How to Run an AI Visibility Check on Your Company

"If someone asks ChatGPT: “best app designer in London” …it doesn’t show a list of links. It just recommends a few businesses."

That's how a small business owner put it on Reddit in March 2026. They'd tested ChatGPT, Gemini and Claude with their own buyer's question, and their company wasn't named once.

Most companies have never run that test. It's not hard, and you don't need a paid tool to do it. You need a list of questions your buyers really ask, an afternoon, and a spreadsheet.

This is the hands-on version. I'll go through what to ask, which AI tools to use, how many times to ask, what to score, which free reports from Google and Bing to add, and what to do when the answer is wrong. If you want the wider picture first, read AI visibility for B2B brands, then come back here.

What an AI Visibility Check Is, and What It Can Tell You

An AI visibility check is a structured test of how AI assistants describe your company and category when buyers ask them questions. It's the first thing to do before you spend anything on AI search work, because it tells you whether you have a problem at all.

It answers five questions. Are you mentioned? Are you recommended, or just listed? Is what's said about you accurate? Which competitors are named instead? And which web pages does the assistant cite as its sources?

It's worth separating those, because they fail in different ways. Ahrefs makes a useful distinction in its AEO course: you can be cited with a link, mentioned without a link, or not visible at all, and there's a fourth case where you get recommended without being named. This short lesson shows real examples of each across AI Overviews, ChatGPT, Perplexity and Gemini:

Play Video (Loads YouTube)

Who should run it matters more than people expect. You need someone who knows the company's facts well enough to spot a subtle error: a price that's a year old, a service you quietly dropped, a client sector you've never worked in. That's usually a founder, a sales lead or a senior marketer. Hand it to someone new and they'll mark wrong answers as accurate because they sound plausible.

What a check can't tell you is why a model said what it said, or what it'll say next week. It's a sample. A good one, run the same way each time, is still the most useful starting point you'll get.

Before You Start: Questions, Competitors and Settings

Spend most of your preparation time on the questions. The rest is admin.

Write 10 to 15 buyer questions

Pull them from places where buyers speak in their own words: sales call notes, enquiry emails, your Search Console queries, People Also Ask boxes. Then sort them into four types, and make sure you cover the people in a buying group who'd ask them.

Keep brand questions to a minority. An assistant asked about you by name will nearly always mention you, so too many of them will flatter the result. The category and problem questions are where the real test is.

If you'd like a ready-made starting list, the free AI Visibility Self-Audit worksheet has twelve buyer prompts, a scoring grid and a fix list you can date and re-run.

Name three to five competitors

Pick the companies you actually lose deals to, not the biggest names in your industry. You'll record every brand that appears, but these are the ones you'll compare yourself against.

Fix the tools and settings

Use the tools your buyers use. For most B2B companies that's ChatGPT, Gemini, Claude, Perplexity and Google (AI Overviews or AI Mode). Use a fresh chat for every question, so earlier messages don't colour the answer. Write down whether you're logged in, which model or mode is selected, and your country. Then keep all of that the same every time you re-run the check. Change a setting and you can't compare the results.

Run the Check, Step by Step

Ask every question three times in every tool, and record each answer word for word. With 12 questions and five tools, that's 180 answers. It sounds like a lot. Most of the time goes on copying and pasting, and it's worth it.

1. Open a fresh chat for each question. In ChatGPT, Gemini, Claude and Perplexity, start a new conversation every time. In Google, run the search and note whether an AI Overview appeared at all.
2. Ask the question exactly as written. Don't add your company name or nudge the model. You're trying to see what a buyer sees.
3. Copy the whole answer into your sheet. Add the date, the tool, the setting and the run number. Paste the text in full, because you'll want to re-read it when you score.
4. Record the sources. Note every cited page, and whether any of them are yours.
5. Repeat twice more. Same question, new chat. Then move to the next question.

Set the sheet up before you start, with one row per answer. These columns are enough: date, tool, setting, question, question type, run number, the full answer text, every brand named (in the order named), whether you're mentioned, whether you're recommended, accuracy, and the cited sources. It looks like overkill for the first round. It's what makes the second round comparable.

