How to Check Whether AI Search Is Citing Your Business

You cannot improve AI search visibility until you can measure it, and most businesses have never checked. This is the method we run monthly for clients: twenty prompts, three assistants, thirty minutes, no paid tools required.

If you want the underlying theory first, our GEO hub covers what generative engine optimisation is and how AI systems decide what to cite. This article is the practical companion: how to find out where you actually stand, and what to do about it.

There is no ranking to check, and that changes everything

With Google you can search a keyword and see your position. It is the same for everyone in the same place at the same time, near enough, and you can track it.

Assistants do not work that way. They generate an answer rather than retrieving a fixed list, so the output varies between sessions, accounts, devices and locations. Ask the same question twice and you can get two different sets of businesses named.

This has one important consequence: a single check proves nothing. Asking ChatGPT once, seeing your name, and concluding that AI search is working is exactly as unreliable as asking once, not seeing your name, and concluding it is broken. What you can measure is a rate across a fixed set of questions, tracked over months.

That is what the method below produces.

The three layers of measurement

Start with layer one. It is free, it is the most directly useful, and it is the only one that tells you what a real buyer would see.

Layer 1: the 20-prompt tracking sheet

Build a list of twenty questions a genuine prospect might ask an assistant, then run the same twenty every month against ChatGPT, Gemini and Google AI Overviews.

The rule that matters most: keep the wording identical every month. Changing a prompt resets your ability to compare, which is the entire point.

Spread the twenty across five types. Here is a filled example for a fictional clinic, Sharma Dental Care in Rajouri Garden.

Two notes on building your own. Take the prompts from questions clients actually ask on calls, not from a keyword tool, because people type differently into an assistant than into Google. And include the brand prompts even though they feel like cheating. If an assistant cannot describe your business accurately when asked by name, nothing else on the list will go well.

How to score an answer

Three outcomes per prompt, per assistant. Keep it this simple or you will stop doing it.

• Cited — you are named and linked
• Mentioned — you are named but not linked
• Absent — no mention

Sixty cells a month, twenty prompts across three assistants. Record the date and the competitors named alongside you, which is often the more useful column.

From that you get a simple share of voice: your mentions divided by all business mentions across the sheet, times one hundred. Track that number monthly. Movement in it is the signal; any individual answer is noise.

Practical tips. Use a logged-out or incognito session so your own history does not flatter the result. Set your location to where your customers are, not where you are sitting. And run all three assistants in one sitting, so the comparison is fair.

Layer 2: Search Console's generative AI report

In June 2026 Google added a generative AI performance report to Search Console. It is the first official data any of us have had on this, and it is worth knowing exactly what it does and does not contain.

What it shows: impressions of your pages in AI Overviews and AI Mode, filterable by page, country, device and date. Google says it expects that list of features to grow.

What it does not show: clicks, click-through rate or position. Impressions only. It also excludes Search Labs experiments, inherits the usual 1,000-row limit, and recent data is preliminary and can change.

The catch: it is on limited rollout. Not every property has it, and a site needs enough impressions in AI features to appear at all. If you do not see it, that is not a fault on your site.

Because there is no click data, treat this report as a visibility signal rather than a traffic report. It answers "are my pages being surfaced in Google's AI answers", not "did anyone come to my site as a result". Pair it with the manual sheet, which covers the assistants Google does not report on at all.

Layer 3: when a paid tool earns its place

A category of AI visibility monitoring tools now exists, running prompt sets automatically across ChatGPT, Gemini, Perplexity and others, then reporting share of voice, citation counts and sentiment. Entry pricing starts around thirty US dollars a month and rises steeply from there.

For a single-location business tracking twenty prompts, a spreadsheet does the same job for free. A tool earns its place when you have several locations, a long prompt list, competitors you need to benchmark against, or you want daily rather than monthly data.

One distinction worth understanding before you buy. AI search monitoring tracks what users actually see in AI products, including citations and links. LLM monitoring probes what a model knows from training data and returns no citations. The second is interesting; the first is what affects your enquiries.

