How to Test If Your Content Is Cited in ChatGPT, Perplexity, and Google AI Overviews

By Mark Buraga, Independent SEO Consultant at Growth Engine PH Last updated: 25 June 2026

You can check your Google ranking any time you want. You have almost certainly never once checked whether ChatGPT, Perplexity, or Google AI Overviews actually use your content. Same kind of question, no dashboard required, and you can answer it in about twenty minutes.

That gap matters more than it sounds. AI visibility has become the thing everyone has an opinion about and almost nobody has measured. People believe they are cited, or fear they are invisible, on no evidence at all. The fix is a test you run yourself, on your own content, on a cadence. This is that test: the exact prompts to fire, how to read each answer, how to log it, and how often to repeat it.

Cited, mentioned, or absent: the only three outcomes

Every AI-visibility test resolves to one of three outcomes, cited as a linked source, mentioned by name in the answer body, or absent entirely, and they are earned by different work.

Before you run anything, fix the read in your head, because the whole test depends on it. When you ask an AI engine a question and it answers, your content lands in exactly one of three places.

Cited means the engine attributes part of its answer to you with a linked source. That is the sidebar link in Google AI Overviews, the numbered source in Perplexity, the reference list in ChatGPT. A citation is earned structurally: the engine could reach your page, parse it, and lift a clean passage. Mentioned means your brand or name appears inside the synthesized answer text, with or without a link. That is earned differently, through consensus, when enough credible places say the same thing about you that the engine treats it as established. Absent means neither. You are not in the answer at all.

These are not the same outcome with different labels. They run on different mechanisms, which is why you can have one without the other, and why collapsing them into “AI visibility” hides the information you need. The acronyms piece maps the distinction onto the surfaces it shows up on. Hold the three-way read through every step below.

Why your ranking can no longer tell you this

Only 38 percent of pages cited in AI Overviews still rank in the top 10, so position has come apart from citation and a separate test is now the only way to know.

The reason you have to run a dedicated test, rather than just glancing at your rankings, is that the two have decoupled. Only 38 percent of the pages cited in Google AI Overviews still rank in the top 10 organic results, down from around 76 percent eight months earlier (Ahrefs). Position used to predict the outcome. Rank well, get the clicks. That link has loosened.

So a page can sit at position three and never be cited, while a page nobody can find on the results list gets quoted in the answer. Your rank tracker cannot see this, because it was built to measure a different thing. If you want to know whether the engines use your content, you have to ask the engines directly. The 6-Block GEO Diagnostic explains why citation is decided where it is; this post is how you confirm where you stand.

Build your fixed prompt set first

A test is only repeatable if the prompts never change between runs, so write your prompt set once, freeze it, and reuse the same list every time.

The single most common mistake is testing on a whim with whatever query comes to mind, then drawing conclusions from results that were never comparable. Fix that before you run anything. Write a set of five to eight prompts and freeze the list. You reuse the exact same prompts every time you test, so the only thing changing between runs is the engines, not your questions.

Phrase them the way a real buyer would ask, in full sentences, not stripped-down keywords. An engine answers natural questions, so test with natural questions. Build the set across three kinds:

• The money queries. The two or three questions a ready-to-buy customer actually types. These matter most, because being absent here is the expensive kind of absent.
• The category queries. Broader questions about your topic where you would expect a credible source to be named. These tell you whether the engine sees you as an authority at all.
• A branded query. Your own name or company. If you are not even cited on your own brand, that is a Block 3 entity problem, not a content problem, and it is worth knowing first.

Write them down somewhere you will keep. The prompt set is the instrument. Change it casually and you have thrown away your ability to compare this month to last.

Test 1: ChatGPT

In ChatGPT, run each prompt with web search enabled, then read the answer body for your name and the source list for your link, recording cited, mentioned, or absent.

Start with web search switched on, because without it the model answers from memory rather than the live web, and you want to know what the engine retrieves today. Run your first frozen prompt. When the answer comes back, do two separate reads.

First, read the answer text itself. Is your brand or your name written into the response? If yes, that is a mention, whether or not it is linked. Second, look at the sources the answer cites, usually shown as links or a reference list. Is your page in there? If yes, that is a citation. Record one of the three outcomes for that prompt, then move to the next. A small habit that pays off: start a fresh chat for each prompt so an earlier answer does not bleed context into the next one. Work through the whole frozen set the same way.

Test 2: Perplexity

Perplexity surfaces its sources inline and in a list, so a citation is easy to confirm, but you still read the prose separately for an unlinked mention.

Perplexity makes the citation half of the read almost effortless, because it shows numbered sources inline next to the claims they support and gathers them in a list. Run the same frozen prompt and scan that source set for your domain. If it is there, you are cited, and you can usually see exactly which claim you were cited for, which is genuinely useful intelligence about what the engine valued.

Do not stop at the sources, though. Read the written answer on its own. Your name can appear in the prose as a mention without your link making the source list, and that is a meaningfully different result. Cited tells you the engine could retrieve and lift your page. Mentioned without a citation tells you your reputation precedes your retrievability. Log the outcome for each prompt and continue.

Test 3: Google AI Overviews

For Google AI Overviews, run the query logged out or in an incognito window, expand the Overview, and check both the synthesized text and the cited links.

