AI visibility: how to check whether AI assistants can see your business

What does AI visibility mean?

When someone asks an AI assistant “who does emergency plumbing near me” or “which bookkeeping software suits a small agency”, the assistant answers with a few names, sometimes with links. AI visibility is how often your business is among them, and how accurately it is described when it is.

It has three parts, and each can fail on its own.

• Presence: are you named at all when buyers ask the questions your business answers?
• Accuracy: when you are named, are the facts right: your services, your prices, your location, your hours?
• Citation: does the answer link to your own pages, or to a directory, a review site or a competitor’s comparison page?

AI visibility is not a ranking. There is no position one in a written answer, and the same question can produce a different answer an hour later. That makes it harder to measure than search rankings, and it is why a single screenshot of ChatGPT naming you, or not naming you, tells you very little.

Two layers: can AI read you, and does AI name you

The most useful way to think about AI visibility is to split it into two layers that behave very differently.

The first layer is where the work happens: it is the only part you can change directly. The second layer is where the result shows: it tells you whether the work is reaching the answers. Measuring only the second gives you a number you cannot act on. Measuring only the first tells you nothing about results. You need both.

Check whether AI can read your site

Start with the layer you control. Five checks catch most problems, and none of them needs a tool.

• robots.txt. Open yourdomain.com/robots.txt. Lines that say “Disallow: /“ under “User-agent: *“, OAI-SearchBot, Claude-SearchBot or PerplexityBot keep your pages out of AI search answers.
• Text in the HTML. In your browser, choose “View page source” on your home page and search for a sentence from the middle of the page. Most AI bots do not run JavaScript, so text that is not in the source is invisible to them.
• Your details in structured data. Run your home page through Google’s Rich Results Test or the Schema Markup Validator. Look for Organization or LocalBusiness with your name, address, phone and logo.
• One name, written one way. Compare how your business name is written on your site, your Google Business Profile, your social profiles and the main directories of your trade. Different spellings can read as different businesses.
• Answers on the page. Pick the three questions buyers ask most, and check that each has a direct answer on your site, with numbers, under a heading that asks it.

These five are the core of a full audit, which has 24 items in all. The complete list, with how to check each one, is in our GEO audit checklist.

Ask the assistants yourself: a simple test

The second layer can only be measured by asking. Done carefully, a manual test is a good start.

Choose the questions. Write 10 to 20 questions a buyer would ask before choosing a business like yours, in their words, without your name in them. “Best family dentist in Tucson for nervous patients” is a buyer question. “Is Desert Smile Dental good” (an invented name) is a brand question; it tells you how you are described, not whether you are found.

Cover the kinds of question. Mix questions about cost, about choosing, about process and about trust, and add your town or region to the ones where location matters.

Ask all four. Put each question to ChatGPT, Gemini, Claude and Perplexity. Use a private browser window, or log out where the assistant allows it, so your own history weighs as little as possible, and make sure web search is on where the assistant offers a choice.

Record what you see, in a simple spreadsheet: for each question and assistant, whether you were named, which other businesses were named, which pages were linked, and any fact about you that was wrong.

Repeat on the same day each week, with the same questions. One round is a snapshot; four rounds start to show a pattern.

Twenty questions across four assistants is 80 answers a round, roughly an hour of copying and pasting by hand. That is the main reason people automate it.

Why the answers change from one day to the next

If you run the same test twice, expect different answers. That is not a fault in the test; it is how these systems work.

• The models generate text with some randomness. Two runs of the same question can name different businesses, especially when several fit equally well.
• Search results change. Assistants that search the web build the answer from what they find that day.
• Users differ. Location, language, account history and saved memory can all shape the answer one person sees.
• Wording matters. “Best dentist in Tucson” and “good dentist in Tucson for kids” can produce different lists.
• Models are updated. A new model version can change answers across the board overnight.
• Apps and APIs differ. A test run through an assistant’s API, with web search on, is consistent and repeatable, but it can differ from what one person sees in the app.

The practical answer is to stop reading single answers and start reading shares. Over 20 questions, four assistants and several weeks, “named in 9 of 80 answers this week, 11 of 80 last week” is a number you can follow. “ChatGPT did not mention us this morning” is not.

What your analytics can and cannot show

Your existing tools show part of the picture.

