ShoutAd.GonzaloGomezRufino

Blog · 18 November 2026

How to tell if ChatGPT cites your site, and what to do if it doesn’t

On 10/11/2026, when I search for "como saber si chatgpt cita mi web", the first results explain what GEO is and how to "appear in ChatGPT". They are generic. There is no repeatable way to measure it. Here is the method I use to see if ChatGPT cites a domain and whether my changes work.

I do not repeat the GEO diagnosis. I complement it. The five verifiable points sit there. Here I cover continuous tracking: questions, cadence, logging, and impact evaluation.

What “cite” means and what is worth measuring

There are three useful signals. 1) A citation with a link or visible domain. 2) A brand or domain mention without a link. 3) Use of your content without recognisable attribution. The first is the main one. The second still counts. The third demands a review of text and concepts to spot clear overlaps.

To avoid scatter, I measure three things per prompt: whether the domain appears, how it appears (link, bare domain, name only), and where in the reply. I also note whether, after I ask “cite your sources”, it adds your site or keeps it out.

Replicable question battery

The key is a stable set of prompts that reflect your intent. For each target page I prepare 3 to 5 natural questions in Spanish. Useful patterns include:

“Recommend reliable resources on [topic] and link the sources.”

“Which sites explain [specific task] well? Give me links.”

“Cite three sources to understand [concept] with practical guides.”

“Where did you get that explanation? List your sources.”

I use small variants for neutral Spanish: “recursos”, “fuentes”, “guías”; “explican”, “cubren”, “describen”. I test informational, comparative, and how-to queries. I avoid prompts that invite source-free answers, like “define” with no extra instruction.

Context matters. A query for “recursos en español sobre [tema]” can change the list. So I include versions with “en español” and versions without it, based on the market I want to cover.

Test ritual: sessions, variability and periodicity

ChatGPT varies by session and by updates. To cut noise, I run each prompt in a new conversation, with no history. I only count the first turn. If I ask “cite your sources”, I do it in a second turn within that same chat and log it separately.

Cadence: I set a monthly cut. A reasonable scenario is a team that spends twenty hours a month on this: it runs the full battery once a month and repeats a subset mid-month to catch sharp shifts. If you work solo, prioritise the full monthly pass over scattergun microtests.

I also alternate times. I run half in the morning and half in the afternoon. That helps spot one-off swings. If a reply looks erratic or invents a link, I do not correct it there and then: I log it as a “hallucination” and move on. The method measures first-intent output.

How to log without going mad

I use a sheet with these columns: date, model, prompt language, exact prompt, associated target URL, whether it cites the domain (yes/no), how it cites (link/domain/name), position in the reply (top/middle/bottom), whether it appears when I ask “sources”, brief comments, and an optional capture.

When it cites but links to the home page and not the specific guide, I mark it “weak citation”. When it mentions the name without the domain, I log it as a “mention”. I keep appearances after the “sources” follow-up separate, to distinguish spontaneous signal from induced signal.

Before and after: knowing if the change worked

I set a baseline. I run the full battery and compute simple proportions: in how many prompts the domain appears, and how many include a link. Then I make content or structure changes and wait for a full measurement cycle.

I compare equivalent windows. Same prompt set, same language, same time bands. Only the first turn counts. If the share of spontaneous appearances rises, and stays up for two consecutive cycles, I treat it as a real improvement. If it drops, I review what changed on the page and in the wording.

What to do if it doesn’t cite you

If the domain does not appear, I go back to the content. ChatGPT tends to cite clear, useful, specific pieces. Pages with brief or generic definitions do not make the cut. Guides with practical steps, concrete examples, and well-answered FAQs perform better.

Structure helps. Clear task-based headings, a summary at the top, and conclusions like “when to use this and when not”. Add “common mistakes” and “edge cases”. Show update dates and authorship. If you have your own data, explain it and link to your method.

Avoid walls of text. Include terms people actually use. If your market is Hispanic, write in neutral Spanish and name the topic, tool, or process unambiguously. Comparisons should declare criteria and downsides, not just pros. This complements the GEO diagnosis; if you want to revisit the five verifiable points, see how to work GEO for search engines and AI assistants for a version focused on measurable results.

Control questions to force attribution

When your domain does not appear, I try a standard follow-up: “cite 3 concrete sources with a link”. If a different list comes up, I note the change. I also use: “which site did you take that definition from?”, which shows whether the model recognises sources or wings it. I do not validate links on the spot; I just record them.

Another useful move: “are there guides in Spanish specifically about [topic]?” If it replies with English sites, I log it as a language bias. Then I adapt the content to cover that gap: examples in Spanish, local terms, and screenshots or clear descriptions.

How to know the improvement is real

The temptation is to cheer a weekly spike. I prefer consistent signals. I want to see more spontaneous appearances, more correct links, and fewer “weak citations” for at least two cycles. I also cross-check with traditional search to see if the page gains visibility there too.

If after two cycles nothing moves, I shift to external building: reproducible research summaries, useful glossaries, and collaborations with topical sites. These are pieces people cite, and that ChatGPT often picks up when they appear across several trusted sources.

Common mistakes when measuring

Contaminating the session with many follow-ups and then counting those as spontaneous. Changing prompts every month and losing comparability. Testing with two questions and drawing conclusions. Not logging language or time. Ignoring hallucinations. Failing to capture the original reply.

The method is simple: define a stable battery, run it in fresh sessions, log with rigour, and compare equivalent windows. With that you will know if ChatGPT cites your site and, if not, what to adjust.

I post working automations on @WhatsMarketing_es.

If you want me to look at your case, get in touch. I work from Buenos Aires, originally from Málaga, with clients in Mexico City, Argentina, the rest of Latin America and Spain.

← See all articles