01 ·Context
One answer instead of ten links, and no error message when it is wrong.
In 2020 a customer typed "estate lawyer near me" into Google and got ten blue links. A growing share of them now ask ChatGPT or Gemini and get a short paragraph naming one or two businesses. Either the machine names you and gets you right, or you were never in the running. This shift is what the acronyms AEO and GEO refer to: answer engine optimization and generative engine optimization, two names for the work of being the business the AI names, correctly, when somebody asks.
A generated answer can get a basic fact wrong and still sound confident. Check ownership, services, locations and claims against current sources. An answer that flatters the business deserves the same scrutiny as a negative one.
The check on this page works like a mystery shopper. Same script, every visit, written down, so a change between visits means something. You run it once this week and then once a month, and you know exactly what the machines are telling people about you.
02 ·Architecture
The check has four parts.
Part 01 · Five questions, two phrasingsWritten like a customer, not like the owner.
Ownership, what you sell, reputation, the best-of list, and recency. Each asked twice: once carefully, once the way a real customer types. Casual phrasing finds errors that careful phrasing hides.
Part 02 · Two assistants, freshNew conversations, nothing pasted in.
Choose two assistants your customers may use. Start fresh conversations without supplying your answer key. Use an unpersonalized session where available and record the settings so later runs are comparable.
Part 03 · Four checks per answerYes or no, no essays.
Did it name you? Were the facts correct? Did it cite your website? For the recommendation question only, were you listed? Use N/A where a check does not apply and keep each phrasing separate.
Part 04 · One dated file, rerun monthlyThe file is the asset.
Save the questions, answers, sources, date, assistant and search settings. Keep the wording fixed next month. Answers can vary between runs, so a changed score is a signal to investigate, not proof that one website edit caused it.
03 ·Setup
Print the scorecard, or install the skill.
Nothing to install for the manual run: the two copy boxes in the runbook below are the whole thing. If you want paper, the scorecard is one printable page with the questions, the scoring grid, a fix list, and the monthly log.
The downloadable skill can help organize questions, collect your pasted answers and prepare a scorecard. To install a Claude skill, package it as SKILL.md inside a zipped folder and use the current upload flow. A loose file in a project is reference material, not an installed skill. Read Writing your first skill first if this is new.
Product instructions checked October 26, 2026. Install and enable Claude skills ↗
04 ·Decisions
Two calls before your first run.
Decision 01Manual first, or skill first.
If this is your first run, do it manually with the copy boxes below. Reading the machines' answers with your own eyes, in their own confident wording, is worth more than any summary of them. Once you have felt that, let the skill absorb the tedium of the monthly rerun.
Decision 02Which two assistants.
Pick two assistants your customers use and keep the pair fixed. Record the model or mode and whether web search was enabled. If the product or settings change between runs, note that alongside the score.
05 ·Runbook
The ten questions, then the scoring prompt.
Step one: fill in your business, category, and town, then ask each question in a brand-new conversation in each assistant. Copy the answers somewhere as you go.
Replace [business], [category], and [town], then ask each question in a brand-new conversation, one at a time, exactly as written. 1a. Who is the current owner of [business]? 1b. who owns [business]? 2a. What products and services does [business] offer? 2b. what does [business] actually do? 3a. What do customer reviews say about [business]? 3b. is [business] any good? 4a. Which [category] businesses in [town] are most recommended? 4b. who's the best [category] in [town]? 5a. What recent changes or news are there about [business]? 5b. what's new at [business]? The a-version is careful phrasing. The b-version is how a customer types. You need both: casual phrasing finds errors that careful phrasing hides.
Step two: paste the scoring prompt into any AI, add your true facts and the collected answers, and it scores the run and writes your fix list. Use an assistant that did not answer the questions, or a fresh conversation at minimum.
TASK: Score these AI answers about my business and write me a fix list. CONTEXT: I asked two AI assistants ten questions about my business. Below are the true facts, then the answers. Score every answer separately with four checks, yes, no or N/A: NAMED (it mentioned my business at all), FACTS (details right, starting with the owner), CITED (it referenced my actual website), LISTED (question 4 only: my business appeared on the best-of list). FORMAT: One score table, one row per question phrasing (1a through 5b) per assistant. Keep all twenty answers separate. Under it, quote every wrong sentence verbatim. Then a fix list: for each error, the source only where the answer cites it or you can verify it. Label any suspected cause as a hypothesis. Propose a fix, ranked by damage to the business, worst first. Flattering errors rank high too. TRUE FACTS: [owner, what the business sells, locations, anything else a customer might check] ANSWERS: [paste each assistant's answers here, labeled]
Step three: save the whole thing, questions, answers, scores, and fix list, in one file with today's date. Work the fix list. Next month, run the identical questions again and compare.
06 ·Gotchas
The four that catch everyone.
Asking like the owner.
"Tell me about [business]'s excellent services" invites agreement. Customers ask "is [business] any good?" Test with their phrasing or you will grade yourself on questions nobody asks.
Rewriting the questions each month.
Change the questions and you changed the input, so the changed answer proves nothing. The wording freezes after run one. That is what makes month three mean something.
Trusting a flattering answer.
One business I scanned was credited with an award a direct competitor won. Errors that flatter survive longest, because nobody complains about them. Check the nice claims as hard as the bad ones.
Running it once and moving on.
A single run is a snapshot. Repeat the same check, retain the raw answers and investigate changes. Models, sources and search behavior change too, so a higher score alone does not prove which edit helped.
07 ·What's next
Most of what you find is fixable in an afternoon.
When I ran this on three real businesses, nearly every error traced to a page the business itself controls: a years-old press quote naming the previous owner, a homepage giving a machine 41 words of readable text, an empty description field the website platform generated. No budget required, just the fix list worked in order.
Before you call it done.
Check the result against the job you set out to do.
Product notes and sources · Checked October 26, 2026
Features and interface labels can vary by account. Use the current product guidance for installation, access and scheduling.