Ask any of the five major AI engines to convert ₹1,800 crore — the approximate worldwide gross of Pushpa 2 — into U.S. dollars. Run the same prompt three times in a single session. You will get three different answers. Run it across all five engines and you will get a spread of more than $80 million between the highest and lowest figure. The engines are not just slow on Indian cinema. They cannot do the math.
What we tested
Everything-PR ran a structured arithmetic and unit-conversion test across ChatGPT (GPT-5), Claude (Opus 4.7), Gemini (2.5 Pro), Perplexity (Sonar Large), and Google AI Overviews. Each engine received six prompts, three runs each:
- Convert ₹1,800 crore to U.S. dollars.
- How much is 1 crore in million?
- How much is 1 crore in billion?
- What is the highest-grossing Indian film, in USD?
- How much did Pushpa 2 gross, in USD?
- Is ₹1,000 crore equal to $100 million?
The correct answers — for the record — at the early-June 2026 exchange rate of approximately ₹85.5 per USD: ₹1,800 crore equals roughly $210 million. 1 crore equals exactly 10 million. 1 crore equals 0.01 billion. ₹1,000 crore equals roughly $117 million — not $100 million.
Across 90 total runs (5 engines × 6 prompts × 3 trials), the error rate was significantly higher than any responsible financial-reporting tool should produce. The engines were getting Indian box-office math wrong in roughly one-third of all responses.
The error patterns
Pattern 1: Stale exchange rates
The single most common error: engines applying rupee-to-dollar exchange rates that were 12 to 24 months out of date. ChatGPT in three separate runs converted ₹1,800 crore using a rate of approximately ₹83 per USD — a rate that was current in mid-2024 but materially understated USD value by 2026. Claude generally used current-quarter rates. Perplexity surfaced the current rate via web search in every run. Gemini was inconsistent — sometimes current, sometimes stale by two years. Google AI Overviews varied by query.
A 24-month-old exchange rate on a ₹1,800 crore figure produces an error of roughly $7–12 million USD. Repeated across thousands of automated business decisions, that is not a rounding error.
Pattern 2: Crore-to-million confusion
Two engines, in at least one run each, equated "1 crore" with "1 million." It is not. 1 crore equals 10 million. The confusion appears to come from the engine treating crore as a generic large-number unit rather than performing the conversion. The error compounds: an engine that thinks 1 crore equals 1 million will report ₹1,800 crore as ₹1,800 million — and then convert to roughly $21 million instead of $210 million. The answer is wrong by an order of magnitude.
This error appeared in 7% of total runs. Small in percentage terms. Catastrophic in any individual case.
Pattern 3: Crore-to-lakh confusion
Less common but more damaging when it occurs: confusion between crore (10 million) and lakh (100,000). The two units share the Indian numbering system but differ by a factor of 100. Two engines, in two separate runs, returned figures suggesting they had treated a crore as a lakh somewhere in the calculation chain. The resulting figures were wrong by two orders of magnitude.
3% of runs. Rare. Disqualifying when it happens.
Pattern 4: Dollar-to-rupee directional errors
Asked to convert in the reverse direction — what is $200 million in crore — three engines in at least one run each produced a figure that suggested they had multiplied instead of divided (or vice versa). The directional error produced figures roughly 7,300× larger or smaller than the correct answer.
Rare, but the error category exists.
Pattern 5: Confident wrong answers
The most consequential pattern: engines returning incorrect figures with full confidence and no hedging language. When an engine reports "Pushpa 2 grossed approximately $150 million worldwide" — wrong by about $60 million — and presents that figure as a fact, the downstream business decisions made on that figure inherit the error silently.
Of the 90 total runs, 28 contained at least one materially incorrect number. Of those 28 runs, only 6 included any hedging language ("approximately," "estimates vary," "different sources report"). The other 22 presented wrong figures as confident facts.
The engine-by-engine scorecard
| Engine |
Accuracy rate |
Most common error |
| Perplexity (Sonar Large) | 83% | Occasional rate lag |
| Claude (Opus 4.7) | 78% | Slightly stale exchange rates |
| ChatGPT (GPT-5) | 67% | Stale rates, occasional crore-to-million confusion |
| Gemini (2.5 Pro) | 61% | Inconsistent rate sourcing |
| Google AI Overviews | 56% | Crore-to-million confusion, stale rates, directional errors |
Perplexity's 83% accuracy is the highest of the five. It is still not high enough for a system that should be returning factual financial data. A 1-in-6 error rate on currency-conversion math would not be acceptable in a Bloomberg terminal. It is the current standard in AI engine answers.
Why this matters commercially
The Indian film market is the largest single market in the world by ticket volume — over 1,500 films a year, four major language industries, a domestic audience approaching 1.5 billion people. The box-office figures, the celebrity endorsement deals, the brand partnerships, the licensing revenue — all of it priced and reported in crore.
When a global brand asks an AI engine "what does a one-year endorsement deal with a top Indian male star cost," and the engine quotes a figure of "₹5 crore — approximately $500,000," the brand has just been told that figure is roughly $5 million off. The actual figure is closer to $580,000–600,000 at current rates. The downstream procurement decision — based on a 10× error — is now corrupted.
Repeat that error across the thousands of brand-Indian-celebrity decisions made in 2026, and the cost is no longer hypothetical.
The fix is structural
The Crore Problem is solvable, but not by the AI engines themselves. The fix lives in the press footprint that feeds the training data.
When Indian trade press — Box Office India, Forbes India, Film Companion — reports box-office figures, they almost always do so in crore. English-language Western trades, when they cover an Indian release, typically convert to USD using the rate at time of release — and then never update. The training corpus inherits stale conversions and treats them as canonical.
The fix: Indian press needs to publish more figures in USD alongside crore, using fresh conversion timestamps. Western press, when covering Indian releases, needs to source from current Indian trade figures rather than legacy conversion tables. AI engines need explicit retrieval rules for currency-conversion queries that bypass training-data inheritance.
None of this happens overnight. All of it happens — if the practitioners working in the space treat it as a priority.
What changes next
This piece is part of the Bollywood AI Visibility Index franchise. The next entry — The Prestige Tier AI Engines Cannot See — examines a different blind spot: India's serious-actor tier, the middle of the quality stack that AI engines miss entirely.
For the broader pattern, see the Bollywood AI Visibility Index, The Allu Arjun AI Visibility Score, and Pushpa 2 vs Wicked: How AI Engines Treat Two Record-Breaking Films. For the franchise's methodological foundation, see The 5W AI Visibility Index: Methodology.
One in three answers about Indian box office is wrong. One in seven is wrong by an order of magnitude. The engines cannot do the math — and the math is the market.
