give it nine specific ones.
The same applies to geography. A business working across four suburbs needs a page for each, with real differences between them. Four pages of the same copy with the city name swapped is one source repeated, and search engines and AI systems have discounted that pattern for years.
Plain-language descriptions
Service pages written as sales copy give a model nothing to extract. Pages that explain what the work involves, how long it takes, what it typically costs, and what happens first give it a paragraph it can quote.
The test is whether a page, if it were the only thing an assistant could find, could answer a customer’s question. Most service pages can’t, because they were written to persuade, which is a different job.
Name, address, and phone number
These belong in the footer on every page and again on the contact page, formatted exactly as they appear on the profile. Suite numbers, abbreviations, and phone formatting all count. Fifteen minutes, usually left undone for years.
Crawlability and robots.txt
A site that loads slowly or blocks crawlers doesn’t get read, and a page that isn’t read can’t be cited. Text locked inside images or rendered only by JavaScript has the same problem.
The part most owners have never checked is robots.txt, a file at the root of the domain that lists which automated visitors are allowed where. Anyone can read it by typing the domain followed by /robots.txt. The AI crawlers split along the same line as the two ways an assistant knows things: training crawlers collect text for future models, retrieval crawlers fetch pages when somebody asks a question, and blocking one has nothing to do with blocking the other.
- GPTBot collects content for training OpenAI’s models. OAI-SearchBot finds and links pages for ChatGPT’s search. OpenAI’s documentation states the two settings work independently.
- PerplexityBot indexes for Perplexity.
- Google-Extended controls whether content is used for Gemini training and grounding. It doesn’t affect ordinary Google Search.
The common failure is accidental: a security service treats the retrieval crawlers like scrapers, and the site quietly disappears from AI answers. Blocking all of them on principle is defensible for a publisher selling subscriptions and a self-inflicted wound for a plumber who wants the phone to ring. Just look for a line reading Disallow: / under any of the names above.
A related file, llms.txt, comes up often in AI search advice. It is a proposed plain-text summary of a site, placed at the root of the domain for AI systems to read. No major assistant has committed to reading it, and nothing in this section depends on it. It is harmless to add and not a substitute for a readable site and an open robots.txt.
When a rebuild is the faster path
Some of this is page-level work any site can absorb. But a site on a template that can’t add pages, or one where the service copy lives inside image files, costs more to patch than to rebuild. That’s when a website rebuild is the practical answer rather than an upsell, with structured data, location pages, and FAQ blocks built in. For a structurally sound site, none of that is necessary.
What is structured data, and does it actually help?
Structured data is a block of code on a page, invisible to visitors, listing the page’s facts in a fixed format. A person reading a service page works out that this is a plumber in St. Charles open until six; structured data says so outright. It’s also called schema, and the terms mean the same thing. It sits in the page’s underlying code, so it’s added by whoever built the website or through a plugin on platforms like WordPress, not typed into a page editor.
Four types cover almost everything a local service business needs: LocalBusiness or a specific subtype like Plumber, Service for each offering, FAQPage for question-and-answer content, and Organization to tie the business to its profiles elsewhere. The detail is in the guide to schema markup for local businesses. One caution: a business can’t mark up reviews it wrote about itself and expect Google to show star ratings; only third-party ratings count.
What it does, and what it doesn’t
Structured data removes ambiguity. It doesn’t create facts or raise a ranking on its own, and a page with perfect markup and nothing to say is still a page with nothing to say.
The rule that matters most is that structured data can hold nothing that isn’t visible on the page. Google’s guidelines say so even when the hidden information is accurate.
What questions should a website answer?
The Q&A box on a profile, covered earlier, works because it stores a real question with a direct answer under it. The same logic applies to the website. It’s the least technical of the signals in this article and often the fastest to show up in an answer.
Where the questions belong
On service pages first, since that’s where a purchase decision happens, then location pages. A standalone FAQ page is fine for policies, but questions about a specific service belong on that service’s page, where a model finds them in context.
Finding the real questions
The good ones are already being asked: on sales calls, in reviews, in the People Also Ask box on a relevant search, and in whatever ChatGPT says back when asked.
What doesn’t work is inventing questions to hold keywords. “Why choose our team for water heater repair” is marketing copy wearing a question mark. Nobody asks it, so nothing matches it.
What a quotable answer looks like
Direct answer first, complete in two or three sentences, then as much detail as the question deserves. An assistant that gets a full answer from the opening sentence can quote it. One that has to assemble it from three paragraphs moves on to a source that made it easier. The most common mistake is burying the answer under a warm-up line, one that repeats across every FAQ section on the site.
