CrawlCheck

Directory · Inference and GPU compute

Vast.ai — what a crawler receives from vast.ai

Website measurement, not a product review. This page records what machine clients received from vast.ai. It does not evaluate Vast.ai software, features, pricing or support.

89
Agrade

AI assistants like ChatGPT can read your site, but find little to quote about your product.

AI visibility 89/100, measured 2026-10-27 14:34 UTC from outside its network, as each named AI and search crawler and as a browser. An engine can reach and read this site. What is left is giving it something specific to say.

Open the full report vast.ai ↗

961. ReachCan a named answer-engine crawler get your pages at all?882. ReadOnce it has the bytes, can it find the words?793. QuoteIs there a specific fact it can state and attribute?

3

findings

✓

llms.txt on its host

✗

entity map on its host

8

scans on record since 2026-09-18

The three stages, in the order an engine hits a site

Each stage gates the next: an engine that cannot reach a page never reads it, and one that cannot read it never quotes it.

96

Stage 1 · 40% of the score

Reach

Can a named answer-engine crawler get your pages at all?

Every named crawler was served the same page a browser gets.

88

Stage 2 · 30% of the score

Read

Once it has the bytes, can it find the words?

The delivered HTML is mostly content, and the machine files point at the rest.

79

Stage 3 · 30% of the score

Quote

Is there a specific fact it can state and attribute?

An assistant would have to paraphrase your page instead of quoting a fact.

The stage to fix first is quote, the lowest of the three.

What to fix

3 findings on this site. Every one is named below; the top one is shown in full, free. What we saw, why each matters, how to fix it and the bytes behind it are what a licence buys — or we implement the whole list for you.

1 Low 2 Note

1The instruction files contradict each otherContent-Signal under User-agent: * says ai-train=no, but GPTBot, ClaudeBot, Google-Extended each have their own group that allows them in with no signal of theiLow

What we saw

Content-Signal under User-agent: * says ai-train=no, but GPTBot, ClaudeBot, Google-Extended each have their own group that allows them in with no signal of their own. A crawler reads only its own group; the refusal never reaches them. /robots.txt

Why it matters

A crawler obeys only its own User-agent group. A Content-Signal or Disallow written under * says nothing to an agent that has its own group, and a group that both invites and forbids leaves the crawler to pick. Put the signal in every group it is meant for, and let the rules under it agree.

How to fix it

Read the evidence below, then change what the server sends at that path. Re-scan to confirm: done is measured, not asserted.

On Next.js (detected)

  1. App Router: add app/robots.ts that returns a MetadataRoute.Robots object (rules per userAgent, sitemap URL). Pages Router or static export: put robots.txt in public/.
  2. Ship one of the two, not both: the route handler and a public/robots.txt collide.

All Next.js fixes

Site deployed from GitHub

  1. Add this workflow file as .github/workflows/crawlcheck.yml (change robots-path if robots.txt lives elsewhere).
  2. Run it from the Actions tab: it opens one pull request with the corrected robots.txt, every change listed. Review, merge, deploy.

How the action works

Fix it with a coding agent

Paste into Claude Code, Cursor or Codex, opened on the code that serves this site.

Show prompt
You are working in the codebase (or server/CDN configuration) that serves vast.ai. A crawler audit found the problem below. Fix it at the source, keep the change as small as possible, and do not change anything a visitor sees unless the fix requires it.

Finding: The instruction files contradict each other (INSTRUCTION_CONFLICT, low severity)
Page: https://vast.ai/robots.txt
What was measured: Content-Signal under User-agent: * says ai-train=no, but GPTBot, ClaudeBot, Google-Extended each have their own group that allows them in with no signal of their own. A crawler reads only its own group; the refusal never reaches them.
Why it matters: A crawler obeys only its own User-agent group. A Content-Signal or Disallow written under * says nothing to an agent that has its own group, and a group that both invites and forbids leaves the crawler to pick. Put the signal in every group it is meant for, and let the rules under it agree.
How to fix it: Change what the server sends at that path so the finding no longer applies.
Detected platform: Next.js. Make the change where this platform controls it.
Rule and examples: https://crawlcheck.io/rules/INSTRUCTION_CONFLICT
Full report (markdown): https://crawlcheck.io/r/47rhqyi9d5.md

When you are done:
1. Tell me which files or settings you changed and why.
2. If the fix is a server, CDN or cache setting you cannot change from the code, say exactly where it is set.
3. After deploying, re-scan at https://crawlcheck.io/?url=vast.ai . The fix counts when INSTRUCTION_CONFLICT is gone from the new report.
Why this was decided
RuleINSTRUCTION_CONFLICT · rule pack sv34 · scope core
Evaluated2026-10-27 14:34:58 UTC
Decisiond1:7a556adb7f30ab87f062 · finding f1:c0d362cbb9516ad752e5 resolve either at /api/graph/<id> with a key
Observercrawlcheck-worker build c1d5ec39 from EWR · audit v4

What it rests on

IdentityStatusBytessha256Kept
read: /robots.txt2001,280baf5aec80bdaalready_held
read: /llms.txt2006,1364dd080c06749already_held
read: /200101,440ff00e279cce5kept

22 more in the manifest.

