Call · 30 min
MarketingMattia EspositoUpdated 10 November 202613 min read

Do AI crawlers render JavaScript? Bot by bot, what has been measured and what has not, and how to check your site’s pages in half an hour.

A site that looks complete to you can reach ChatGPT with no text, no prices and no contact details, if a script adds them. Here is which bots run JavaScript, which don’t, the source for each, and an eight-point checklist, in Word and PDF.

In brief

In the only public study, the crawlers of OpenAI, Anthropic, Perplexity and Meta read the HTML without running JavaScript. Googlebot and Bingbot, which serve AI Overviews and Copilot, do run it.

The study is by Vercel and MERJ, dated 17 December 2024. No owner of an AI crawler has said in its documentation whether its bot runs JavaScript, and Google itself warns that not all bots do.

The rule that holds regardless: text, prices, contact details, title and canonical in the initial HTML. Every reader gets them.

The eight-point checklist can be downloaded below, in Word and PDF, with no email required.

Written by: Mattia Esposito, AI and marketing consultant, founder of Itria AI. From Puglia, based in Turin since 2023, he works remotely in Italy and abroad. LinkedIn profile.

The “Readable without JavaScript” checklist, to download

A sheet with the table of bots that run JavaScript and bots that don’t, eight checks with where to look, and a log for the five pages that matter, in Word to fill in and PDF to print. A browser is enough: Ctrl+U opens the page source. It takes half an hour; point 8 needs access to your logs.

The Readable without JavaScript checklist, with the table of AI crawlers that run JavaScript and the page-by-page log
FileContentsLink
Wordto fill in, in Italian

The table of bots, the eight checks, the page-by-page log.

javascript-e-crawler-ai-checklist.docx

PDFto print, in Italian

The same checklist, to tick off with a pen.

javascript-e-crawler-ai-checklist.pdf

N.What to checkWhere to look
1

Title, H1 and main text are in the raw HTML.

Ctrl+U or curl

2

Prices, address, phone number and email address are written in the HTML, not added by a script.

Ctrl+U, search for the email

3

The title and meta description are in the initial HTML.

Ctrl+U

4

The canonical is in the initial HTML, not injected with JavaScript.

Ctrl+U, search for canonical

5

Menu links are real links with an address, not buttons driven by a script.

Ctrl+U

6

The page returns 200, with no redirect chains.

Developer tools, Network

7

Near-identical versions of a page carry a canonical or an hreflang.

Ctrl+U

8

In the logs, which bots requested the page and whether they also requested the .js files.

Server or CDN logs

On 25 September 2026 I read the itria.io contact page the way an AI assistant reads it, as the HTML the server delivers, without running the scripts. The email address was missing. It was written out in full, but Cloudflare’s obfuscation took it out of the HTML and put it back with a script in the browser: a person could see it, a crawler could not.

I fixed it the same evening by excluding the address from obfuscation, and wrote up the case in the lab note on the email address AI assistants can’t see. The next day, across the websites of 121 food producers, I found the same flaw on 14: seven obfuscated the address, seven did not write it in the code at all. The study has the figures page by page.

On 7 September, after looking at the websites of 142 companies that also sell abroad, I put it like this: “An assistant doesn’t browse a site: it reads it, and reports what it finds written there.” The full post. If a script adds what is written there, the assistant reports an empty page.

Do AI crawlers render JavaScript? The short answer

In the only public study with traffic data, none of the main AI crawlers ran JavaScript, except Gemini and AppleBot. Vercel and MERJ measured it on the logs of the Vercel network and wrote on 17 December 2024: “none of the major AI crawlers currently render JavaScript”. The crawlers of OpenAI, Anthropic, Perplexity, Meta and ByteDance read the HTML as it came.

Google and Microsoft are a different matter. Googlebot, which AI Overviews and AI Mode depend on, runs JavaScript with an up-to-date version of Chromium. Bingbot, which serves the index behind Copilot, has used the Microsoft Edge engine since 2019. For them a page built by a script is readable, with a delay; for the others it is not.

The weak point of the answer is its date. Almost two years have passed, and no owner of an AI crawler has said in its documentation whether its bot runs JavaScript. OpenAI’s and Perplexity’s pages on their crawlers, read on 10 November 2026, do not contain the word. Google, which does run it, warns in its guide that “not all bots can run JavaScript”.

