How to tell if ChatGPT was trained on your content

You wrote the words. You ran the analysis. You produced the original work. And now a chatbot answers questions about it as if it had read everything you ever published — without ever sending you a referral, a citation, or a cheque.

Was your work used as training data? This is the practical answer. Five concrete tests, ordered from "fastest" to "most definitive," that will tell you whether OpenAI, Anthropic, Google, Perplexity, or any of the other major AI labs have scraped your content. Some of them take 30 seconds. The most thorough takes an evening.

The short answer

There is no official lookup. OpenAI does not have a Search Console. Anthropic does not publish a "was your site in our training set?" tool. There is no API that returns a yes/no for "is this URL inside GPT-5's weights?" The reasons are partly competitive (publishing the corpus would let competitors replicate the model) and partly legal (admitting specific training-data inclusion creates litigation exposure).

But the question can still be answered, indirectly, with high confidence. Five ways:

1. Check your server logs for the named AI user-agents.
2. Check Common Crawl — the public dataset that most major models train on.
3. Probe the model with specific, low-frequency phrases from your work.
4. Check the OpenAI training-data opt-out roster (and the equivalents from other labs).
5. Use a third-party scrape detection service.

The first three are free. The fourth tells you whether you've been excluded going forward. The fifth is a paid product that wraps the others.

Test 1 — your server logs (definitive for "have they ever visited?")

If your server logs are intact, this is the cleanest evidence available. AI crawlers identify themselves by User-Agent string when they visit your pages. If they ever scraped you, the visit is in your logs.

The user-agents to look for as of mid-2026:

• GPTBot — OpenAI training crawler
• ChatGPT-User — OpenAI in-session browsing (a user asked ChatGPT to fetch your page)
• OAI-SearchBot — OpenAI search-index crawler
• ClaudeBot and Claude-Web — Anthropic
• anthropic-ai — older Anthropic agent, still in some traffic
• Google-Extended — Google AI training (separate from Googlebot)
• PerplexityBot and Perplexity-User — Perplexity
• Bytespider — ByteDance / Doubao
• Applebot-Extended — Apple AI training (separate from Applebot)
• Meta-ExternalAgent and Meta-ExternalFetcher — Meta
• Amazonbot, cohere-ai, CCBot, Diffbot, DuckAssistBot, YouBot — others

Pull the last 12 months of access logs and grep for any of these strings. Most hosting providers (Cloudflare, Netlify, Vercel, AWS, anything Apache or Nginx) keep logs by default.

The shortcut on Cloudflare: Security → Events → filter by Bot Category: AI Crawler. Last 30 days are free; 90 days on Pro.

What the result means:

• Many visits, recently — they're actively crawling. Your robots.txt and server-side blocks may not be effective.
• Many visits, then stopped — they finished, probably ingested, and now treat your site as a known source rather than a new one.
• No visits, but content known to be in the model — they got it indirectly through Common Crawl (see test 2) or a third-party dataset.
• No visits, content not known — you're either too obscure, behind a paywall, or your blocks worked.

Test 2 — Common Crawl (the smoking gun for most models)

Common Crawl is a non-profit web archive that publishes monthly snapshots of public web pages. It's free, it's enormous, and it's the foundation of the training data for most major large language models — OpenAI's GPT family, Anthropic's Claude family, Meta's Llama family, and many open-source models all start from a Common Crawl base.

The implication: if your URL is in Common Crawl, it's almost certainly in the training data of the major models, regardless of whether the labs ever sent their own crawler to your site. Common Crawl is the ambient supply.

To check whether your site is in it, the easiest tool is the Common Crawl URL Index search at index.commoncrawl.org — paste your domain, get back the list of every URL that's been captured, when, and how many monthly snapshots it appears in.

We built a friendlier version of this check, scoped to creators and combined with a robots.txt audit and AI-bot visit summary: the scrape check — paste a domain, get a single-page answer in about 30 seconds.

The answer divides creators into two groups:

• In Common Crawl — the major labs already have your work. You can block future crawls, but the back catalogue is in their training data, and your leverage going forward is over new work.
• Not in Common Crawl — your work is rare training data and worth materially more in a licensing conversation, because the buyers can't get it any other way. This is the better situation to be in.

