Why an AI answer isn't the same as a Google ranking, and what decides whether your business gets named.
When someone asks ChatGPT or Gemini for the best plumber in Salem, they get back a short list of names and a paragraph explaining the choices. No ten blue links. The useful question for a business owner isn't whether this is happening. It's how those names get chosen, because that part you can actually do something about.
What is an AI assistant?
A program that generates text. It was trained on an enormous amount of writing, and what it does when you ask it something is produce the words most likely to follow your question. ChatGPT, Gemini, Claude, Perplexity, and Copilot all work this way, as does the AI summary Google now puts above its own search results.
It has no directory of businesses. There's no list inside it, no database of who's good in your town, no record of your hours. It isn't looking you up in a table, because there's no table.
It sounds equally sure either way. The answer comes out smooth and specific whether or not it knows anything real about you. When a person is guessing, you can usually tell. Here you can't. So the question was never whether the answer will sound convincing. It will. The question is what it had to go on.
So the answer is only as good as what the assistant could find. That's the part you can do something about: not the model, but the sources it reaches for.
AEO and GEO describe two different things
The terms get used interchangeably, which makes them less useful than they should be. I keep short definitions of both on the method page, but the distinction worth understanding is this.
Answer Engine Optimization is about direct questions with a correct answer. "Who does emergency plumbing in Salem on a Sunday?" has a right answer. AEO is the work of being it, and it depends on whether your basic details agree with each other and are easy for a machine to read.
Generative Engine Optimization is about the paragraph the assistant composes. When it writes "a few well reviewed options include," GEO is what determines whose name appears in that sentence. It depends less on having your facts complete and more on whether anything out there says something specific about you.
They overlap, but they fail differently. A business can have a flawless Google Business Profile and still never get described, because nothing anywhere says what makes it worth choosing.
Retrieval is the step that matters
Traditional search returns documents and asks you to pick. You optimize a page, it ranks, someone clicks, they arrive on your site. You control the destination.
An AI assistant answering a local question usually does something different. It runs a search behind the scenes, pulls up a handful of pages, reads them, and writes its answer out of what it just read. The customer often never visits your website.
That changes two things.
First, you're competing to be retrieved and quoted, not to be clicked. The relevant question is no longer "where do I rank" but "when the assistant searches this, is my business among the sources it pulls, and does the page it lands on state clearly enough what I do for the model to repeat it."
Second, much of what gets said about you isn't on your site. Assistants pull from your Google Business Profile, your reviews, directories, local news, and anyone else's page that happens to list businesses like yours. When I audit a site for AI visibility, I often find the assistant describing a business accurately but working from somebody else's page to do it, usually a "best plumbers in Salem" style list published by a competitor or a directory. Whoever sits at the top of that list usually gets named first.
Training data and retrieval are separate questions
These two get mixed up constantly, and mixing them up wastes real effort.
Some crawlers gather text to train future models. Others fetch pages at the moment someone asks a question. OpenAI runs both: GPTBot collects training data, while OAI-SearchBot and ChatGPT-User handle search and live fetching. Anthropic and Perplexity make similar distinctions.
What that means in practice: if your site blocks the crawlers that fetch pages, you can be missing from answers today no matter how good your pages are. Blocking only the training crawlers is a fair thing to want, and much slower-acting, since it shapes what future versions know about you rather than what today's assistant can look up. Both are reasonable positions. They're just not the same one, and a site's robots.txt often picks by accident. It happened here: my own host was adding a block on the retrieval crawlers above the file where I'd explicitly welcomed them, so this site was arguing with itself. I only caught it by reading the file as a crawler would. The audit I ran on this site covers what a full pass looks like.
What actually moves the outcome
Your Google Business Profile, completely filled in. Hours, categories, services, service area, attributes, photos, the description. It's the most reliable record of the plain facts about a local business, and it happens to hold exactly the details people ask about. An incomplete profile is the single most common thing I find, which is why it turns up again in the patterns that show up in almost every audit.
Reviews that say something, and keep coming. The star average matters less than the text. Assistants summarize what reviews say, so reviews that name specific services give a model something concrete to repeat. Recency matters because a model working from five year old reviews has less to say about you, and vague praise reads as a weaker recommendation than specifics.
Pages that answer a question in the first sentence. A model pulling an answer off your page needs it stated plainly, near the top, in the words a customer would use. Long preambles tend to get skipped. It's the same instinct as writing for the answer box that sometimes sits at the top of Google results, applied more broadly.
Structured data that matches what the page says. Schema markup for your organization, services, location, and FAQs spells out what your business is and where it operates. It doesn't make you more trustworthy. It removes guesswork, which is a different thing and a more achievable one.
Being on the lists that already get quoted. Local directories, chambers, association listings, and the "best X in town" lists that rank for your category. If the assistant reliably reaches for one local list when answering questions in your industry, not being on that list costs you more than most website work would gain.
What's genuinely uncertain
Some of this isn't knowable yet, and it's worth saying so.
Nobody outside these companies can see how the ranking actually works, and it changes without notice. It's also harder to measure than ordinary search: two people asking the same question get different answers, so checking once tells you almost nothing. You have to ask the same set of questions over and over, across each assistant, and watch which way it drifts.
What holds up is unglamorous. Accurate, consistent, well structured information about a business tends to get retrieved and repeated. That was true before these tools existed, which is why most of this work looks like local SEO done carefully rather than something new.
Where to start
Before changing anything, find out what the assistants currently say when someone asks about businesses like yours, and which sources they're using to say it. That's the first thing I check in a Clarity Review, along with whether your site is reachable by the crawlers that matter. There's a sample review if you want to see the format first. And if things are already working, I'll tell you that too.