The same page read by two kinds of bot: AI crawlers that read only the initial HTML, and Googlebot and Bingbot, which run the script after a rendering queue The serversends the initial HTML AI crawlerGPTBot, ClaudeBot, Perplexity Reads only the HTMLthe script never runs Googlebot, Bingbotand AppleBot Rendering queueseconds or longer Rendered pagethe script has run
The same page arrives in two ways. What sits in the initial HTML, everyone reads; what a script adds, only bots that run it can see. Sources: Vercel and MERJ, 17 December 2024; Google Search Central, updated 4 March 2026.

The Vercel and MERJ study: what it measured, and when

For several months the study read the logs of nextjs.org and the Vercel network, plus two job-listing sites built on different technologies, to see which bots ran scripts. The authors are Giacomo Zecchini, Alice Alexandra Moore, Malte Ubl and Ryan Siddle. Vercel is a hosting company that sells server-side rendering: this is a vendor study, with real traffic data.

The most useful finding is a subtle one. ChatGPT and Claude downloaded JavaScript files, 11.50% and 23.84% of their requests, but did not run them. Content already present in the initial HTML, including JSON data inside the page, could still be read. Copilot was left out of the study, because it has no user agent of its own.

What the study measuredThe figurePeriod
Googlebot requests

4.5 billion

one month, late 2024

GPTBot requests

569 million

one month, late 2024

Claude requests

370 million

one month, late 2024

AppleBot requests

314 million

one month, late 2024

PerplexityBot requests

24.4 million

one month, late 2024

JavaScript files downloaded by ChatGPT, not run

11.50% of requests

several months

JavaScript files downloaded by Claude, not run

23.84% of requests

several months

ChatGPT requests for 404 pages

34.82%, plus 14.36% on redirects

several months

Claude requests for 404 pages

34.16%

several months

Googlebot requests for 404 pages

8.22%, plus 1.49% on redirects

several months

The last rows say something else: the AI crawlers of the time wasted about a third of their requests on pages that did not exist. A site with many broken links or redirect chains wastes their visits on useless addresses. The study makes a single recommendation: main content, titles, descriptions and navigation rendered on the server, with scripts kept for secondary elements.

Bot by bot: which ones render JavaScript

ChatGPT did not run JavaScript in the study; Gemini, Copilot and AI Overviews rely on bots that do. The table sets each bot beside its source and date. Where the source is the owner the row is firmer; where it is Vercel it holds for late 2024. The exact names of every bot, and what to block, are in the guide to robots.txt and AI crawlers.

Who readsRuns JavaScript?Source, date
Googlebot (AI Overviews, AI Mode)

Yes, up-to-date Chromium, with a rendering queue

Google, updated 4 March 2026

Gemini

Yes, through Googlebot’s infrastructure

Vercel and MERJ, 17 December 2024

Bingbot (Copilot)

Yes, Microsoft Edge engine

Bing, October 2019

GPTBot, OAI-SearchBot, ChatGPT-User

No: they download JS files without running them

Vercel and MERJ, 17 December 2024

ClaudeBot

No: it downloads JS files without running them

Vercel and MERJ, 17 December 2024

PerplexityBot

No

Vercel and MERJ, 17 December 2024

Meta-ExternalAgent, Bytespider, CCBot

No

Vercel and MERJ, 17 December 2024

AppleBot

Yes, a browser-based crawler

Vercel and MERJ, 17 December 2024

Claude’s web fetch tool (API)

No, Anthropic says so

Anthropic, read on 10 November 2026

Two rows need a careful reading. Google-Extended, often mistaken for Gemini’s crawler, is a robots.txt token: it decides whether pages are used to train Gemini, but the reading is done by Googlebot. And the only written sentence from Anthropic concerns its API tool: “The web fetch tool currently does not support websites dynamically rendered with JavaScript.” About ClaudeBot and Claude-SearchBot it says nothing.

