Does ChatGPT run JavaScript
ChatGPT does not browse your site the way a person does. OpenAI publishes three crawlers, and a measurement from Vercel says none of them run the JavaScript that fills in a client-rendered page.
Check what AI crawlers see on one of your URLs. The result is the HTML before JavaScript runs, plus any robots.txt block for GPTBot, ClaudeBot, PerplexityBot and Googlebot.
The three OpenAI crawlers
OpenAI documents the split on its crawler page. GPTBot collects pages that may be used to train generative models. OAI-SearchBot is the crawler behind ChatGPT search: a site that disallows it is not shown in search answers, though OpenAI says it can still appear as a navigational link. ChatGPT-User fetches a page when a person, or a custom GPT, asks for one. OpenAI says ChatGPT-User is not an automatic crawl, and that robots.txt may not apply to it.
The three settings are independent. Allowing OAI-SearchBot and disallowing GPTBot is a supported combination. OpenAI also says that if you allow both, one crawl may be reused so they do not fetch the same URL twice, and that a robots.txt change can take about a day to show up in search. The published user agents are on that same docs page, and the crawler check sends the GPTBot string from it.
None of those docs say the crawler executes scripts. They describe a fetch. What happens to the JavaScript file after it is downloaded is a separate question, and the vendor page does not answer it. For that, the public measurement is Vercel and MERJ.
| Bot | Company | What it is for | JavaScript | Docs |
|---|---|---|---|---|
| GPTBot | OpenAI | Collects pages that may be used to train generative models. Blocking it is separate from ChatGPT search. | Does not run itVercel's study found it fetches JavaScript files and does not run them. | Docs |
| OAI-SearchBot | OpenAI | Surfaces sites in ChatGPT search. OpenAI says a site that opts out is not shown in search answers, though it can still appear as a navigational link. | Does not run itNamed in the same Vercel study as a crawler that does not render JavaScript. | Docs |
| ChatGPT-User | OpenAI | Fetches a page when a person asks ChatGPT or a custom GPT to look. OpenAI says it is not an automatic crawl, and robots.txt may not apply. | Does not run itIncluded in the Vercel list of OpenAI crawlers that do not execute JavaScript. | Docs |
What the Vercel study measured
Vercel published The rise of the AI crawler with MERJ. In the month they measured on their network, GPTBot made 569 million fetches, Claude 370 million, Applebot 314 million and PerplexityBot 24.4 million, beside 4.5 billion from Googlebot. The rendering result was the same across the apps they checked: OAI-SearchBot, ChatGPT-User, GPTBot, ClaudeBot, Meta-ExternalAgent, Bytespider and PerplexityBot did not render JavaScript.
They did download script files. JavaScript was 11.50 percent of ChatGPT fetches and 23.84 percent of Claude fetches. The study says those files were not executed, so a download of app.js in your logs is not evidence that the bot built the page. Content that was already in the HTML, including JSON in the document, can still be read. A div that stays empty until a fetch in the browser cannot.
Two crawlers in that study did render. Gemini uses Googlebot, which renders JavaScript in a later wave. Applebot uses a browser. Common Crawl, which many models train on, did not render. The study is one network at one time. It is the best public measurement we have, and it is why the safe page is the one whose words are in the first HTML response.
Fetches in one month on Vercel's network
- Googlebot
- 4.5B
- GPTBot
- 569M
- Claude
- 370M
- Applebot
- 314M
- PerplexityBot
- 24.4M
Claude and Perplexity
Anthropic publishes ClaudeBot for training, Claude-SearchBot for search results, and Claude-User for a fetch a person triggered. Their help center says the bots honor robots.txt, and that a rule naming ClaudeBot beats a general no-training signal. Vercel measured ClaudeBot with the crawlers that do not run JavaScript. The study did not name Claude-SearchBot, so this page does not invent a rendering result for it.
Perplexity publishes PerplexityBot for search and says it is not used to train foundation models. The example user agent is Mozilla/5.0 (compatible; PerplexityBot/1.0; +https://perplexity.ai/perplexitybot). Perplexity-User is the user-triggered fetch, and Perplexity says it generally ignores robots.txt. Vercel included PerplexityBot in the no-render list. A client-rendered page is an empty document to that crawler even when robots.txt allows it.
| Bot | Company | What it is for | JavaScript | Docs |
|---|---|---|---|---|
| ClaudeBot | Anthropic | Collects web content that may be used to train Claude. Anthropic says a disallow is a signal to keep future material out of training sets. | Does not run itVercel measured ClaudeBot fetching pages without executing JavaScript. | Docs |
| Claude-SearchBot | Anthropic | Browses to improve Claude search results. Anthropic says disabling it can reduce how often the site shows up in those results. | Not measuredAnthropic documents the token. The Vercel study named ClaudeBot, not this search bot, so we do not claim a rendering measurement for it. | Docs |
| PerplexityBot | Perplexity | Surfaces and links sites in Perplexity search. Perplexity says it is not used to train foundation models. | Does not run itVercel included PerplexityBot with the crawlers that do not render JavaScript. | Docs |
Rendering numbers are from Vercel, The rise of the AI crawler.
