How to Check If ChatGPT Cites Your Website

Checking a citation is diagnosis. Publishing is later. If you generate articles to "get cited" before you know whether the search crawler ever fetched the candidate URL, you are paying for fiction.

Checking a citation is diagnosis, and publishing is a later step. The order matters because the cost of generating articles before you know whether the search crawler fetched the candidate URL is the cost of fiction, and fiction does not cite back. The first job is to find out whether the page was even in a position to be cited, then to find out whether it was, and only then to write.

Cited means a URL you control showed up as a source. Named without a link is a mention. utm_source=chatgpt.com is a referral. Google Search Console will not answer this. Overviews are Google.

The definitions are the part to hold before any tooling. Cited means a URL you control showed up as a source in the answer. Named without a link is a mention, which is a different KPI with no click attached. utm_source=chatgpt.com is a referral, a session that came in through ChatGPT. Google Search Console will not answer this question, because Overviews are Google and GSC does not store ChatGPT answers.

OpenAI: Publishers FAQ, bots. SearchBot for Search surfacing. GPTBot for training. ChatGPT-User for live user fetches. Different jobs. Robots.txt can take about 24 hours. Allow published SearchBot IPs. A 403 is an infra ticket.

The OpenAI source pages are the basis for the bot work. The Publishers FAQ and the bots page name three crawlers with three different jobs. SearchBot is for Search surfacing, which is the crawler that puts a page in a position to be cited. GPTBot is for training, which is a separate legal decision. ChatGPT-User is for live user fetches, which is a third job. Robots.txt can take about 24 hours to land, so a change is not a change until a day has passed. Allow the published SearchBot IPs, which is the access step that precedes any content argument. A 403 is an infra ticket, not a content problem, and it belongs to the hosting or firewall owner.

ChatGPT-User may ignore robots.txt because a user asked. It is not Search index proof. Atlas may show title plus link for a URL that arrived via a third party. noindex if you want that gone, and the crawler must read the tag.

ChatGPT-User is the crawler that gets ignored most often in this analysis. It may ignore robots.txt because a user asked, which means a blocked URL can still be fetched in a live session. It is not Search index proof, so a ChatGPT-User fetch does not mean the page is in the Search index. Atlas may show a title plus link for a URL that arrived via a third party, which is a pickup rather than a citation you earned. noindex is the tag if you want that gone, and the crawler must read the tag for the tag to work.

Look at errors before copy

Promptwatch Agent Analytics: ChatGPTBot and related, errors, crawl-to-citation. Cloudflare through custom HTTP. Professional is where 25M logs show up ($245/mo). Essential $95/mo is the prompt desk. Explore (10 ChatGPT prompts, free) will not replace logs.

Errors come before copy, and the tool that shows them is Agent Analytics on Promptwatch. Agent Analytics logs ChatGPTBot and related crawlers, errors, and the crawl-to-citation path, across Cloudflare through custom HTTP. Professional is where 25M logs show up, at $245/mo, which is the tier where a miss can be read as a CDN block. Essential at $95/mo is the prompt desk, the seat that runs the standing prompt list. Explore, with 10 ChatGPT prompts and free, will not replace logs, so the free tier is a taste, not a diagnosis.

No 200, no citation argument. Stop Content Agents. Do not open the review inbox.

The rule that follows from the error view is short. No 200, no citation argument. If the crawler never fetched the page successfully, there is no content case to make, and generating more content will not change the fetch. Stop Content Agents. Do not open the review inbox, because the review inbox is for pages that were fetched and lost, not for pages that were never fetched.

Look at the source URL

Freeze buyer questions, including vs-queries. Run ChatGPT Search, not only chat without browsing. Save: named / cited to you / cited to a third party / absent. One session is noise.

The source URL is the second view. Freeze buyer questions, including vs-queries, because those are the questions where a citation is a customer. Run ChatGPT Search, not only chat without browsing, because Search is the surface that returns sources. Save four outcomes per answer: named, cited to you, cited to a third party, or absent. One session is noise, because answers shift between runs, and a sample of one is not a measurement.

Citation analytics: your page vs domain vs Reddit vs YouTube. Trends, not one scrape. Named without a link is not a cite.