Crawlers and browser agents: two different things

A crawler downloads pages for an index or for training; a browser agent opens the site the way a person does, to carry out a task, and runs the scripts. Claude-Agent, the browser Anthropic hosts so that Claude can do tasks on the web, is of the second kind: its requests carry the user agent of an ordinary Chrome with its own name added.

The difference matters because the two arrive at different moments. For an assistant to cite a page, the page usually has to go through a crawler or a fetch tool first, and there the script does not run. The browser agent comes later, when someone asks it to make a booking or compare a price. How to prepare a site for these agents is in the guide to sites ready for AI agents.

AI crawlers and JavaScript in 2026: what has changed, and what is unconfirmed

No official announcement: as of 10 November 2026 no owner of an AI crawler has said its bot runs JavaScript. Two claims are going round, pointing in opposite directions. A test published in September 2026 says it saw only raw HTML reads; some developers, since late September, say they have seen OAI-SearchBot and GPTBot run scripts in their own logs.

Neither can be verified from outside, and OpenAI has said nothing. This page presents them for what they are, unconfirmed reports. For a site owner the conclusion does not change: if the text sits in the initial HTML, both the bots that run scripts and those that don’t will read it. It holds either way, so you need not wait to find out which report is true.

There is also a delay that applies to Google. Its guide describes three phases, crawling, rendering and indexing: pages returning 200 are queued for rendering, which can take a few seconds or longer. Text that appears only after the script gets into the index later than text already written into the page.

How to check your site in five minutes

Open the page, press Ctrl+U and look in the source for the main text, the price and the email address: anything you can’t find there, a crawler that doesn’t run JavaScript can’t see. The same check, more rigorous, can be run from a terminal with curl followed by the page address. I describe this check in the lab note on the email address.

From a terminal: the HTML as a crawler receives it, and a word to search for in it

curl -sL https://www.yoursite.com/contact | grep -i "info@"

If the command prints nothing, the email address is not in the HTML. Repeat the test on the home page, services, prices and one product page, which are the five rows of the log in the checklist. Often the problem doesn’t affect the whole site: a catalogue, a configurator or a price list loaded by a script is enough to lose that one page.

The last check is in the logs, if your hosting shows them. Search by bot name: GPTBot, ClaudeBot, PerplexityBot. You see which pages they request and whether they also request the .js files. A bot that asks for the page and not the files certainly reads it as it is; one that asks for the files does not necessarily run them, as the study showed.

Near-duplicate pages: why AI cites the wrong version

AI systems group near-identical pages together and pick one to represent them, and it may be the old one. Bing wrote it on 19 December 2025: “LLMs group near-duplicate URLs into a single cluster and then choose one page to represent the set.” Bing talks about grouping, not a penalty.

Bing’s remedies are the canonical, consistent metadata and IndexNow to flag changes at once. Google ranks the signals by strength: redirects and canonicals are strong, inclusion in the sitemap is weak. It also advises against using noindex to choose the main page within a single site. The Copilot case is in the guide on how to show up in Copilot.

The link with JavaScript is the canonical. Google writes that it can be injected with a script but does not recommend it. If the canonical arrives only after the script, Googlebot sees it after rendering; a crawler that reads only the HTML, inferring from the study, does not see it at all. So the checklist asks for it in the initial HTML, at point 4.

Two cases measured on itria.io

The email address that wasn’t there. On 25 September 2026 the HTML served for the itria.io contact page held no address, because Cloudflare’s obfuscation only put it back with a script. Fixed the same evening; the case is in the lab note on the email address AI assistants can’t see.

Fourteen sites out of 121. In the study of food producers’ websites of 26 September 2026, 7 sites obfuscate their email address and another 7 do not write it in the code of the pages read: for an assistant, 14 companies with no address to give.

The result. Between 11 October and 7 November 2026 itria.io pages appeared 3,818 times in Google and Microsoft AI answers: 3,550 impressions in AI Overviews and AI Mode according to Search Console, 268 citations in Copilot according to Bing Webmaster Tools.

AI crawlers and JavaScript for small businesses: what to do

Put everything a customer needs to find in the initial HTML, with server-side rendering, pre-rendering or static generation, and leave scripts only for what serves people using the site. Google advises it, and so does the Vercel study. Google calls dynamic rendering, one version for bots and one for people, “a workaround and not a recommended solution”.