Each tool shows its sources differently, so it's worth knowing where to look. OpenAI's help page says ChatGPT search responses "may include citations", and that you can select Sources "to view cited sources and other relevant links." Perplexity says "Each answer includes numbered citations linking to the original sources." Google describes AI Overviews as "an AI-generated snapshot with key information and links to dig deeper."

Why three runs? Because answers change. In SparkToro's January 2026 study, 600 volunteers ran 12 prompts through ChatGPT, Claude and Google's AI 2,961 times, and got the same list of brands less than once in 100 runs. Three runs won't make your result precise. It will stop you mistaking one odd answer for the truth. Rand Fishkin, SparkToro's co-founder, has been blunt about it:

If you'd like to watch someone do a manual audit like this end to end, Backlinko's walkthrough is a good one. It covers tracking mentions and citations across ChatGPT, Google AI Overviews, Claude and Perplexity, spotting the sources that keep coming up in your niche, and turning the findings into a 90-day plan:

Play Video (Loads YouTube)

What to Score in Each Answer

Score every answer on the same five points. Paid tracking tools use much the same structure. You're just doing it by hand at a smaller scale.

Don't skip the accuracy column because the answers sound sure of themselves. Google's own help page for AI Overviews says they "can and will make mistakes". Check every factual statement about you against your own records: what you sell, who for, where you're based, your prices and your founding date.

Once you've scored everything, add up three numbers: the share of answers that mention you, the share that recommend you, and the share of your mentions that are accurate. That's your baseline. If you want to turn it into share of voice against named competitors, my guide to measuring AI share of voice has the formulas and a scorecard template.

Add the Free Data From Google and Bing

Google and Microsoft now both report some AI data for free, and it's worth adding to your check. It tells you which of your pages are being used as sources, which your manual test only samples.

In Google Search Console, look for the generative AI performance reports. They launched in June 2026 and reached every site worldwide on 31 August. They show impressions, pages, countries, devices and dates for your appearances in AI Overviews, AI Mode and Discover.

In Bing Webmaster Tools, open AI Performance, which launched in February 2026. It covers Copilot, AI summaries in Bing and some partner integrations. Glenn Gabe, the SEO consultant, summed up what you get on launch day:

Grounding queries are the most useful part for this check. They show the phrases the AI searched for when it went looking for content to cite. Compare them with your buyer questions. If the grounding queries use words your pages never use, that's a gap you can close.

Neither report tells you whether your brand was named, or what was said about you. That's still the job of your manual check.

Reading the Results

Look for patterns across runs and tools, not single answers. Three patterns cover most of what you'll find.

You're missing from category and problem questions. This is the most common result, and the hardest to fix quickly. It usually means AI tools haven't seen you described often enough, in the words buyers use, on sites other than your own. The small business owner from the top of this guide hit exactly this:

Their post goes on to say the same competitors came up every time. That's a useful clue in itself. Look at the sources cited in those answers, and you'll usually find the directories, comparison articles and review sites where your competitors are described and you aren't.

You're mentioned, but described wrongly. Old prices, a product you've retired, the wrong category. This is usually the quickest to fix, because the cause is often on a page you control or a profile you can edit.

You're mentioned, but someone else is recommended. Read those answers closely. The reason given for the recommendation ("better for smaller teams", "more established") tells you what the model believes about each option, and where that belief probably came from.

Your pages are cited, but you're not named. This one surprises people. An assistant uses your article as a source for a general explanation, then recommends three competitors. It means your content is good enough to read but doesn't make clear that you're a provider of the thing being explained. Say plainly, on those pages, what you do and who it's for.

Write a one-line verdict for each question type. Something like "Category: missing in 4 of 5 tools. Brand: accurate in ChatGPT, outdated price in Gemini." That's the summary you'll share internally, and the thing you'll compare against next quarter.

What the Research Says Gets Content Cited

The most rigorous public evidence on what makes content more visible in AI answers comes from GEO: Generative Engine Optimization, by researchers from Princeton University and IIT Delhi, presented at KDD 2024. They built GEO-bench, a benchmark of 10,000 queries, and tested nine ways of rewriting a web page to see which made a generative engine show more of it in its answer.

