VISTACITE
All docs

Concepts

What is the Crystal Ball feature?

Every audit asks Claude, verbatim, what it knows about your domain. If the answer is "nothing," that's the entire diagnosis in one sentence.

Last updated April 9, 2026

The Crystal Ball is the simplest feature in VistaCite and it's also the most useful one. Before we run any scored checks, we ask Claude this question about the domain we're about to audit:

What is {domain} known for? Answer in exactly 2 sentences. If you do not have reliable information about this domain in your training data, say so directly in the first sentence — do not guess or speculate.

The answer is your Crystal Ball. It's shown verbatim on the results page and emitted into the downloaded .fix_it_plan.md file. We do nothing else with it — no scoring, no transformations, no cleanup.

Why it's the most useful thing we do

Most on-page audit tools measure whether your page has the right HTML signals. Crystal Ball measures something different and more important: does the AI actually know you exist?

For about a third of the sites we've tested, the answer is "no." The LLM says something like "I don't have reliable information about this domain" or "I'm not familiar with this site." That's not a problem with your HTML — it's a problem with the training data. Every on-page fix you make after that is necessary but not sufficient, because even perfect AEO and GEO scores don't help you if the model has never heard of your brand.

When the Crystal Ball fails, the diagnosis is:

  1. Your brand has zero or very weak presence in high-authority sources the LLM's training data includes (news, Wikipedia, academic papers, high-DR blogs, podcasts, forums the model indexed).
  2. No amount of on-page optimization alone will fix this — you need off-page signals.
  3. The path forward is getting mentioned (not just linked) in those high-authority sources.

Why we use Claude specifically

We use Anthropic's Claude for the Crystal Ball query because it has a published training-data cutoff and a consistent, well-documented behavior when it doesn't know something: it says so plainly. Other models tend to hallucinate plausible-sounding answers for domains they've never seen, which makes them useless for this specific question.

The fact that Claude IS an LLM doesn't mean its answer represents "what all AI engines think." It represents what Claude specifically knows. ChatGPT, Perplexity, and Gemini will have different answers based on their own training data. In practice they correlate tightly for well-known domains (if Claude knows you, ChatGPT usually does too) and for unknown ones (if Claude doesn't, the others usually don't either). For borderline cases, Claude is a conservative proxy.

Why we don't score it

Technically we could score the Crystal Ball: +100 if the LLM wrote a substantive answer, 0 if it said "I don't know." We chose not to because:

  • The result is already unambiguous without a number attached.
  • Scoring the Crystal Ball would double-count it with the on-page AEO score, which is the wrong framing — the Crystal Ball is an off-page signal.
  • Presenting it raw, verbatim, with the model's own hedging language, is more useful as a diagnostic than any summary we could compute.

What to do if your Crystal Ball is red

The strongest off-page signals for getting mentioned in LLM training data:

  • Wikipedia — the single biggest lever. If your brand qualifies for a Wikipedia article (notability threshold), an active page is the strongest possible training-data signal.
  • Published research, whitepapers, or data that other sites cite. Appears in scraped academic corpora + news aggregators.
  • Podcasts and interviews where your brand is named verbally — these end up transcribed and scraped.
  • High-DR tech publications (TechCrunch, The Verge, Stratechery-class newsletters) that include your brand in trend pieces.
  • Trusted industry publications in your vertical, even without a huge DR.
  • GitHub if you have an open source project — GitHub is heavily scraped by every major LLM.

What doesn't help:

  • More backlinks without brand mentions (the model has to ingest text about you, not just links)
  • SEO content marketing on your own blog (your own site usually isn't in the training data at the same weight)
  • Social media posts in isolation (posts decay; lasting mentions live in static HTML)

Related