How to test your own URL
Paste the URL into the crawler check. It loads the page with JavaScript off, counts the words, and requests the same URL as GPTBot, ClaudeBot and PerplexityBot so you can see a status code and a challenge. It also reads robots.txt for those bots plus OAI-SearchBot, ChatGPT-User, Claude-SearchBot, Google-Extended, Googlebot and Bingbot, and shows the rule that matched.
Or do it with curl. One user agent per command, then read the body. A 200 whose body is a root div is still an empty page. For Google-Extended, read robots.txt. That token is not a separate downloader. Googlebot fetches the page, and Google-Extended only says whether the fetch may be reused for Gemini.
bash
curl -s -A "Mozilla/5.0 AppleWebKit/537.36 (KHTML, like Gecko); compatible; GPTBot/1.4; +https://openai.com/gptbot" "$URL"bash
curl -s -A "Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/131.0.0.0 Safari/537.36; compatible; OAI-SearchBot/1.4; +https://openai.com/searchbot" "$URL"bash
curl -s -A "Mozilla/5.0 AppleWebKit/537.36 (KHTML, like Gecko); compatible; ChatGPT-User/1.0; +https://openai.com/bot" "$URL"bash
curl -s -A "Mozilla/5.0 (compatible; ClaudeBot; +https://www.anthropic.com)" "$URL"bash
curl -s -A "Mozilla/5.0 (compatible; Claude-SearchBot; +https://www.anthropic.com)" "$URL"bash
curl -s -A "Mozilla/5.0 (compatible; PerplexityBot/1.0; +https://perplexity.ai/perplexitybot)" "$URL"bash
curl -s -A "Mozilla/5.0 (compatible; Googlebot/2.1; +http://www.google.com/bot.html)" "$URL"bash
curl -s -A "Mozilla/5.0 (compatible; bingbot/2.0; +http://www.bing.com/bingbot.htm)" "$URL"bash
curl -s "$ORIGIN/robots.txt"How to fix an empty page
Put the words in the HTML. Server rendering and static generation do that for every client, which is what Google recommends instead of dynamic rendering. A rendering service does it for crawler user agents only, and leaves the JavaScript app for people. Pagerender is that second path: try it free for a week, then decide if a rebuild is the real project.
If the HTML is fine and the status is 401, 403, 429 or 503, the firewall is the problem. OpenAI, Anthropic and Perplexity publish the tokens and, for some bots, IP lists. Allow the ones you want cited. A robots.txt disallow is a different switch, and it is per bot. Blocking GPTBot does not hide a site from ChatGPT search. That switch is OAI-SearchBot.
What a crawler request passes through
Firewall
The host answers the bot
Stops on: 401, 403, 429 or 503robots.txt
The bot's token is allowed
Stops on: A matching DisallowHTML
Words arrive before JavaScript
Stops on: Under 60 wordsnoindex
The page asks to be indexed
Stops on: A noindex tagRead and indexed
Search can rank it and AI answers can cite it
A chat session is not the crawler
People test this by pasting a URL into ChatGPT and asking what the page says. That is ChatGPT-User, or a browse the product does for that person, and OpenAI says robots.txt may not apply. It is one fetch, for one question. It is not GPTBot building a training set, and it is not OAI-SearchBot deciding what search can cite tomorrow. A page can answer in a chat you triggered and still be absent from search if OAI-SearchBot is blocked or the HTML is empty.
The same split exists at Anthropic and Perplexity. Claude-User and Perplexity-User are the fetches a person caused. ClaudeBot and PerplexityBot are the automatic ones. Perplexity says the user fetch generally ignores robots.txt. A successful answer in the chat box is not a clean bill of health for the crawler. OpenAI says a robots.txt change can take about a day to affect search, so a test you run the minute you save the file can still show the old rule.
Read the body, not only the status. curl with the GPTBot user agent from OpenAI's docs will often return 200 from a host that is happy to talk to bots, and the body will still be a shell. Count the words you can see in that body. Under 60 is the checker's nearly invisible line, taken from the analyzer in the product, not from a ranking blog. Between 60 and 250 the page is only partly there: a title and a slogan, with the real content still behind a request the bot will not make.
Google is the other direction. Googlebot can render JavaScript after the first fetch, so a URL that looks fine in Search can still be an empty document to GPTBot. Do not use "Google indexed it" as the test for ChatGPT. Use the raw HTML. The crawler check and the curl commands on the checker page are that test, with the user agent string OpenAI publishes for GPTBot.
How the word count turns into a verdict
- Nearly invisible
- Under 60 words
- Partially visible
- 60 to 250 words
- Readable
- 250 words or more, with a meta description
Common questions
Does ChatGPT run JavaScript?
OpenAI documents fetches, not script execution. Vercel and MERJ measured GPTBot, OAI-SearchBot and ChatGPT-User fetching pages without running JavaScript. Ship the text in the HTML if you want those crawlers to read it.
Does blocking GPTBot hide a site from ChatGPT search?
It does not. OpenAI says search opt-out is OAI-SearchBot. GPTBot is the training crawler. The two robots.txt tokens are independent.
Does Googlebot work the same way?
Googlebot fetches the HTML first and can render JavaScript later. The AI crawlers in the Vercel study did not. A page that only works after the second wave is a page those AI crawlers do not see.