Before rebuilding the site, find out how big the problem is: the five rows of the log are enough to tell you. If text, prices and contact details are already there, the work lies elsewhere: an llms.txt file that introduces the company to assistants, the doors in robots.txt, pages that answer customers’ questions. The overview of techniques is in the guide to generative engine optimization.

For ChatGPT, Claude, Brave and the rest, each assistant’s doors are in the guides on how to show up in ChatGPT and in Claude; every guide is in the marketing for small businesses index. How this work is done for a real company is on the page about presence in search engines and AI assistants.

Questions and answers

Do AI crawlers render JavaScript?

Mostly not, according to the only public study, by Vercel and MERJ, dated 17 December 2024: the crawlers of OpenAI, Anthropic, Perplexity and Meta read the HTML without running scripts. Googlebot, which also serves AI Overviews and AI Mode, and Bingbot, which serves Copilot, do run JavaScript.

No owner of an AI crawler has said in its documentation whether its bot runs it.

Can ChatGPT crawl JavaScript sites?

In the Vercel and MERJ study, GPTBot, OAI-SearchBot and ChatGPT-User downloaded some JavaScript files without running them, so they read only the initial HTML. OpenAI’s page on its crawlers, read on 10 November 2026, does not mention JavaScript. Since late September 2026 there have been reports of a change, unconfirmed by OpenAI.

Do Google and Gemini render JavaScript?

Yes. Google writes that Search runs JavaScript with an up-to-date version of Chromium, after a rendering queue that can last seconds or longer. Gemini, according to Vercel and MERJ, goes through Googlebot’s infrastructure. Google-Extended is a robots.txt token, not a crawler.

How do I check whether my site can be read without JavaScript?

Open the page, press Ctrl+U to view the source and look for the main text, the price and the email address. If they are not there, a crawler that does not run JavaScript cannot see them. You can run the same check with curl followed by the page address.

Does dynamic rendering solve the problem?

Google calls it a workaround, not a recommended solution. Its advice is server-side rendering, pre-rendering or static generation, which put the content in the HTML for everyone, people and bots alike.

Notes on sources

  1. Vercel, The rise of the AI crawler, by G. Zecchini, A. A. Moore, M. Ubl and R. Siddle (MERJ), 17 December 2024: method, bots, volumes, downloaded files, 404s and redirects; the source of the quoted sentence on crawlers. A study by a hosting vendor. Read on 10 November 2026.
  2. Google Search Central, Understand JavaScript SEO basics, updated 4 March 2026: Chromium, the three phases, bots that do not run JavaScript, the injected canonical. Dynamic rendering as a workaround, updated 10 December 2025.
  3. Google Search Central, canonicals and duplicate URLs, updated 10 July 2026: strength of signals, noindex.
  4. Bing Webmaster Blog, the evergreen Bingbot, October 2019; duplicate content and AI search, 19 December 2025: the source of the quoted sentence on duplicates.
  5. OpenAI, Overview of OpenAI Crawlers, and Perplexity, Perplexity Crawlers: no mention of JavaScript, read on 10 November 2026.
  6. Anthropic, web fetch tool (the sentence on JavaScript) and Anthropic-hosted browsers, updated 4 November 2026: Claude-Agent.
  7. The 2026 reports of a change at OpenAI come from posts by individual developers and an independent test, which cannot be verified from outside; no announcement from OpenAI as of 10 November 2026.
  8. The itria.io cases: lab note of 25 September 2026 and study of 26 September 2026 on 121 sites; the figures come from Search Console (generative AI report) and Bing Webmaster Tools, window 11 October to 7 November 2026.
·Who does it for you

Putting the text in the HTML is the first step. Getting cited instead of your competitors is the work that comes after. Positioning is built starting from the market.

Mattia Esposito, who wrote this guide, also works with Itria on positioning: in-depth market analysis to find where competitors are exposed and take that ground. He did it on his own site first. In August, asked who had founded itria.io, ChatGPT couldn’t find him. From 11 October to 7 November 2026 itria.io pages appeared 13,803 times in Google and Bing results and 3,818 times in AI answers: Google’s AI Overviews and AI Mode, and Microsoft Copilot.