We pulled 8,400 brand-pricing answers — three models, 2,800 brands, one question each: what does this cost? It is the single highest-disagreement category we measure, by a wide margin. Models miss pricing roughly 6.5× more often than they miss founding facts.
Why pricing is hard for models
Pricing breaks AI for reasons that are structural, not incidental:
- It's the fact most likely to be hidden. "Contact us" boxes, gated quotes, and PDF rate cards are invisible to a model. No machine-readable price means no correct answer.
- It changes. A founding date is true forever. A price is true until the next pricing-page update, and stale cached pages outlive the change.
- It's expressed inconsistently. $49/mo, $588/yr, "starting at $49," "from $49 per seat" — all the same price, all parsed differently.
What the three models do when they don't know
This is where the providers diverge, and the divergence is diagnostic.
| Model | Most common failure mode |
|---|---|
| ChatGPT | Confabulates a plausible price |
| Gemini | Quotes a stale cached price |
| Perplexity | Refuses — "I can't find pricing" |
Perplexity's refusal is the honest failure. The other two will hand a user a wrong number with full confidence, which is worse than saying nothing — because the user acts on it.
How to fix it
The fix is unglamorous and reliable: make your price machine-readable. Product schema with priceCurrency and priceValidUntil. A plain pricing table, not a "Contact us" wall. A line in your llms.txt pointing at the canonical pricing page. When Christmas or Not did exactly this, pricing-question accuracy moved 6.5× in a single scan cycle.
The lesson generalizes: the questions models get most wrong are the questions whose answers you've hidden behind a form. Unhide the answer and the consensus follows.