What AI tells customers about your business, and the 15-minute check

Happy Wednesday.

Sometime this week, somebody probably asked Claude, ChatGPT, or Gemini about your business. Who owns it. Whether it's any good. Who's the best in your category nearby. The machine gave them one answer, you never saw it, and nobody checked whether it was right.

This issue explains the two acronyms your next marketing invoice may carry, AEO and GEO, in plain English. Then it hands you a fifteen-minute check that shows exactly what the machines say about your business, with a printable scorecard so you can rerun it every month. I ran the check on three real businesses in July. Two came back with the wrong owner.

Also in here: what happened when I ran it on my own company, three things worth your time, and one vendor pitch decoded.

Here is how a customer found a business in 2020: they typed “estate lawyer near me” into Google, got ten blue links, and you paid an SEO person to be in the top three. A growing share of customers now do it differently. They ask Claude, ChatGPT, or Gemini “who should I use for this around here” and get a short paragraph naming one or two businesses. No list, no page two. Either the machine names you and gets you right, or you were never in the running.

The new acronyms are names for that shift. SEO was search engine optimization, the work of ranking in Google's list. AEO is answer engine optimization. GEO is generative engine optimization. Both describe the same job: being the business the AI names, correctly, when somebody asks. Two names exist because the field is barely two years old and the vocabulary hasn't settled. You don't need to pick a favorite. You need to know what the line item means when a vendor puts either one on an invoice, and that line item is coming.

The difference from SEO that can actually hurt you: Google's list could bury you, but it couldn't lie about you. A generated answer can. When these systems are missing a fact about your business, they don't say “I don't know.” They fill the gap with something plausible and deliver it in the same confident tone as everything else. Ask about a business the machine has barely seen and you still get a complete, assured answer. Some of it is invented.

Where do the answers come from? From whatever the machine can reach: its training memory, your website if the text on it is readable by a machine, your Google and Yelp profiles, directories, old press. If the reachable version of you is stale, the answer is stale. In my testing, the fact machines get wrong most often is the first one a customer would check: who owns the business today. A ten-year-old press mention and your current about page carry equal weight in the machine's reading, and the old mention often wins.

One more mechanic before the check: phrasing changes the answer. Asked carefully, like a researcher, a machine will often get you right. Asked casually, like a real customer, the same machine on the same day can get you wrong. A useful check has to use customer phrasing.

The check works like a mystery shopper. Same script every visit, written down, so a change between visits means something. Fifteen minutes, five steps:

1. Write five questions a stranger would ask. Who owns your business. What it actually sells. What people say about it. Who's the best in your category in your area. What's changed there lately. Write them the way a customer talks, not the way you'd describe your own company.
2. Open two assistants, fresh. Claude and ChatGPT, or Gemini and Perplexity. New conversation, nothing pasted in. Ask the five questions one at a time, exactly as written.
3. Score each answer with four checks. Named you at all? Facts right, starting with the owner? Cited your actual website? Put you on the best-of list? Yes or no, four marks per answer.
4. Save it all in one file, with the date. Questions, answers, scores.
5. Rerun the same file next month. Same wording, same assistants. What moved is the finding.

The fixed script is what makes step five work. If the questions drift between runs, a changed answer tells you nothing, because you changed the input. Hold them still and a changed answer means the machines' picture of you moved. The same discipline runs through everything you'll ever hand an AI: instructions precise enough to produce the same result twice.

When does this matter? If you pay a marketing, SEO, or IT vendor a monthly retainer, before your next call with them, because the pitch is coming (decoded below). Beyond that, a little more each quarter. Traffic from AI answers is a trickle today, but a person arriving from one has already been told an answer and already picked. Nobody has won this yet, including your competitors; most businesses have never run this check once.

The manual check above is where I started. After a few rounds I taught my own system to run it, and the built version does five things in sequence. It reads the business's own website first and freezes the true facts: owner, locations, services, claims. It writes the five questions in two phrasings each, one careful, one the way a customer would type it. It runs all ten across the assistants and scores the four checks. It ranks every wrong answer by how much damage the error does. Then it writes each finding next to its fix, so what comes out is a work list, not a report card.

I pointed it at three real small businesses. Two of the three came back with the wrong owner. One was described as belonging to somebody who sold it years ago: present tense, under a heading about leadership, current owner absent. The other was placed in a business partnership that ended in 2024, also present tense, and the machine cited the company's own current website in the same answer. It had looked at the right page and still repeated the old fact, because an outdated mention somewhere else outweighed it.

One of the two also got credited with an award it has never won. The award belongs to a direct competitor. Flattering errors survive longest; nobody calls to complain that the machine said something too nice.

The third business came back clean. Right owner, right locations, ranked in its category, on a tidy modern site with real reviews. It had never done a minute of deliberate work on this. Clean is reachable.

Most of what was broken was cheap to fix. A years-old press quote naming the old owner, sitting on a page the business controls. A homepage that gives a machine exactly 41 words of readable text. An empty description field the website platform generated and nobody filled. Nearly all of it sat on pages the businesses own, fixable in an afternoon, no budget required.

If you ever hire this out, the spec is one sentence: same questions, two phrasings, scored, run monthly, compared against last month. The scorecard is how you check the homework.

Before pointing this at anyone else, I aimed it at Praxis. Six other AI companies share the name. Asked about mine, the machines confidently described one of theirs instead. The one model that actually read my website got everything right, which told me the site was fine and the noise around the name was the problem. Different diagnosis, different fix.

The newsletter did worse. I renamed it in July, and days later the machines had never heard of the new name. Pressed, they didn't admit that. They invented a plausible newsletter with the wrong audience attached. Both findings went onto my own fix list, same grid as everyone else's.

Yelp is licensing its reviews, ratings, and photos to OpenAI. Announced July 23. ChatGPT now pulls Yelp data directly into answers about local businesses, and as of August 10 it can book a table or join a waitlist through Yelp without leaving the chat, with quote requests for service businesses still to come. The Yelp profile you half-finished years ago is now source material for the machines, photos and review count included. Worth thirty minutes of cleanup this month.

Google's AI Mode now connects to apps and completes tasks. Instacart, Canva, and YouTube Music at launch, US users first, announced July 16. The box that used to answer questions is becoming a place where the errand gets done. Being the business it names gets a little more valuable each quarter, because the machine is moving from recommending to transacting.

Two thirds of US small businesses now use AI, and 70% of owners say they need more training. A 561-owner survey published July 15 puts adoption at 66%, up eleven points in a year, with 86% saying they are somewhat to extremely comfortable using AI. Comfortable and effective are different claims, and the distance between those two numbers is where most wasted AI spend lives. Written-down, repeatable checks close it faster than new tools do.

Your SEO vendor is about to offer an AI visibility package, if they haven't already. Some of these are real work: fixing the readable text on your pages, cleaning up old mentions, tightening your profiles. Some are last year's retainer with a new label. The label won't tell you which. Two questions will.

First: “Show me what the assistants say about us today.” Anyone selling improvement can show a starting point. If there's no before, there's no measurement, and you're buying activity.

Second: “Which exact questions will you rerun every month, and can I see the wording?” A fixed, written question set means you'll see what moved. A vague promise to optimize for AI means nobody will ever know whether it worked.

You've run the check yourself, so you already know what good answers to both look like.

Run the five questions this week, or grab the playbook and do it properly: the scorecard, the prompts, and the skill are all there. Then reply with the strangest thing a machine said about your business. I read every reply.

If you want to learn how to build this, and other useful tools, there are a few options here.

See you next Wednesday.