Before you change anything, know what the machine is reading
Ask an assistant who fixes commercial fridges in your town and it will not hand you ten blue links. It will write a short paragraph, name one or two businesses, and list the sources it leaned on. That paragraph is assembled at the moment you ask, from documents the system went and fetched. It is not recalled from memory.
That single fact explains most of what follows in this course. Large language models do not hold a reliable, current record of your opening hours. What they hold is language. The specifics come from retrieval: the system finds documents that look relevant to the question, reads them, and writes an answer grounded in what it found. The industry calls this retrieval-augmented generation, usually shortened to RAG, and grounding is the term for tying the generated sentence back to a retrieved source.
So the question stops being “how do I rank” and becomes “what documents exist about my business, and do they agree with each other”.
The four stages, and where you sit in them
Every local recommendation goes through roughly the same sequence.
Query. The assistant takes the customer’s phrasing, works out the intent, and often rewrites it into something more searchable. “Who can fix a walk-in cooler tonight” becomes a location-scoped query about commercial refrigeration repair with an urgency signal attached.
Retrieval. The system gathers candidate documents. For a local question, that usually means Google Business Profile data, pages from business websites, and third-party sources such as directories, review sites and local press. Retrieval is typically semantic rather than keyword-exact: the system compares meaning using vector search, which is why a page phrased in customer language can be pulled even without an exact keyword match.
Grounding. The retrieved passages are attached to the prompt so the model writes from them rather than from memory. This is the stage that decides whether your facts or somebody else’s end up in the sentence.
Answer. The model writes a short response and, in most interfaces, attaches citation attribution: the little numbered links or source cards under the paragraph.
You cannot edit stages one, two or four. You have a great deal of control over what is available to be retrieved at stage two and grounded at stage three, which is where this course spends its time.
Three sources, unequal weight
For a local business, retrieval almost always draws on three kinds of source.
Your Google Business Profile is the densest single record of your facts: name, address, phone, category, hours, services. Engines read it directly and lean on it heavily for local questions. Lesson 2 is entirely about it.
Your website is where nuance lives. A profile cannot explain that you service walk-in coolers but not domestic fridges, or that your emergency callout covers three towns and not a fourth. Only your pages can answer that, and only if they say it plainly. Lesson 3 covers the shapes of page that get pulled.
Third-party mentions are everything else: directory listings, review sites, chamber pages, local news, a forum thread where somebody recommended you. These matter more than owners expect, because what other sites say about you is treated as independent corroboration. Lesson 6 covers them.
| Source | What it is best at | Where it falls short |
|---|---|---|
| Google Business Profile | Core facts: name, address, phone, category, hours | Cannot express nuance or exclusions |
| Your website | Scope, exclusions, pricing logic, service detail | Only counts if the page states it plainly |
| Third-party mentions | Independent corroboration of your facts and reputation | You do not control when they are updated |
The common failure is not absence, it is contradiction. One directory carries an old phone number, the profile says you close at five, the website footer says six, and a listing from two addresses ago is still live. Faced with conflicting sources, an engine hedges, picks the majority reading, or names the competitor whose facts are unanimous.
Being named is not the same as ranking
Classic SEO chases a position. AI search produces a different unit of value: your business appears inside the sentence, sometimes with a link and sometimes without. That is the zero-click discovery problem in one line. A customer can learn your name, your speciality and your hours without ever visiting your site.
The vocabulary around this is messy. Some people call optimising for it Generative Engine Optimization, or GEO. Others call it Answer Engine Optimization, or AEO. The guide on what GEO means untangles the terms, and the AEO versus GEO versus local SEO comparison settles which one you are actually buying when a vendor pitches you. The labels matter far less than the work, which is consistent facts, citable pages and machine-readable data.
One practical consequence: your measurement changes. Sessions and rankings no longer capture the whole picture, because some of the value happens inside an answer you never see. Lesson 7 deals with that honestly, including what still cannot be measured.
The crawlers that make any of this possible
Retrieval needs a copy of your page. That copy arrives through named crawlers, and they do not all do the same job. GPTBot and ClaudeBot collect content for training. OAI-SearchBot and PerplexityBot fetch pages to support search and live answers. ClaudeBot and its search-side counterpart behave similarly for Anthropic’s assistants.
The distinction matters when somebody suggests blocking them. Blocking the search-side crawlers removes you from the answers you are trying to appear in. The guide on which AI crawlers read your site lists the user agents so you can recognise them in your server logs, and Lesson 5 covers what to allow.
Where you can realistically intervene
Three places, in this order.
Make your facts unanimous. Every place a customer or an engine can read your name, address, phone number, category and hours should say the same thing. This is unglamorous and it is the highest-leverage work available to a small business.
Make your pages answerable. Write passages that stand on their own: a clear question, a direct answer, a specific fact. A sentence lifted out of your page should still make sense and still be true. That is what gets quoted.
Make your data machine-readable. Structured data hands the engine your facts in a form it cannot misread. Lesson 4 covers LocalBusiness schema, including a block you can paste and validate the same day.
If an assistant is naming a competitor instead of you, one of those three is almost always the reason. The diagnostic guide on why AI recommends a competitor walks the three checks in about thirty minutes.
What to do next
Run the test yourself before you read Lesson 2. Log out, ask the question a customer would ask, and write down who gets named and which sources the answer cites. That list is your starting position, and Lesson 7 turns it into a repeatable monthly check.
Then move to Lesson 2, which takes the single most-read source of facts about your business and makes it something an engine can trust.
