How it works

Start with the questions buyers actually ask

Everything starts with a list of questions. Not keywords. Full questions, the way a person types them into an assistant.

I build the list from three things: the services you want to be found for, the areas you serve, and the questions your customers already ask you. Then I add the questions people ask right before they decide.

• Asking for a pick: “Who is a good roofer for an older home?”
• Comparing options: “Should I repair or replace my water heater?”
• How to choose: “What should I look for in a bookkeeper?”
• About you by name: “Is [your business] any good?”

Once we agree on the set, it stays fixed. Same wording every month. If the questions kept changing, you couldn't tell whether your visibility moved or the test did.

Run every question in a clean session

Assistants remember things. If I asked from an account with a long history, that history could shape the answer. So each question runs in a fresh session with no past chats behind it.

I run the set in ChatGPT, Claude, Gemini, Grok and Perplexity. I also check Google AI Overviews and AI Mode when they appear for a question.

I log the date and the exact wording every time. Answers can vary from one day to the next, even for the same question. A clean, logged method is how I keep that noise from being mistaken for a result.

Record the rung, the names and the sources

For every answer, I write down four things.

• The rung. Absent, cited, mentioned, recommended, or recommended against.
• Who got named instead. The businesses the assistant put in front of your buyer.
• The sources. Which pages and sites the answer leaned on.
• Anything wrong. Old facts, a wrong phone number, services you don't offer.

The sources matter most. They show which sites an assistant trusts for your kind of business. That's where the work goes.

Audit the foundations

Before anything else, I check that assistants can reach your site and read it. These problems are common, and many are quick to fix.

Crawler access in robots.txt

Your robots.txt file tells crawlers what they are allowed to read. I check it for each of these, by name.

OpenAI's help center says any public website can appear in ChatGPT search, and that sites should not block OAI-SearchBot if they want to be included. Read OpenAI's note.

Files and markup I check

• robots.txt

  Does it let the crawlers above in? One wrong line can shut an assistant out of your whole site.

• Raw HTML

  Is your main content in the page source, or only drawn in later by JavaScript? Some crawlers only read the raw HTML.

• Schema

  Organization, Service and FAQ markup, filled in with real data that matches what the page says.

• llms.txt

  A short plain text guide to your key pages. It's optional and no assistant has promised to use it, but it costs little to add.

• Name, phone, description

  The same business name, phone number and description on your site and everywhere else.

Google says AI Overviews and AI Mode have no extra technical requirements beyond being indexed and eligible to show in Search with a snippet. That's why the ordinary basics matter so much. Read Google Search Central on AI features.

Fix the entity and listings layer

Assistants check what the rest of the web says about you. If your name, phone or services differ from one site to the next, that's a reason for doubt. So I make the facts match.

• Google Business Profile, where your business is eligible
• Bing Places
• Apple Business Connect
• The directories and review platforms that assistants cite for your type of business

That last item comes straight from your ledger. I don't work from a generic list of a hundred directories. I work on the ones that actually showed up in the answers.

Write pages that answer the real questions

Many sites say what a business does, but never answer what buyers ask. Answer pages fix that.

• Question pages that answer the exact questions in your prompt set, in plain words, near the top
• Comparison and “how to choose” pages for the decisions buyers make before they call
• Sourced facts: claims you can stand behind, with quotes and links to where they come from

This follows published research. In the GEO paper by Aggarwal et al. (KDD 2024), adding citations, quotations and statistics raised visibility in generative engine answers in the authors' tests. It's one study, not a promise. But it points the same way as common sense: be clear, and back it up.

Re-test every month and report in writing

Each month I run the same prompt set, the same way, and compare it to the month before. Then I send a short written report. Here is what it covers.

Monthly AI visibility report

Business: Your business / Month: Month

Sample

1. Summary

   Three or four plain sentences. What moved, what didn't, and anything that needs a decision from you.

2. Prompt by prompt

   • Each question, in each assistant
   • This month's rung beside last month's rung
   • Any answer that got facts about you wrong

3. Who got named instead

   The businesses that showed up where you didn't, and how often.

4. Sources the answers leaned on

   The sites and pages assistants used, and whether you are on them.

5. Work done this month

   What I changed: files, schema, listings, pages. Each item listed plainly.

6. Plan for next month

   What I'll work on next, and why, based on what this test showed.

No dashboard to log into. No charts dressed up to look like progress. Just what changed, what didn't, and what I'll do next.