Tips
How to check whether ChatGPT cites your site
Write down the questions your buyers ask. Ask each engine several times with web search on, and log every source. The method we ran on our own site.
Write down the questions your buyers ask. Ask ChatGPT each one several times with web search on, and log every source it cites. Then do the same on the other engines your buyers use. One run is an anecdote. A fixed list, asked the same way each time, is a measurement.
We ran this on our own site on 24 September 2026, the day after we relaunched it. 30 questions, four engines, three runs each. That made 360 answers. Here is the method, with the mistakes we made on the way.
1. Write the questions a buyer would type
Our list has 30 questions in four groups. Sort yours the same way, because the groups give very different numbers.
| Group | What it asks | Ours | Example |
|---|---|---|---|
| Discovery | who provides the service, without your name | 9 | “Which agency does answer engine optimization for SaaS companies?” |
| Comparison | advice on choosing or buying | 8 | “How do I measure whether AI assistants are recommending my product?” |
| Brand | about you, by name | 6 | “What does synapsereality.io do?” |
| Trust | whether you can be trusted | 7 | “Is synapsereality.io a legitimate company?” |
Discovery and comparison are where new buyers find you. Brand and trust show what the engines say once someone already has your name.
Write the questions the way a buyer types them, not the way your homepage talks. And never make a competitor the subject of a question. You’ll see who the engines name instead in the sources.
2. Decide what counts as a citation before you look
Ours is a link to synapsereality.io or github.com/bensynapse, in the answer or in its list of sources. A mention of the name without a link is counted apart. The two are not the same number. 155 of our 360 answers named us. 106 linked to us.
If another company shares your name, expect trouble. “Who founded Synapse Research and when?” got no citation in 12 answers. Every engine answered about a different company called Synapse, or asked which one we meant.
3. Ask each question several times, with search on
Turn web search on. Without it the model answers from its training data, and last month’s page isn’t in there. In our run OpenAI, Gemini and Perplexity searched before every answer. Claude searched in 82 of its 90 and fetched a page without searching in 3 more.
Ask each question at least three times. Answers change between runs, and one run tells you nothing about that.
Ten questions by hand in the ChatGPT app is an afternoon. 360 answers is a script. Ours called each engine’s API with its search tool switched on. The 270 API answers cost $1.39 in total: $0.76 for OpenAI, $0.26 for Gemini and $0.37 for Perplexity. The Claude runs used a subscription.
The API is not the ChatGPT app. If you script it, ask a few of the questions in the app as well and compare.
4. Log the sources, not only the answer
Keep one row per answer. It holds the question, the engine, the run number, the full answer text and every source URL. We write them to one JSON Lines file, so the rates can be counted again later.
Keep the failures in the file too. Perplexity returned HTTP 429 three times, because we started a second run while the first was still going. We asked those three again with nothing else running. The error rows stayed in the file and out of the rates.
The source list shows who the engines read instead of you. In our 204 answers that didn’t name us, a forum site was cited in 25. Google’s developer documentation was cited in 20. The most-cited agency appeared in 15. That tells you where to be present, as well as which pages to write.
5. Read what the answers say about you
A count of citations won’t show you an answer that is wrong about you. 41 of our 156 brand and trust answers mentioned NFT, metaverse or Web3 work from our old site. About 15 of them described it as what we do today. One gave our founding year as 2015, the year a directory profile still lists. We were incorporated in 2018.
Look up the sources behind each wrong answer. The fix goes there, and often it isn’t your own site.
6. Measure the noise, then run it again the same way
Before you trust a change, rerun a few questions the same day. We asked five of ours again on the three API engines, three runs each, for $0.26. All 15 pairs landed within one citation of the first run. So a change of one citation on a question is inside our noise.
Keep the questions, the engines and the run count fixed between readings. We run ours again on 5 November, six weeks after launch.
Use an account with no link to your company. Our Claude runs went through a tool that passed Ben’s git username to the session. In 21 of its 39 brand and trust answers, it told the reader the company was probably their own.
What our first reading showed
Asked about us by name, the engines cited us in 105 of 156 answers. Asked for a provider or for advice, they cited us in 1 of 204. That one was a Perplexity answer about MAP monitoring, where we were the fifth source. We compare that with our search rankings in what we measured in launch week.
Can Search Console or Bing show you this?
In part. Google Search Console reports generative-AI impressions. Bing’s AI Performance report shows citations and cited pages across the surfaces it supports. Neither is a transcript of every buyer’s answer. Use them next to your question list, not in place of it.
Where to start
First check that the engines can read your pages at all. Our free crawler access check reads your robots.txt and the HTML a crawler gets, without running JavaScript. It doesn’t ask any engine a question.
Then write ten questions and ask each one three times. If you want the list run for you, with the sample size on every reading, that is what our AEO work does.