Google is the one with a wrinkle: not every query triggers an AI Overview, and the ones that do are personalized by your account history. So run these in an incognito or private window while logged out, which strips most of the personalization and shows you something closer to what a stranger sees. Enter your frozen query and wait for the Overview to render at the top.

When it appears, expand it fully, including the carousel or panel of cited sources it shows alongside the answer. Same two reads as before. Is your name written into the synthesized paragraph, a mention? Is your page among the linked sources, a citation? If no Overview appears for a given query at all, note that too, because an absent Overview is information: the engine did not think the query warranted a synthesized answer, which changes how you weight that prompt. Record each result.

Log every result the same way

Record each test as one row, engine, prompt, date, and outcome, so the log becomes a trend you can read rather than a pile of screenshots.

A single round of testing is a snapshot. The value is in the trend, and you only get a trend if you log consistently. Keep it boring and keep it small. One row per engine, per prompt, per date, with the outcome in plain words.

That is the whole template. A note app or a single sheet is plenty. Screenshot only the results worth keeping, the wins you want to prove and the losses you want to fix. The log is what turns “I think we got cited that one time” into a record you can act on three months from now.

How often to test, and what the pattern tells you

Test on a fixed cadence, monthly for most sites and weekly when you are actively building a topic, because AI answers vary run to run and only a trend is trustworthy.

Here is the fair objection, and it is worth taking seriously: AI answers wobble. Ask the same question twice and you can get two different answers, with different sources. So does a single test prove anything? On its own, no. That is exactly why the method is a logged cadence and not a one-time glance. You would never judge your rankings off a single refresh on a single device, and you should not judge your AI visibility off a single answer either. You are reading a trend across repeated runs, and the trend is what is trustworthy.

For most sites, monthly is the right rhythm. When you are actively building out a topic, publishing into it and working its off-page surface, move to weekly so you can see the change while it is happening. Then read the pattern, not the single result:

• Cited but not mentioned: the engine can retrieve and lift you, but the market has not yet built enough consensus to name you in the prose. A reputation gap.
• Mentioned but not cited: your name carries, but the engine could not cleanly retrieve a passage from your page. A retrievability gap, usually on-page and usually fixable fast.
• Absent across runs: neither signal is firing yet on that query. A both-gap, and the place to start work.

What to do with what the test shows you

Cited but not mentioned is a consensus gap, mentioned but not cited is a retrievability gap, and absent is a both-gap, and each points at a different fix.

The test is not a verdict. It is the first step of the fix, because it localizes the failure. Once you can name the gap, you know where to work, and there are only a handful of places to look. Who gets cited breaks down the decision the engine is actually making, and the diagnostic walks the six specific blocks where citation is won or lost.

If you are mentioned but not cited, the work is on your page: clean retrieval, an answer the engine can lift, a verifiable entity behind the site. Four of the six blocks live there, and you can usually move them inside a week. If you are cited but not mentioned, the work is off your page, in the slower business of building enough credible third-party validation that the engine names you by default. If you are absent, you are doing both. The test did not just tell you that you have a problem. It told you which problem, which is the only useful kind of answer.

The thing to take from all of this is small and durable: you do not have an opinion about your AI visibility until you have run the test. Until then you have a hope. Twenty minutes and a frozen prompt set will tell you more than a year of reading about AI search, and the only way to know whether you are cited, mentioned, or absent on the query that wins you work is to go and look.

If the test surfaces a gap you cannot close on your own, that is the work Engine and Engine Pro do on a cadence, so the read happens every week without you running it by hand. If your AI visibility is not matching the effort you have put in, let’s talk.

FAQs

How do I know if ChatGPT is using my website? Run your real buyer questions in ChatGPT with web search enabled, then do two reads of the answer. Check the response text for your name, which is a mention, and check the cited sources for your link, which is a citation. Record one of three outcomes for each question: cited, mentioned, or absent. Repeat on a cadence, because a single answer is not reliable.

What is the difference between being cited and being mentioned in AI search? A citation is a linked source the engine attributes part of its answer to, earned by structural retrievability. A mention is your brand named inside the synthesized answer text, with or without a link, earned by market consensus. They run on different mechanisms, so you can be cited without being mentioned, or mentioned without being cited, and each gap points at a different fix.

How can I check if I am in Google AI Overviews? Run the query in an incognito or private window while logged out to reduce personalization, then expand the Overview when it appears. Read the synthesized paragraph for your name, a mention, and check the cited sources for your link, a citation. Note that not every query triggers an Overview, and an absent Overview is itself information about how the engine treats that query.

Why does the AI answer change every time I ask? AI answers are non-deterministic, so the same prompt can return different text and different sources across runs. That is why a single check is unreliable and the method here is a logged cadence rather than a one-time look. You are measuring a trend across repeated tests, the same way you would never judge a ranking off one refresh.

Do I need a paid tool to track AI citations? No. The manual test costs nothing and gives you the read that matters most, cited versus mentioned versus absent, on the queries you choose. A paid tracker helps once you are testing at scale, but it answers a narrower question and often hides the citation-versus-mention distinction. Run it by hand first so you understand what any dashboard is actually counting.

How often should I test my AI visibility? Monthly for most sites. Move to weekly when you are actively building out a topic and working its off-page surface, so you can see the change as it happens. Keep the prompt set frozen between runs so each test is comparable, and read the trend rather than any single result.

Related reading: What is SERP optimization and the 6-Block GEO Diagnostic.