GEO for local businesses: what agencies can measure and improve

Generative engine optimization, or GEO, is the work of making a business’s public information easier to find, understand, and use in AI-generated answers. For an agency serving local businesses, the useful starting point is a set of real buyer questions and the answers they produce.

This guide stays on that ground. It does not promise a ranking, a citation or a lead. It describes what you can observe about a local business in AI answers, what you can change on the business’s own pages, and how to evaluate whether a change did anything.

Start with the questions a buyer asks

A buyer who uses an AI assistant does three things in roughly this order: they look for a provider, they compare a shortlist, and they check one name before calling. Each moment is a different question, and each produces a different answer.

• Finding: “Who does emergency plumbing in the north of the city?”
• Comparing: “Is X or Y better for a family dental practice?”
• Checking: “What do people say about X?” or “Does X take new patients?”

These are illustrative questions written for this guide. They are not observed queries and carry no search volume.

Write your own set before you look at any answer, in the words a buyer in that market would use, and keep them fixed. A question you rewrite after seeing the answer is no longer a test.

Separate being named, being described, and being recommended

An AI answer can name a business, describe it, or recommend it, and these are three different observations. A business can be named in a list and described with a service it does not offer. It can be described accurately and never recommended. A mention proves neither accuracy nor preference, so record the three separately:

• Named. The business appears in the answer at all.
• Described. What the answer says about it, compared with the facts you can verify.
• Recommended. Whether the answer puts it forward when asked for a choice, and the reason it gives.

Keeping them apart matters because the fix is different for each. A business that is never named has an information problem. One that is described wrongly has an accuracy problem. One that is named and described but never recommended has a proof problem: the answer found nothing to prefer it for.

Check answers with and without web search

Most assistants can answer in two configurations: with web search disabled, from what the model already holds, and with web search enabled, where it can read pages while answering. Run every question in both, and keep the results side by side.

The comparison is useful because the two configurations often disagree. A twelve-person firm may be absent when search is off and present when it is on, or the reverse. That tells you where the work is: a business missing with search enabled usually has pages the model could not read or did not find; a business missing with search disabled has simply not been written about enough to be retained.

What the comparison does not tell you is exactly what a model has memorised, or that a provider actually opened a specific page because search was on. Treat the two modes as test conditions you control, not as a window into the model.

Work on information you can actually improve

Most of what an AI answer can say about a local business comes from pages the business controls and from third-party pages that describe it. Start with the first kind, because that is where an agency can act this month:

• A page per service, saying exactly what it covers and what it does not.
• The area served, stated plainly, with the neighbourhoods and towns named.
• One consistent identity: the same name, address, phone and opening hours everywhere.
• Real proof: dated results, named credentials, reviews that can be traced to a source.
• Internal links between these pages, so a crawler reaches each from the home page.
• Availability for crawlers: no page that matters behind a script, a login or a robots rule.

What not to do: generate a page per town with the same text and a different place name. Google’s guidance on helpful content is explicit that pages made for engines rather than people are the problem, not the solution, and its notes on AI features say the same practices apply there.¹ That guidance is Google’s; do not assume the other providers publish the same rules.

Two things that are often sold as shortcuts: structured data and an llms.txtfile. Both can help a page be understood; neither makes a business get cited. Publish them if they describe something true, and do not expect them to substitute for a page a buyer would actually read.

Turn a finding into client work

A finding becomes work when you can say what was observed, what you checked, what you propose, and who does it. The reasoning runs in one direction:

1. Observation. “With search enabled, four of the seven models answered the ‘who does X in this area’ question without naming the business.”
2. Check the page. Does a page exist that says the business does X in this area? Can a crawler reach it? Does it say so in the words the buyer used?
3. Propose. If the page is missing, propose it. If it exists but says something else, propose the correction. If it exists and is fine, say so and look elsewhere — the finding may not be a page problem at all.
4. Assign. The agency writes and publishes; the audit provider measures. Keep that line clear in the proposal.

This is a conditional chain, not a case study. The point is that each step depends on the previous one being checked, and that a proposal without a checked page is a guess.

Deliverables that follow naturally from it:

• A dated record of the answers, per question, model and configuration.
• The factual errors found, each with the source the model gave, if it gave one.
• The pages to clarify, add or fix, in priority order.
• The check that will be run again, and when.

Retest without promising a ranking

After the changes ship, run the same questions on the same models in the same two configurations, and compare. Frame the comparison carefully: models change between runs, and an answer can move without anything having changed on the site. A retest shows movement; it does not prove that your change caused it.

And keep the measurement separate from traffic and leads. A business being named more often is a real observation; whether that became a call is a different question that this kind of test cannot answer. Say both parts to the client.

Start with an observable problem

Start with an observable problem and a change your agency can explain. A useful GEO report makes that connection clear and keeps the limits of the measurement visible. If you want to see what that looks like on a real business, the sample report shows one run, with the answers behind it.

1. Google Search Central, “Creating helpful, reliable, people-first content” and “AI features and your website” (developers.google.com/search). Both describe Google’s own systems only.