Test 3 — probe the model directly (qualitative, but revealing)

The least technical and most accessible test. Take a low-frequency phrase from your work — something specific enough that it shouldn't appear in generic web content but distinctive enough that a model trained on you would have seen it. Paste it into the chatbot and see what it says.

Good probe phrases:

• A specific factual claim you made in an article that wasn't covered anywhere else
• An idiosyncratic turn of phrase you've used repeatedly
• A non-public detail from a podcast interview transcript
• A piece of original analysis with a specific conclusion

Ask the chatbot a question that would only be answerable if it had read your work. For example: "What did [your name] argue about [your specific topic] in the [year] essay published on [your site]?" If the answer is detailed, accurate, and uses your phrasing — they've trained on you. If the answer hallucinates plausibly but inaccurately — they haven't.

Three caveats:

1. The model may have been trained on someone quoting you, not on the original. If a third-party site referenced your work in detail and that site was scraped, the model knows your work indirectly. Look for verbatim sentence reuse rather than topical knowledge.
2. Models occasionally "memorise" content even from training data they don't have direct access to, through chains of reference. This is rare for low-traffic content but possible for high-profile work.
3. The probe doesn't tell you which model variant — GPT-3.5 trained on different data than GPT-5, Claude 2 trained on different data than Claude 4. The major labs have done multiple training runs with different corpora.

Test 4 — the opt-out rosters

OpenAI publishes a list of domains that have requested removal from training data going forward (via the GPTBot Disallow declaration in robots.txt) — but it's not searchable from the outside. The closest signal is to check your own opt-out status by looking at your robots.txt and confirming the GPTBot block is live and reachable.

Anthropic, Google, Perplexity, and most other labs offer similar opt-outs through their respective named user-agents. The mechanism is the same: drop a Disallow: / for the named bot in your robots.txt. Going forward, compliant crawlers will skip your site.

Going backward — having content removed from a model that has already been trained on it — is harder. The standard answer from the major labs is some variant of "we can flag your content for exclusion from future training runs, but cannot retroactively remove it from current model weights." There is no clean technical solution to this; once the model has trained on the content, removing it requires retraining, which costs hundreds of millions of dollars.

What this means in practice: the opt-out rosters control your future exposure, not your past. The honest answer about historical content is that the labs have it, and the leverage you have over already-trained models is licensing (paid agreements going forward), not deletion.

Test 5 — third-party scrape detection

A small industry of paid tools has appeared to wrap the above tests with nicer UI and richer alerting:

• Cloudflare's Bot Analytics — free on all plans, surfaces AI-crawler traffic by name with daily totals.
• Dark Visitors — paid, monitors AI bot activity across your site with longer retention and alerting.
• Tollbit — paid, combines bot detection with a paywall/licence layer on top.
• ArchiveBay's the scrape check — free, combines Common Crawl, robots.txt audit, and AI-bot history check into a single page for creators specifically.

The third-party tools mostly do what your server logs already do, more nicely. If you have logs, start there. If you don't, a service is the right call.

What to do with the answer

Whichever combination of tests you run, you'll end up in one of three buckets.

Bucket 1 — your work is in their training data. Most independent creators are in this bucket for content older than ~2 years. The leverage you have now is over new content, going forward, and over attribution and licensing deals on the back catalogue. There's nothing technical that can pull your old work out of GPT-5; what you can do is ensure your future work is either blocked or paid for.

Bucket 2 — your work is partially in their training data. Some content (older, public, in Common Crawl) has been ingested; other content (paywalled, recent, on a private feed) hasn't. Your task is to figure out which parts of your archive are still scarce, because that's the part with real licensing value. Use the scrape check to see the breakdown.

Bucket 3 — your work is not in their training data. Rare in 2026 unless you've paywalled or kept content behind credentials. If this is you, your archive is scarce training data, and the prices it can command are materially higher than the public-web average. The valuation tool gives you a range; the marketplace gets you in front of buyers.

The point of running these tests is not just to confirm the suspicion that yes, AI labs have used your work. It's to identify which parts of your archive still have leverage, because that's where the licensing money lives. Knowing what's been taken is the first step in negotiating to be paid for what hasn't.

---

Run the scrape check for the all-in-one answer, or the valuation tool for what your archive is worth.