Artifacts: manifest:bc8a5070114dd9bd049de816d575d85d6edef10eec39685d047af1518ce1f6ce

Grade effect Does not cap the letter: only medium, high and critical findings do. Moves no points either; the score is built from the measured rows.

Rule INSTRUCTION_CONFLICT under scoring rules v34 · fingerprint INSTRUCTION_CONFLICT|/robots.txt · persisting (8 scans)

Verify the fix Re-scan after the change. This finding is cleared when INSTRUCTION_CONFLICT is no longer reported at /robots.txt; if it comes back after clearing, it is marked a regression, not a new finding.

INSTRUCTION_CONFLICT · full transcript

2The homepage is reached through a chain of redirects🔒 What we saw, why it matters, the fix and the evidence bytes are in the licensed report.Note
3A business record gives an address but no coordinates🔒 What we saw, why it matters, the fix and the evidence bytes are in the licensed report.Note

Open the other 2

What we saw, why it matters, how to fix it, and the evidence bytes for every finding — on this report and every re-scan. From $29 a month.

Unlock the findings

Or have them fixed

We implement the list on your site and re-measure, with a sealed before and after. One site, $100 once.

Get this fixed — $100

The chain a crawler follows

The chain breaks at entity graph. Everything after that point is only reachable by a crawler guessing the conventional path.

✓robots.txtEvery crawler reads this first
✓sitemapNamed in robots.txt and resolves
✓llms.txtPoints onward, links stay on this host
✗entity graphParses as JSON and points home

A machine file that does not parse is worth less than one that is absent, because it looks present. Rows behind this.

Where the bytes go

Of the 101,410 decompressed bytes the homepage delivers, 5.5% is text a reader or a model can actually use. Most of what a crawler downloads here is not words.

  • Readable text 5.5% · 5,548 B
  • Markup & attributes 73.4% · 74,412 B
  • Structured data (JSON-LD) 2% · 2,022 B
  • Inline CSS 4% · 4,029 B
  • Inline JavaScript 4.4% · 4,503 B
  • Inline SVG 10.7% · 10,840 B
  • HTML comments 0.1% · 56 B

The individual blocks that weigh the most, so the fix is a search rather than a hunt:

  • Inline SVG · 3,709 B (3.7%) — #vast-logo-R6ih6mopens <path d="m1.052 16.916 9.539 2.552a21 21 0 0 0 .06 2.033l5.956 1.593a1
  • Inline CSS · 2,354 B (2.3%) — #__jsx-d683830613966beeopens .help-launcher-button-glass.jsx-d683830613966bee{background:rgba(255,2

What changed

  1. 2026-10-27the stale edge copy clearedresolved
  2. 2026-10-13the homepage began taking more than one redirect to reachappeared
  3. 2026-10-02score moved 84 → 85score-up
  4. 2026-10-02a stale edge copy was served againreturned
  5. 2026-09-24the stale edge copy clearedresolved
  6. 2026-09-19a stale edge copy was servedappeared

8 scans on record, first 2026-09-18. Score movements are shown only between readings taken under the same score version; finding codes are comparable across all of them.

Is this your site?

This listing reads the public record. Claim it and Watch keeps that record on your terms: every list this page holds back, finding age, and twice-daily re-measurement. Free for 30 days, no card.

Claim the record for vast.ai

Is this your company?

This listing came from a fetch, not a submission, so nobody was asked. If vast.ai is yours you can prove it in a minute and for nothing — a DNS TXT record or one meta tag on the home page, both checkable by anyone.

Claiming marks you as the verified owner and lets you correct the category and the company name. It moves no score, no rank and no position.

Claim this listing — free or ask about a verified profile

Dispute or re-scan

If a reading here is wrong, it is our instrument that is wrong, and we want to know. Tell us what is wrong and the site is re-measured; the result is published as measured. Nothing about a listing, a payment or a request changes a number.

By email instead: hello@crawlcheck.io. Category “Inference and GPU compute” is our label for what the product is primarily sold as; it is not scored. Back to the directory.