Each question gets its own heading, phrased the way a person would say it. Stacking three related questions under one heading makes all three harder to find.
What kind of content does AI actually cite?
Models cite content that holds facts they can’t get anywhere else. A post about what a repair typically costs in one metro area, written by the person who does the repair, is information that isn’t written down anywhere else. A general article assembled from other general articles contains nothing a model needs.
Specific beats comprehensive
The instinct is to write the biggest article on the broadest topic, which puts a local business in competition with every national publisher at once. Narrow, local, and concrete works better, and the reason is supply. A question like what a sewer line replacement runs in a specific county has almost no competition, because almost nobody has written it down.
First-hand expertise
First-hand knowledge shows up as detail nobody could guess: what fails on a twenty-year-old system in this climate, what it takes to acquire a permit in this county. It also shows up in what a business is willing to say plainly. Content that names the cases where a service isn’t the right answer reads as expertise. The alternative reads as a brochure.
Structure a model can parse
Definitions near the top. Headings that describe what’s under them. Tables for anything factual. Dates that reflect real revision, since models prefer current sources on anything that changes. The details are in the guide to writing blog posts AI search cites.
Planning what to publish
A blog that only answers questions gets read and never converts. A blog that only announces services gets neither. What separates the two is deciding what each post is for before writing it.
Posted Social sorts client content four ways: posts that get found by new people, posts that explain how the work is done, posts that show other people’s results, and posts that ask for the business. The names are Attract, Educate, Prove, and Convert, and the model is called PACE. For AI search, Educate carries most of the weight, because explaining how the work gets done is the one thing an assistant can’t assemble from anywhere else. The full model is in the PACE content allocation model.
Why does AI trust some businesses over others?
Two plumbers can have equally good websites and correct profiles, and a model will still name one more often. The difference is corroboration: whether other credible sources reference the business and agree with each other. It comes in two forms. A mention, where another site names the business. Or a link, where another site names it and points to the website. Links carry more weight, because a link is a site putting its own credibility behind the reference.
Domain authority, in plain terms
Domain authority is a third-party estimate of how much weight a site carries, scored out of 100. Google doesn’t use it; Moz created it, and other tools have their own versions. It’s a rough gauge of whether corroboration is working, not a target. Chasing it is how businesses end up buying links, which costs more to fix than the weakness it was meant to cover.
Listings and consistency
A listing is an entry for the business on a site it doesn’t own, and most exist already, created by the platforms from public records. The core set is small: Google, Bing Places, Apple Business Connect, Yelp, Facebook, the Better Business Bureau, and the two or three directories that matter in a given trade. Consistency matters more than volume, and the audit is in the guide to domain authority and local listings.
Where AI looks for local recommendations
Four places come up again and again. Yelp functions as a direct feed for some assistants, so a neglected Yelp page carries more weight than its traffic suggests. Reddit turns up constantly because it’s real people answering each other’s questions; a business can’t manufacture that, and recommending yourself under an unmarked account is how a brand gets attached to the word astroturfing. Nextdoor matters disproportionately for home services for the same reason. YouTube is the one most local businesses skip, covered below.
Reviews and responses
Review volume and recency both count, and the responses are the underused part. A response is public, permanent, and adds text naming the service, the location, and the outcome in the business’s own words, on a platform these systems read. A thoughtful reply to a critical review does more than a defensive one, and more than none.
Local press and partnerships
A youth-team sponsorship, a chamber membership, a quote in the neighborhood paper, a manufacturer’s certified-installer page: none of those sites gain anything by making it up, which is where their weight comes from.
Do photos and video affect AI visibility?
Yes, though the value is mostly in the text attached to an image rather than the image itself, because text is what a model reads without ambiguity.
File names and alt text are the bigger win
An image saved as IMG_4471.jpg tells a system nothing. A file name that describes what’s in it tells it something. Alt text is the short written description a website stores alongside every image, originally for screen readers and now the only part of a photo a system reads with confidence.
This is tedious and it is genuinely effective, which is an unusual combination. It’s also what gets skipped most, because nobody enjoys renaming ninety files after a shoot.
Geotagged images
Photos can also carry hidden information inside the file, called metadata: where it was taken, what business it belongs to, who produced it. Posted Social embeds these on client shoots as standard practice. It’s a smaller signal than a good file name, worth doing for a business already producing photography and not instead of the basics. The workflow is in the guide to naming and geotagging photos.