Citation analytics is the layer that turns one scrape into a trend. The breakdown is your page versus domain versus Reddit versus YouTube, so a name-drop that cited a listicle on a third-party domain reads differently from a cite that sent traffic to your page. Trends, not one scrape, are the unit, because a single answer is a point and a trend is a signal. Named without a link is not a cite, and the distinction is the one that keeps a mention count from being reported as a citation count.

Otterly.AI from $29/mo will not do this depth on our listing. Mention count, possible week lag. Peec AI screenshots the mention. We still want the URL. We keep Peec last. Profound Starter is ChatGPT-only, 50 prompts, annual $99. HubSpot's grader is a snapshot. Scores are not guaranteed.

The other tools in this row each miss a layer. Otterly.AI from $29/mo will not do this depth on our listing, returning a mention count with a possible week lag. Peec AI screenshots the mention, which is evidence, but we still want the URL behind the screenshot, and a screenshot without a URL is not a citation log. We keep Peec last, as the site convention requires. Profound Starter is ChatGPT-only, 50 prompts, annual $99, so it is not a multi-engine tracker on that SKU. HubSpot's grader is a snapshot, and scores are not guaranteed, which HubSpot states itself.

Then the ticket

404 after launch: redirect. Reddit wins: maybe you are not the source. Your /blog/2019 wins: fix that URL. Missing claim: Content Agents to Webflow or Framer after review, 5 per month on Essential. WordPress: export. Do not pretend the CMS connection is live.

The ticket is the action that follows the diagnosis, and each diagnosis has its own action. A 404 after launch is a redirect, because the page exists somewhere and the URL is wrong. Reddit wins means maybe you are not the source, and the work is offsite rather than on-page. Your /blog/2019 wins means an old URL is picking up the citation, and the fix is that URL, not a new page. A missing claim is the one case for Content Agents, to Webflow or Framer after review, 5 per month on Essential. WordPress is an export, not a live connection, so do not pretend the CMS connection is live when it is not.

Visitor analytics (script or GTM): referral vs citation. Unified Actions: one list. Fetch errors first, then offsite, then draft. Daily checks. No pager SKU.

Visitor analytics, via a script or GTM, is the layer that separates referral from citation, so a session that came in is not reported as a citation that was earned. Unified Actions is one list, not a siren, so the to-do list is a queue, not an alert. The order of work is fetch errors first, then offsite, then draft, because a draft cannot fix a fetch error and an offsite play cannot fix a missing claim. Daily checks are the cadence. There is no pager SKU, so nobody is woken for a missing citation.

Pass looks boring. SearchBot 200s, two weeks of the same prompts, a column for cited URL, a ticket when the cite is missing.

A pass looks boring, and boring is the goal. SearchBot 200s, two weeks of the same prompts, a column for the cited URL, and a ticket when the cite is missing. That is the shape of a program that checked, and it is the shape that does not produce a slide full of numbers that do not mean anything.

Wait about 24 hours after robots.txt. GPTBot is a separate legal call. A single "block AI" rule often kills Search. ChatGPT-User is not index proof. Atlas title-plus-link after a third-party pickup is not a citation you earned. noindex, and the crawler must read the tag.

The access caveats repeat because they are the most common reason a check fails without a content cause. Wait about 24 hours after a robots.txt change. GPTBot is a separate legal call from SearchBot, and a single "block AI" rule often kills Search along with training, which is the unintended consequence that loses citations. ChatGPT-User is not index proof. An Atlas title-plus-link after a third-party pickup is not a citation you earned, and noindex, with a crawler that reads the tag, is the way to remove it.

Explore (10 ChatGPT prompts) finds the hole. Essential $95/mo is the working list. Professional $245/mo when logs matter (25M on that tier). Daily checks. Visitor analytics (script or GTM) next to the citation column. Do not average referral and cite into one score. HubSpot's grader is a snapshot. Scores are not guaranteed. We do not file it as a citation ledger.

The tool ladder is the last piece. Explore, with 10 ChatGPT prompts, finds the hole, which is enough to see whether there is a gap at all. Essential at $95/mo is the working list, with citation analytics and visitor events. Professional at $245/mo is the step when logs matter, with 25M on that tier. Daily checks are the cadence. Visitor analytics, via a script or GTM, sits next to the citation column, not in place of it. Do not average referral and cite into one score, because the average is neither. HubSpot's grader is a snapshot, scores are not guaranteed, and we do not file it as a citation ledger.

That is the check. Product: promptwatch.com.