• Referral traffic. Visits from AI assistants appear in analytics as referrals from their domains: chatgpt.com, perplexity.ai, gemini.google.com, claude.ai, copilot.microsoft.com. Many links from ChatGPT also carry the tag utm_source=chatgpt.com. Visits from the mobile apps sometimes arrive with no referrer and are counted as direct traffic, so these numbers run low.
• Google Search Console. Google counts clicks and impressions from AI Overviews and AI Mode in the overall Web search data. There is no separate AI view, so you cannot tell from Search Console alone which clicks came from an AI answer.
• Bing Webmaster Tools. Bing has added a report of how often your pages are cited in Copilot and Bing’s AI answers, with the cited pages. It covers Microsoft’s assistants, not the others.
• Server logs. Your server or CDN logs show visits from GPTBot, OAI-SearchBot, ClaudeBot, Claude-SearchBot, PerplexityBot and the rest. They prove the bots are reading your pages, not that the assistants are naming you.

None of these answers the main question, whether assistants name you when buyers ask. Traffic only counts the people who clicked, and many read the answer and never click. That is why asking the assistants directly stays part of any serious measurement.

A worked example: reading four weeks of answers

Here is how a tracking sheet might look for an invented dental clinic that asks the same 20 buyer questions to four assistants every week. All figures are hypothetical, chosen to show how to read the numbers, not what to expect.

Three lessons come out of a table like this.

Swings without changes are noise. Weeks 1 to 3 moved between 6 and 10 with nothing changed on the site itself. A single week’s number means little; the average over several weeks, here about 8 in 80, is the baseline.

Wrong facts are the quickest win. The old opening hours came from an outdated profile elsewhere, not from the clinic’s site. Accuracy problems often live off-site, and they are worth fixing before anything else, because a wrong fact can cost a customer outright.

Do not credit one week to one change. Week 4 went up after prices were added, but it is within the range of weeks without changes. Only several more weeks would show whether the change moved the average.

Turning the numbers into fixes

Measurement is only worth the time if it changes what you do next. Read the two layers together, and the next step usually follows from the pattern.

• Not named, and the site check fails. Fix access first: the robots.txt lines, the firewall, text that only JavaScript shows. Until the bots can read your pages, nothing else you write reaches the answers.
• Not named, and the site check passes. Look at the questions you are missing. Is there a page on your site that answers each one directly, with numbers? Often the answer exists only in a brochure, a PDF or someone’s head. Write it on the page that fits, under a heading that asks the question.
• Named, but with wrong facts. Find where the wrong fact comes from. Old hours, a closed location or a former phone number usually sit on a profile or directory listing you forgot, not on your site. Correct it there, and state the right fact plainly on your own pages so there is a clear source to prefer.
• Named, but the link goes elsewhere. When the answer cites a directory or a review site instead of your page, your own page probably lacks the specific answer. Add it, with the figures the cited page gives.
• Competitors named again and again. Open the pages the answers cite for them. Note what those pages state that yours do not: prices, service areas, years in business, licenses, a clear list of services. That list is your content plan, and it is built from evidence rather than guesses.

Two habits keep this honest. Change one thing at a time where you can, and note the date in your sheet, so a later rise or fall can be read against it. And give each change several weeks before judging it: none of the four assistants publishes how soon it reads a page again.

What the MWY checker measures

The MWY checker covers both layers. The free check reads your site the way the AI search bots do, as raw HTML without JavaScript and within the limits of your robots.txt, and scores 24 items in three groups: can AI read it, does AI know who you are, and can AI quote it. You get your score and the five items that matter most in a few seconds, with no sign-up. A sample report shows a full scan of a real site.

Pro adds the second layer. MWY writes 20 questions a buyer of your business would ask, from the content of your site, and you can edit them. Every week each question goes to ChatGPT, Gemini, Claude and Perplexity through their APIs with web search on, and the report shows which answers name your site, which other businesses they name, which pages they cite, and what changed since the week before. Answers through an API can differ from what one person sees in the app, which is why the report reads weeks, not single answers. See what Pro includes.

What MWY does not do is promise that an assistant will name your business. Nobody outside the AI companies controls that. What can be done is to remove what stops them from reading you, and to measure honestly whether they do.