Here's how the main methods scored on the paper's headline measure, which combines how much of a source's text appears in the answer and how early it's cited:

The authors sum it up themselves: "The best methods improve upon baseline by 41% and 28%" on their two measures. Adding quotations and statistics helped most. Keyword stuffing, the old SEO habit, scored below doing nothing at all. They also re-ran the test on Perplexity, a live AI search engine, and saw the same direction: quotations and statistics up, keyword stuffing down.

I'd treat these as lab results. The researchers chose the queries, and your market isn't in them. But the practical lesson is sound, and it matches common sense. A page that says "we verify claims against named public registers, and publish our method" gives a model something specific to quote. A page that says "we're the leading verification platform" gives it nothing.

It's also worth being clear about what the study didn't test. It measured how much of a page's text turned up in an answer. It didn't measure whether a company got recommended for a buying question, which is what most B2B firms care about. Recommendations seem to depend far more on how you're described across the web than on any single page. So use the findings to improve the pages you own, and don't expect them to fix a "missing" result on their own.

That's also why I'd focus on facts that others can check. If you want more on that, I've written about getting proof claims on B2B websites right.

What to Do About a Wrong or Missing Answer

There's no single "report an error" button that corrects every AI assistant. Each one draws on a mix of training data and live web retrieval, and neither updates instantly. Four actions are worth taking, roughly in order of effort.

1. State the right fact clearly on your own site. Specific, dated and in plain words, on a page that's easy to find. A vague "about us" paragraph gives a model very little to work with.
2. Make every third-party profile match. LinkedIn, Companies House, Crunchbase, G2, industry directories, partner pages. An inconsistency between your site and a public profile is exactly the kind of thing that produces a wrong answer.
3. Check that AI crawlers can reach the page. If your robots.txt or firewall blocks a search crawler such as OAI-SearchBot or Claude-SearchBot, the corrected page may never be read. My B2B guide, linked at the top of this page, has a full crawler checklist.
4. Earn mentions where the cited sources are. If every answer cites the same two comparison sites, get yourself fairly described on them. This is the slow part, and usually the one that moves "missing" to "mentioned".

Then be patient. Corrections show up over weeks to months, as AI tools recrawl pages and refresh their sources. Re-checking every day won't speed it up.

How Often to Re-Run the Check

Quarterly is right for most established companies. Monthly makes sense if your category is moving fast or you're actively fixing things and want to see the effect.

Also re-run it straight after any of these: a rename, a new product, a pricing change, a big website rebuild, or a new competitor in your space. Those are the moments when AI descriptions are most likely to be stale.

When you report it internally, keep it short. One line per question type, the three headline numbers against last quarter, the worst wrong statement you found, and the fixes you're making. Leadership doesn't need 180 answers. They need to know whether buyers asking an AI about your category are hearing about you, and whether what they hear is true.

Keep the questions, tools, settings and number of runs identical each time. A check is only useful when you can compare it with the last one. If you want to add questions, add them as a separate group so the original set stays comparable.

Where ThriveFinity Fits

If you'd rather not run the check by hand, we can do it for you as part of our AI search optimisation (GEO) service.

• AI Visibility Signal (free, under an hour): 12 buyer questions across GPT (OpenAI), Claude, Gemini and Perplexity, scored 0 to 100 and graded A to E. It's AI-only and unsigned, and the full question list is published.
• AI Visibility Diagnostic (£349, 3 working days): the diagnosis part of the Retrofit on its own, fully credited to the Retrofit within 30 days.
• AI Search Readiness Retrofit (£1,950, 10 working days): 100 to 150 buyer questions across five AI tools, your 10 most valuable pages fixed, and a re-test of the same questions at day 30. Signed off by a person.

You can see the format of our reports in the anonymised sample AI visibility report. If you're weighing us against other options, the GEO agency buyer's guide has the questions I'd ask anyone, including us.

We're not the right choice if you need continuous, daily monitoring across hundreds of prompts. A dedicated tracking platform is built for that. And if your check shows you're named, recommended and described correctly, you might not need anyone. Just re-run it next quarter.

Pranav Unni

Founder · ThriveFinity Connect on LinkedIn →

Pranav Unni is the founder and lead verifier of ThriveFinity. He reads and signs every paid deliverable personally, and writes about go-to-market decisions for established B2B companies.