How to Check If ChatGPT Cites Your Website (Free, in Under 10 Minutes)

Someone on your team says “ChatGPT recommended a competitor”. Someone else says “we’re all over AI”. Neither claim has evidence behind it. You need a factual answer today, without a procurement cycle or a paid tool.

You can get one in under ten minutes. What follows is the exact procedure, including the mistakes that produce false confidence in both directions.

What does “ChatGPT cites your website” actually mean?

Three different things get confused here, and separating them is most of the work.

• A citation is a linked source shown with the answer. It’s the strongest signal: the engine fetched your page and attributed the content.
• A mention is your brand named in the text with no link. Valuable, but invisible to your analytics.
• A recall is the model repeating something absorbed during training. It may be years out of date and isn’t evidence that your current site is reachable.

Record all three separately. A company that’s mentioned constantly but never cited has a very different problem from one cited on a single blog post and nowhere else.

Step 1: Write ten real questions (three minutes)

Don’t use keywords. Answer engines are asked full questions, so your test must use full questions. Pull them from sales calls, support tickets and the emails prospects send before a first meeting.

A balanced ten-question set works in any sector:

• Two category questions — “what is [your service] and how does it work?”
• Two best-of questions — “best [product category] for [your customer type]”
• Two comparison questions — “[you] vs [known competitor]” and “alternatives to [category leader]”
• Two problem questions — the symptom a customer describes before they know the solution exists
• Two brand questions — “what does [your company] do?” and “is [your company] any good?”

The brand questions are the fastest accuracy check you’ll ever run. They routinely surface a discontinued product, a stale description, or a confusion with a similarly named firm.

Step 2: Run them clean (four minutes)

Open ChatGPT in a logged-out session or a private window. This matters: a logged-in account carries memory, custom instructions and chat history, all of which bias the answer toward what you’ve discussed before. Many people “discover” they’re cited when in fact they told the model about themselves an hour earlier.

Then, for each question:

1. Paste the question exactly as a customer would ask it. No brand hints in the category or best-of questions.
2. Note whether your domain appears as a linked source, whether your brand is named in the text, and which other organizations appear.
3. Ask one follow-up: “which sources did you use?” This often exposes retrieval the visible answer didn’t surface.
4. Repeat each question once more in a fresh chat. Answers vary run to run, and the difference between “cited once” and “cited both times” is real information.

If you’ve three more minutes, repeat the same ten questions in one other engine. Perplexity is a useful contrast because it cites aggressively and shows sources plainly; Google’s AI Overviews are a third useful sample. Differences between engines tell you whether the problem is your content or one engine’s retrieval preferences.

Step 3: Record it so you can compare later

A five-column table is sufficient: question, engine, cited (Y/N), mentioned (Y/N), competitors named, accuracy notes. Date the sheet. Without a dated baseline you’ll never be able to show that anything you do afterwards worked.

This is the structure our free ARIA Citation Tracker automates — same prompts, multiple engines, repeated runs, results stored so month-over-month movement is visible. Doing it by hand once is still worth it, because reading the raw answers teaches you things a dashboard will not.

How to read what you find

Four patterns cover most results.

Cited on informational questions, absent on comparisons

The most common outcome. Your explanatory content is doing its job; you’ve nothing that positions you against alternatives. Honest comparison and alternatives pages close this gap faster than anything else.

Mentioned but never linked

The model knows the brand but is not retrieving your site. Check crawler access first — robots directives and bot-blocking at the CDN or firewall level are a frequent and silently self-inflicted cause.

Competitors cited on questions you should own

Read the page that was cited. Usually it isn’t better written; it’s better structured — direct answer near the top, clear headings, specific steps or criteria, and passages that stand alone.

Cited, but the description is wrong

Treat this as urgent. An engine repeating an outdated claim about your pricing, coverage or capabilities does commercial damage. Fix the canonical source pages, then check the third-party profiles and directories the engine may be corroborating against.

What not to conclude

• Don’t conclude anything from one run. These systems are probabilistic. A single miss is noise; a miss across four runs is a finding.
• Don’t test only your brand name. Almost every company is “visible” for its own name. The value sits in unbranded questions.
• Do not assume ChatGPT represents all engines. Coverage varies substantially, and the engine your buyers use may not be the one you tested.
• Don’t confuse a plausible answer with a sourced one. If no link is shown, assume nothing was fetched.

The next ten minutes

If the test shows gaps, the highest-leverage fix is structural rather than volumetric. Take your three most commercially important pages and rewrite the opening of each section so the direct answer comes first, headings are phrased as real questions, and each passage names its subject instead of relying on pronouns. Running a draft through the free AI Citability Scorer before publishing tells you whether the structure will hold up.

If you would rather have the baseline, the prompt set and the remediation plan produced properly, that’s what our AI visibility service exists for.

Frequently asked questions

Does ChatGPT actually browse my website?

Sometimes. When a question benefits from current information it may fetch live pages; other times it answers from training data alone.

This is why a linked citation is meaningful evidence and an unsourced mention isn’t, and why you should always ask which sources were used.

Why do I get a different answer every time I ask?

Answer generation is probabilistic and retrieval isn’t fixed, so the same question can pull different sources on different runs. That’s normal, and it is the reason to measure rates across repeated runs rather than treating any single answer as the verdict.

How would I know if I am blocked from being cited?

Check your robots.txt and any bot-management rules at your CDN or firewall for the AI crawler user agents. Blocking them is a legitimate choice, but many organizations have done it accidentally through a security default and then wondered why they never appear as a source.

Is ten minutes really enough?

It’s enough to establish whether you’ve a visibility problem and roughly what kind. It’s not enough to manage the problem over time, which requires a larger prompt set, several engines and a monthly cadence. Start with the ten minutes; the case for the rest usually makes itself.