What to change when a prompt comes back Absent

Nearly always, the page exists and is simply not quotable. Assistants lift self-contained passages that answer a question directly. Marketing prose does not survive that treatment.

Here is the same clinic described two ways.

Before, unquotable:

"We are one of the leading providers of comprehensive dental solutions in the region, with a dedicated team committed to excellence and a patient-first philosophy that has made us a trusted name for families."

After, quotable:

"Sharma Dental Care is a two-chair dental clinic in Rajouri Garden, West Delhi, open since 2014. It offers root canals, implants, braces and paediatric dentistry. A single root canal costs ₹4,000 to ₹7,000 and takes one or two visits."

The second version names the business, the locality and the city, states what it does, and attaches real numbers. Lift it out of the page and it still makes sense, which is precisely what an assistant needs.

The fixes that move Absent to Cited, in the order we apply them:

1. Put a direct 40 to 60 word answer immediately under each heading, before any preamble
2. Write headings as the questions people actually ask
3. Name your business, service and city explicitly instead of writing "we" and "our team"
4. Replace vague claims with specific numbers, prices, timelines and dates
5. Add a comparison table, which gets extracted disproportionately often
6. Add FAQPage, Organisation and LocalBusiness schema
7. Show a visible author and a last-updated date
8. Make sure the content is in the HTML, not rendered by JavaScript after load

Off the site, the things that feed assistant answers are the same signals that have always built reputation: consistent listings, genuine third-party mentions, industry directories that real people use, and discussion threads where your business comes up. Our page on improving ChatGPT visibility goes further into what is and is not controllable there.

How the three surfaces differ

Local questions lean heavily on your Business Profile, which is why the profile checklist is the cheapest AI-visibility work available to most Indian businesses.

What this method cannot tell you

Being honest about the ceiling here matters more than usual, because this is the part of the industry where overselling is most common.

You cannot see how many people asked an assistant about your category, so there is no volume figure. You cannot attribute an enquiry to an assistant reliably, because most arrive as direct or branded traffic afterwards. You cannot control the output, and nobody can. And your sample of twenty prompts is small, so treat single-month movements with suspicion and look at three-month trends.

What you can do is establish whether you are present or absent for the questions that matter, watch that ratio move, and know which pages to fix. That is a great deal better than guessing, which is where most businesses are.

Anyone offering guaranteed AI citations is selling something that does not exist.

Frequently asked questions

Can I see AI Overview traffic in Google Search Console?

Partly. Search Console has a generative AI performance report covering AI Overviews and AI Mode, but it shows impressions only, with no clicks, click-through rate or position. It is also on limited rollout, so not every property has it yet.

Why do I get a different answer from ChatGPT every time I ask the same question?

Assistants generate answers rather than retrieving a fixed ranking, so output varies between sessions, accounts and locations. This is why one lucky check proves nothing. Track the same prompt list monthly and read the trend, not any single answer.

How many prompts should I track to measure AI search visibility?

Twenty is enough for most small businesses. Spread them across category discovery, comparison, problem-led, recommendation and brand-name questions. Keep the wording identical every month, because changing a prompt resets your ability to compare.

Do I need a paid tool to track AI search visibility?

Not to start. A twenty-prompt sheet run manually takes about thirty minutes a month and costs nothing. Paid monitoring earns its place when you track many prompts, several competitors or multiple locations, and want daily rather than monthly data.

What should I change if an assistant never mentions my business?

Usually the page exists but is not quotable. Add a direct forty to sixty word answer under each heading, name your business and city explicitly instead of writing "we" and "our team", and replace vague claims with specific numbers and dates.

Where to start this week

Build the twenty prompts, run them once, and save the sheet. That first run is your baseline, and it usually takes under an hour. Then check whether your Search Console property has the generative AI report yet.

If most of your sheet comes back Absent, the fix is on your pages rather than in a tool, and the eight changes above are the order to work through. Our on-page work covers most of them, and the GEO hub explains the reasoning behind each.