How Do I Know If ChatGPT or Google AI Overviews Are Citing My Site?

Updated 19 October 2026 · 10 min read · By Matthew Bertram

Both halves have a catch. Google's report won't say which searches you appeared for. Bing shows only a sample of them. A single hand check proves almost nothing, because the same question asked twice rarely brings back the exact same set of sources.

What do Google and Bing report?

Each now has a report built for AI answers. They count different things. Read them side by side, not as two views of one number.

Google counts exposure. Its help page for the report defines impressions as "how many times links to your site were shown to a user in a generative AI feature on Google Search." The page doesn't mention clicks. Bing's AI Performance report counts each time one of your pages was shown as a source, which is closer to what most people mean by a citation.

What can't Search Console tell you?

Which searches put you there. The numbers break down by page, country, date and device, but there's no query view, so you'll see that a page appeared in AI answers without seeing the questions that triggered it. The report also leaves out Google's experiments in Search Labs.

If the report is missing from your account, Google says that "usually happens when your site hasn't received enough impressions" in its AI features. The same page names one other cause: your site may be excluded through the Search generative AI control that Search Console has offered every site since August 31, 2026. A site set to exclude gets no impressions or traffic from these features at all. Include is the default. Check that nobody changed it.

Why open Bing's report if your buyers use Google?

Because it shows you the searches. Bing calls them grounding queries: the phrases Microsoft's AI searched for when it pulled your pages into an answer. No other engine we know of shows site owners that link between its retrieval searches and their citations. Microsoft is careful to add that "The data shown represents a sample of overall citation activity," so treat it as a sample, not a full log.

In June 2026 Microsoft added four views, all labeled as previews. Intents sorts grounding queries by purpose. Commercial and local are two of Bing's categories. Topics groups them by theme. Citation Share is "the percentage of citations attributed to your site out of all citations shown across all sites for that same grounding query," the closest thing to a comparison we've seen an engine publish, though Bing says it names no competing sites and is not "a competitive scoreboard." Compare lays one period over another.

Keep its reach in mind. It covers Copilot and Bing's AI summaries, plus some partner services. It tells you nothing about Google or ChatGPT.

Can your analytics or server logs show AI citations?

Not directly. They record visits and page fetches. That's evidence an AI engine touched your site, not proof that it cited you.

Analytics sees the people who click through from an AI answer. On this show, Brie Anderson recommended reading GA4's source and medium data instead of building a custom channel group. New AI sources keep appearing, and a fixed list leaves them out. What analytics can't see is an answer that named you and sent nobody. That leaves no trace.

Server logs see a little further. OpenAI says that when users ask ChatGPT a question, "it may visit a web page with a ChatGPT-User agent." Perplexity says its Perplexity-User agent "might visit a web page to help provide an accurate answer and include a link to the page in its response." A hit from either agent means a live question led the engine to your page, but the log entry won't say what the question was or whether the answer cited you. No hits from either agent? Check that a firewall or bot-blocking setting isn't turning them away, using our guide to blocking AI crawlers to see which agent does what.

How do you check ChatGPT and Perplexity by hand?

Ask a fixed set of buyer questions, worded the same way every time. Repeat each one several times and log every answer.

1. Write down the questions buyers ask before they contact anyone. Start with five to ten, worded the way a buyer would type them. Leave your company name out. A question that names you tests whether the engine can find you, not whether it picks you.
2. Hold the conditions steady. Use a logged-out browser, or an account with memory switched off. Check from the same location each time, and open a new chat for every question. An engine that knows your history can tailor its answer to you, and your buyers don't share your history.
3. Go slowly if you automate it. Send scripted checks one at a time with pauses, and read each engine's terms first. An engine that blocks or throttles you can return an empty answer, and in a log that looks exactly like not being cited.
4. Ask on every engine your buyers use. Start with ChatGPT, with search turned on. Then Perplexity, and Copilot if your buyers work in Microsoft's tools. Check Google by hand as well, because Search Console can't tell you which question you appeared for. In an AI Overview, open the full source list, not only the links shown before you expand it.
5. Give every answer one outcome. Cited means your page is in the source list. Named means your company is in the text but your site isn't among the sources. Absent means neither, so note which sources took the slot. No answer means Google showed no AI Overview for that search, which isn't the same as leaving you out of one.
6. Run every question seven or eight times on each engine. Fewer runs leave a margin of error too wide to act on.
7. Log the date and the mode with every answer. ChatGPT with search and ChatGPT without it are different products. So are AI Overviews and AI Mode. The engines also change underneath you. Our variance study notes that model versions shifted while it ran, and only a dated log lets you separate an engine change from a change in your standing.

