TL;DR
- JavaScript can hide key facts from a reader that does not run scripts. The risk depends on what the server sends, not just which framework built the site.
- A homepage study found large text gaps on some sites, while the median gap was small. Test the pages that answer your buyers' questions.
- A browser view, a raw response and a saved AI answer are different records. Each resolves a different question.
- Keep product scope and limits in readable page text. Treat a rendering fix as an access improvement until later evidence shows an answer change.
Start with the fact the buyer needs
This matters for generative engine optimization (GEO) when a buyer needs to check a product's fit. A polished browser page may show a support limit that a plain fetch misses. Start with the missing fact and its delivery path. A framework name alone does not tell you whether that fact is available.
What Does a No-JavaScript Test Actually Show?
QuanticData's August 2026 study compared plain HTTP and rendered homepage views. It started with 800 top-ranked domains; 374 returned 200 on both requests and formed the analysis sample. In that group, 17.4% lost more than half their words without JavaScript. Readable content existed only with scripts on 6.4%.
| Homepage finding | Share of the 374-site sample |
|---|---|
| Lost more than half its words | 17.4% |
| Lost more than 80% of its words | 11.5% |
| Readable content only with JavaScript | 6.4% |
Source: QuanticData, August 25, 2026. One plain and one rendered homepage fetch per included domain. These groups overlap; do not add them together.
The median plain-to-rendered word ratio was 0.99. A large loss was not the typical result. This was a one-day homepage comparison, not a B2B product-page survey or a test of AI citations. Failed requests were excluded. It cannot tell you how often buyers' answers miss a fact.
Word totals are a useful alarm, but they are a blunt test. A page could retain most of its navigation and marketing copy while losing the one limit that matters to a buyer. Review the actual claims, not only the size of the text gap.
QuanticData, August 25, 2026. One plain and one rendered homepage fetch; 374 domains returned 200 in both views out of an initial top-800 selection. Groups overlap and must not be added. This is not an AI citation test. QuanticData, August 2026.
Does Every Search Reader Need the Same Fix?
No. Vercel and MERJ's December 2024 analysis observed named AI crawlers that fetched scripts without running them. On nextjs.org, JavaScript files made up 11.50% of the reported ChatGPT crawler's requests and 23.84% of Claude's. A script request did not establish that the script ran.
| Direct source | Observed result | What it helps explain | Limit |
|---|---|---|---|
| QuanticData, August 2026 | 17.4% lost more than half their words | A plain fetch can miss page text | Selected homepages, not AI answers |
| Vercel and MERJ, December 2024 | JavaScript fetch shares of 11.50% and 23.84% | Downloading code differs from running it | Historical crawler observations, not every current AI mode |
| Google Search documentation | Google Search uses Chromium to render pages | A plain fetch is not its only view | Rendering and indexing still have conditions |
These sources cover different sides of the problem. QuanticData measured its own paired page fetches; it cites Vercel for crawler behavior. It is not an independent retest of those crawlers. Nor should a historical crawler result become a claim that no AI tool can use a browser today.
Google's current guide describes a rendering queue and supports server rendering because not all bots can run scripts. That favors a clear text response for core facts. It does not justify stripping useful features out of a website or promising that one delivery method wins citations.
Which Missing Facts Deserve Attention First?
Start with the facts that could change a buying decision. Compare the same URL and page state, keeping the time, locale and access conditions. Ask the web team to distinguish text in the server response from text created by scripts. Keep that check separate from a tool's later use of the page.
The table below is an editorial review aid. It asks for visible evidence of a fact, rather than a percentage of words retained.
| Buyer fact | Compare in both page views | Why a mismatch matters | Useful completion check |
|---|---|---|---|
| Supported products | Exact model or plan names | A buyer could assess the wrong offer | Names and linked details agree |
| Scope limits | Exclusions beside the relevant claim | A broad headline can overstate fit | The qualifier survives with the claim |
| Requirements | Data, system or site prerequisites | A buyer may miss a dependency | The same prerequisites are readable |
| Next step | Correct page links and route | Missing links can block verification | The linked evidence opens correctly |
Illustrative example: a software page says that the product supports single sign-on. A script adds a note that this feature needs the Enterprise plan. The plain response has the broad claim but lacks the plan limit. The problem is not an empty page; it is an incomplete product claim.
A useful repair would put the approved claim and its exclusion together in the server-sent page text. The product owner checks the wording against the specification. The web team then compares both views again and confirms that the same condition is readable. The team has repaired that mismatch. It has not yet measured a change in AI answers.
This is more precise than a request to “make the site AI-ready.” Use the product evidence in your docs to define what must survive. Our AI crawler traffic discussion explains why more requests alone do not settle whether a buyer saw it.
JavaScript and B2B AI Search: Common Questions
Must we replace a React or Next.js website?
Not on this evidence. Those tools can send readable HTML from the server. First identify a real content gap. A focused change to how a page delivers core facts may be enough; architecture decisions belong with the web team.
Does a 200 response mean the product facts are available?
No. A successful response can contain a shell, the full text or a partial account. Inspect the relevant content. Check a failed response separately, since it may involve access or server errors rather than rendering.
Does readable HTML guarantee a Google AI citation?
No. Google's AI feature guide requires a supporting page to be indexed and eligible for a search snippet. Meeting requirements does not guarantee crawling, indexing or display. Other products need their own checks.
What should we save after fixing the page?
Save the URL, test conditions, both page views and the approved facts that now match. If you later test AI answers, save the prompt, product mode, date and exact answer too. Report the access repair and the answer result separately.
Discuss the facts missing from a page view
If your web and marketing teams disagree about what AI search can read, bring the affected URLs, approved product facts and paired page captures to a SemShift AI visibility audit discussion. We can discuss which gaps matter to buyer questions and how to define a useful verification scope.