Video, and where it goes
Video pays off in one place, YouTube, because Google owns it and Gemini can draw on it directly. A short video answering a common customer question, titled the way a customer would ask it, sits where most local competitors have never looked.
Social content counts too
Social profiles are public, indexed, and read, and they’re where service descriptions and phone numbers drift out of date fastest. A website update that skips them leaves the old details in place. The guide to social media for local service businesses covers the rest. For businesses shooting regularly, naming and tagging belongs in the content production process, not bolted on afterward.
How do you measure AI search visibility?
Google now reports part of this. Everything else gets checked by hand, on a schedule, and recorded so one month can be compared to the last.
Manual testing
The manual check starts with five questions a customer would ask, in a customer’s words, including one about the service that makes the most money. Each goes into ChatGPT, Gemini, Google’s AI Mode, and Perplexity, with a record of whether the business was named, what was said about it, and which competitors came up instead. The same five run again each month. One month means little, because these systems return different answers to the same question on different days. The pattern across three months is the signal.
Check the business by name
Asking each assistant what it knows about the business surfaces whether the information is current and whether it has confused the business with another one. Wrong information is more common than none, and it’s fixable at the source, usually a stale listing or an outdated page rather than the assistant itself.
Search Console and Google Business Profile
Google Search Console is a free tool that shows how a site performs in Google search: which queries brought it up, how many people saw it, how many clicked. On June 3, 2026, Google added generative AI performance reports to it, rolled out to every site by August 31. They show how often a site’s pages appeared in AI Overviews, AI Mode, and Google’s AI features in Discover. Impressions only, no clicks or question data yet, and Google only. Nothing equivalent exists for ChatGPT or Perplexity, which is why the manual check stays the larger half of the job.
The same update added a switch that blocks a site from AI Overviews and AI Mode. It doesn’t affect ordinary rankings, and it gives up exactly the impressions the report exists to measure. Worth knowing it’s there, not worth using.
Google Business Profile insights cover calls, direction requests, and website clicks, the actions that follow a mention, so they’re the closest thing to a conversion measure available.
Tracking tools
A category of software automates the manual check. Otterly, Peec, Semrush’s AI toolkit, and Profound are among the names, at published prices from about $29 a month to several hundred for multi-assistant plans. Different tools report different numbers for the same business in the same week, because they measure different things, so the number is a trend line rather than a fact. For a single-location business, a spreadsheet does the job.
Where Posted Social fits
Most of what this article covers is work a business owner can do without hiring anyone. The profile audit, the Q&A box, the file renaming, the monthly check: all of that costs time rather than expertise.
What tends to stall is the rest. A site that can’t hold the structure. Content that needs publishing for longer than a month to matter. A monthly check that everyone means to run and nobody runs. That’s the work Posted Social’s SEO practice takes on, and when a site can’t support location pages or structured data, the web design side goes first. For a business that would rather hand over the whole sequence, that conversation starts here.
But the first step doesn’t require any of it. The phone number on the Google Business Profile and the one on the website either match or they don’t.
Either way, that is the honest starting point.
Common questions
Is AI search visibility the same thing as SEO?
Mostly, yes. Being findable in AI answers means being indexed, consistent, specific, and corroborated, which is what local SEO has always meant. What’s new is that one answer replaces ten links. AEO and GEO describe the same job.
How long does it take to show up in AI search?
Profile and listing fixes can register within weeks, because assistants look information up fresh. Content and corroboration take months, because they depend on other sites and on publishing consistently. Nothing here works in days.
Can a business correct wrong information an AI gives about it?
Not directly. Wrong information almost always traces to a public source the business does control: a stale listing, an old page, an unclaimed profile. Fix the source and the answer generally follows within weeks.
Does blocking AI crawlers hurt a local business?
It depends which ones. Blocking the retrieval crawlers, the ones that fetch pages to answer questions, removes the business from AI answers. Blocking the training crawlers keeps content out of future models without affecting whether it gets cited today. The two are separate decisions, and only the first affects visibility.
[Implementation note for Kris: this block needs FAQPage structured data matching the visible text exactly. See the log about FAQ rich results before promising one.]
Sources
- Google Search Central, AI features and your website
- Google Search Central, guide to optimizing for generative AI features
- Google, new opportunities, control and insights for website owners, June 3, 2026, updated August 31, 2026
- Google Search Central, general structured data guidelines
- Google Search Central, intro to structured data markup
- OpenAI, overview of OpenAI crawlers
- Perplexity, Perplexity crawlers
- Schema.org, LocalBusiness vocabulary reference (link to be confirmed at publish)
- Moz, Domain Authority explainer (link to be confirmed at publish)