How to run an AI search audit on your own company

Say you ask ChatGPT who replaces windows in Round Rock, on your phone in the truck, and your company isn't on the list. At the office you ask again on the laptop, and the list has changed.

Neither list is a verdict. A sheet of dated answers is something you can work from, and it needs no paid tool. Here's the one I'd fill in, as of October 2026. If the term is new, start with what an AI visibility audit is. For how assistants choose local companies in the first place, read our guide to AI search for home service companies.

What do you need before you start?

A notebook or a spreadsheet, an afternoon and a fact card. Write the card first: your company name as it's on the truck, your phone number, hours, services and towns. Every answer and listing gets compared with it.

Which ten questions should go on the sheet?

Write them the way a homeowner would type them: the service, the town and, for some, the situation. Take the words from your phone calls and estimate notes.

Eight leave your name out, because a name in the question hints the answer, and two name your company. Our open benchmark also mixes unbranded and branded prompts.

Swap in your service and towns, number the ten, and keep the wording the same once you start.

How do you ask the questions and record the answers?

Ask each question in a new chat, so the answers before it aren't part of the conversation. Do it in each assistant you pick, and write one row per answer. Here's what each help page says:

• ChatGPT may search on its own, and answers that use search may include citations.^[1]
• Claude searches when a question benefits from current information, and its responses then include citations.^[2]
• Perplexity numbers its citations.^[3]

If an answer shows no sources, turn on search and ask again, and note that. Claude's page says to include "Search the web" in the prompt if you want to ensure it searches,^[2] so if you add that, add it to every Claude question. Gemini, Microsoft Copilot and Google's AI Overviews and AI Mode can join the sheet by hand.

Go signed out where the assistant allows it. OpenAI says people who aren't signed in can use ChatGPT's web search, and that when memory is on, saved memories may be used to rewrite a search.^[1] Put the town in every question, since ChatGPT may use an approximate location from your IP address.^[1]

Our benchmark says the same prompt can return a different answer minutes later, and Google says AI Overviews and AI Mode may use different models and techniques, so they can differ from each other.^[4] One run proves little, so every row carries a date.

On Google, AI Overviews often don't trigger,^[4] so write none shown when there isn't one. How to check Google's AI answers covers that side.

What should you check when an assistant doesn't name you?

Check your own site and profile first. In one study of local AI answers, business websites and Google Business Profiles supplied much of what was cited:

1. Crawler access. Open yourdomain.com/robots.txt, a public file that tells crawlers, the programs that read your pages, where not to go. Look for a rule that blocks every bot, or just OAI-SearchBot, Claude-SearchBot or PerplexityBot, the search crawlers named by OpenAI, Anthropic and Perplexity.^[5][6][7] Our two-minute check shows how.
2. Your details. Lay your fact card beside your site, your Google profile and the listings the answers cite. Google's AI features page counts text content and an up-to-date Business Profile among the basics.^[4]
3. The cited pages. Open the sources an answer cites. OpenAI says to open a cited source to check that it supports the answer.^[1] Then see whether any mentions you, and note which are directories, "best of" lists or competitors' service pages.
4. One page per question. Does a page on your site answer the exact question, with the service and town in plain words? Our guide to writing a service page covers it.
5. Your reviews. Read the last ten. Do any say what the job was?

Can ChatGPT audit your website for you?

It can help you write the audit, but I wouldn't let it do the audit. OpenAI says ChatGPT can sound confident when it's wrong, and that its mistakes include made-up citations and references to sources that don't exist.^[8] It may also be unable to get information from a website because of technical issues, paywalls or robots.txt.^[8] Without search, its answers come from training.^[8]

So an "audit" can include findings about pages it couldn't read. Keep a finding only if you can see it on the page yourself.

I'd use it to draft your ten questions from call notes and to turn a filled-in sheet into a fix list. OpenAI's own advice is to use ChatGPT as a first draft, not a final source.^[8] Keep customer names and phone numbers out of what you paste.

What can a free AI visibility checker tell you?

A checker that asks the assistants can show you what they answered on the day it asked. The reasons you were left out are yours to check, with the five checks above. For Google's results, Google says third-party tools don't have access to its internal ranking data.^[9]

Before you trust a score, ask three things:

• Which assistants and questions did it use, or is it one template with your town swapped in?
• On what date, and did it ask more than once?
• Does it show the full answer text and the sources, so you can check the score yourself?

Our own Free AI visibility check is a different job: a review of your website and Google profile by hand, with your first fixes, usually within two business days. It isn't this sheet done for you. What an audit costs is in its own post.

How do you turn the sheet into a fix list?

Count, don't score: for each assistant, how many of the ten questions named you correctly, how many named you with a mistake and how many didn't. A result might read 3 of 10 in one assistant and 0 of 10 in another. Cost and HOA questions may name no company, so mark those "no names" rather than as a miss.

Then count the other names. The companies that come back most are the assistants' idea of your peers, and our benchmark says that list is usually more useful than your own number.

Sort what you found into three piles: a block (a crawler turned away), a mismatch (a detail that's wrong or disagrees) and a gap (no page, no listing, no mention). Start with the blocks. For a mismatch, trace the wrong detail. For a gap, compare your evidence with the companies named.

Pick the three cheapest fixes and make them this month. Ask the same ten again in a month and compare the two sheets. The monthly routine covers the rest.

Where does a do-it-yourself AI search audit stop?

It stops at what you can see. Ten questions on one day are a dated sample, not a ranking. OpenAI says placement in ChatGPT search isn't guaranteed,^[1] and Google says indexing and serving isn't guaranteed.^[4] The sheet shows what to fix, not what will happen.

Several of these pages show no exact date, so the sources list gives the day each page was opened: October 15, 2026.

If the sheet shows gaps and you want a working session on what to fix first, the Expert Action Plan is $149. It's the complete report from our AI search visibility audit, which runs the buyer questions, plus a 30-minute consultation, for owners who'll make the fixes themselves. Our AI search optimization plans run the same buyer questions across the AI assistants we test (currently ChatGPT, Claude and Perplexity) every month.