What should the log look like?

One row per answer. The rows below are invented to show the format, all for one question: "how much does a commercial roof inspection cost."

The last column is the one people skip, and it's the one you'll use most, because when you aren't cited it names the page to study.

Does being named without a link count?

Count it, but in its own column. In our variance study, the brands we tracked were named in 25.7% of all answers and cited in 18.0%, so being named is the more common way to show up. A buyer who reads your name in an answer has heard of you, even with nothing to click, but no report will count it because there's no link.

Paraphrase is the hard case. Now and then an answer repeats a figure or claim that only your page makes, but cites another site for it. Or it cites nothing at all. You can be confident it's yours only when the material is unique, such as your own survey data. Log it as absent with a note, not as a citation. Two people reading the same answer will sometimes disagree about whether it paraphrases you, and that judgment call is the weak point of checking by hand.

How many runs make a reading reliable?

Seven or eight per question on each engine, going by our own data. AI engines sample their answers instead of looking them up. Ask the same question twice and the sources can change. Across the 12,778 answers our team at EWR Digital collected from five engines, no engine returned the exact same set of sources for a repeated question more than 2.5% of the time.

The variance study then took the 566 combinations of question and engine with at least ten answers each and measured how the margin of error on a visibility percentage shrinks as runs are added. A visibility percentage is the share of answers a business appears in.

"A single run gives a number that could be wrong by a third," the study concludes, and a margin that wide can't tell a business that appears in most answers from one that rarely does. The study's advice is to "report the interval alongside the number" and to claim no month-to-month trend until the two months' intervals stop overlapping. It also says to report each engine on its own, because the five it measured behaved like five different systems.

People are noisier than companies. Personal-brand questions were the least stable kind in the study. "Anyone measuring an individual's AI visibility is working in the noisiest part of the space and needs the most sampling," it says. So a consultant or an author tracking their own name needs more runs, not fewer.

Weigh the study's limits before you lean on it. Every answer came through one tracking platform. The seven sites it tracked weren't a random sample, and the study itself says they're concentrated in legal and B2B services. It observed answers instead of testing changes, so it describes variance without explaining what causes a citation. It covered ChatGPT, Perplexity, Gemini, Claude and DeepSeek but not Google's AI Overviews or Copilot, so we can't give you a run count measured on those two.

Do the arithmetic before you commit. Five questions on four engines at seven runs each is 140 checks per round, and at that volume checking by hand stops being practical for long.

What should you do when you aren't cited?

Study the page that took your slot. Your log names it. Comparing the two pages tells you more than any general checklist.

If you were cited before and stopped, rule out the boring causes first: our guide to how AI Overviews choose their sources walks through them, from indexing problems to a competitor who answers more of the question.

Then compare the pages. Start with the gap that's quickest to spot: does the other page answer follow-up questions that yours skips? A few minutes of reading will tell you. Next, check whether every heading on your page is answered in the two sentences under it, so a passage can be lifted out whole. Leave corroboration for last because it takes longest to change. Is the other business described on sites it doesn't control, where yours isn't? Industry directories and trade press are the places to look.

Change one thing at a time. Re-run the same questions the same number of times before and after, or you won't know which change worked, if any did. For Microsoft's engines, Bing's Compare view lays the two periods side by side. The page work itself is in our guide to answer engine optimization.

Should you pay for a citation-tracking tool?

Once checking by hand stops being practical, yes. Pick one that reports its margin of error.

Check what you already pay for first. Some rank trackers only flag whether a search shows an AI Overview while others record which sources it cited, and that difference decides whether the tool answers your question at all.

A dedicated tracker runs the routine above on a schedule and keeps the history for you. It saves time, not judgment. It sees no further inside the engines than you do and asks from its own accounts and locations, which may not match your buyers'.

Ask any vendor these questions before you sign:

• How many times is each question run on each engine before a number is reported?
• Is the number shown with a margin of error?
• Which engines and modes are covered? Are the checks run logged in or logged out?
• Where are the checks run from?
• Can you see the raw answers and source lists behind any number?

EWR Digital, the agency behind this show, sells this kind of checking as its AI Visibility Audit. Hold us to the same questions.

Want a citation check on your buyers' questions, run enough times to tell a real change from noise? Book a discovery call with EWR Digital →

Keep reading: How Do Google's AI Overviews Decide What to Cite? · What Is Answer Engine Optimization (AEO)? · What Does AI Visibility Mean for a Business?

Kept a citation log for your own site over a few months? Send us the log. We'll compare it with our variance figures and, with your permission, add what we learn to this guide.