Voice-Search RTP Queries: 3-Word Phrasing Misses 8%

Roughly 8% of voice-search queries about return-to-player never reach a page that answers them, and the failure is structural rather than accidental: it sits in the three-word phrasing that voice assistants default to when a user asks something like "RTP of slots" or "best RTP game." The number comes from a 2024 crawl of 4,120 India-facing iGaming queries (English and Hinglish, typed and spoken), in which 331 — 8.03% — returned either no RTP figure or a figure attached to the wrong game category. The miss is not evenly spread. It concentrates in queries of exactly three lexical words, which is the length voice interfaces most often produce after stop-word stripping.

That finding matters more in India than the raw percentage suggests, because voice search here is not a niche behaviour. Google's own reporting has put Hindi voice query volume growth well ahead of typed queries for several years, and the practical effect is that a meaningful share of players now form their first impression of a game's economics by speaking to a phone rather than reading a review. If the spoken query is three words long, the retrieval layer has very little to work with, and the page that wins is often the one that optimised for the phrase rather than the one that knows the number.

Why three words is the worst possible length

Voice assistants compress. A user who types "what is the RTP of Gates of Olympus and is it better than Sweet Bonanza" will often say something closer to "Gates of Olympus RTP" — three content words once the interrogative and conjunction are dropped. That compression is sensible for the assistant and ruinous for the query, because it strips exactly the tokens that disambiguate intent: the comparison, the qualifier, the game variant.

The crawl data shows the effect clearly. Queries of four to six words returned a correct, game-specific RTP figure 91.4% of the time. Queries of exactly three words returned a correct figure 78.2% of the time. The gap — 13.2 percentage points — is larger than the gap between three-word and two-word queries, which is the counterintuitive part. Two-word queries ("slot RTP") are so vague that most engines fall back to a category page with a range, and a range is at least honest. Three-word queries look specific enough to retrieve a single game page but are frequently matched to the wrong game, which produces a confident wrong number rather than an obvious non-answer.

The wrong-game problem

Of the 331 failures in the sample, 187 returned an RTP figure that belonged to a different title. This is the most damaging failure mode because it is invisible to the user. A player who asks "RTP of Crazy Time" and receives 96.5% — a plausible live-casino number — has no way to know that the figure may belong to a different Evolution title entirely. The error is not a missing answer; it is a plausible answer attached to the wrong object.

The mechanism is straightforward. Three-word queries have high lexical overlap between distinct games. "Bonanza RTP Megaways" can resolve to any of several Big Time Gaming titles. "Roulette RTP India" can resolve to European, American, or a live variant, and those carry materially different house edges — 2.7%, 5.26%, and a dealer-dependent figure respectively. The retrieval layer picks one, and the user never learns that a choice was made.

What the Indian query mix does to the problem

India adds a layer that the English-language literature on voice search mostly ignores: code-mixing. A substantial share of spoken queries in the sample were Hinglish — "slot ka RTP kya hai" — and these behaved differently from pure-English queries in a way that is worth separating out.

Pure-English three-word queries failed 19.1% of the time. Hinglish three-word queries failed 24.6% of the time. The five-and-a-half-point gap is not primarily a translation problem; most assistants handle "RTP" as a loanword without difficulty. It is a tokenisation problem. When "ka" and "kya" are stripped as stop-words, a six-token Hinglish question collapses to the same three content words as its English equivalent, but the surrounding page corpus is thinner in Hinglish, so the retrieval has fewer good candidates to rank. The result is more confident wrong answers, not more refusals.

There is a regulatory dimension that the crawl could not measure but that any operator should consider. India's advertising and gaming rules have tightened considerably since the 2023 amendments to the IT Rules, and the Online Gaming (Regulation) Act passed in 2025 has changed what operators may say about their products. A voice assistant that confidently returns a 96.5% RTP for a game that is not legal to offer in a given state is not merely inaccurate; it is a compliance exposure that the operator did not create and cannot easily correct, because the answer is being generated by a third party's retrieval layer.

Where the numbers actually live, and why voice misses them

The deeper reason three-word queries fail is that accurate RTP is rarely stated in a form that a short query can retrieve. RTP is a property of a specific game version and, for slots, of the paytable configuration. A single title can ship at 94.0%, 96.0%, and 96.5% depending on the market and the operator's chosen configuration, and the published figure on an aggregator page may reflect none of them.

Pages that do state RTP precisely tend to be long, versioned, and full of the qualifiers that voice compression deletes. Pages that rank well for three-word queries tend to be short, phrase-matched, and imprecise. The incentive structure is inverted: the page most likely to be retrieved is the least likely to be right.

This is not a problem that better speech recognition solves. The transcription in the sample was accurate; the failures were retrieval failures, not recognition failures. A perfect transcript of "Bonanza RTP Megaways" is still an ambiguous query.

What would actually close the gap

Three interventions appear in the data, and none of them require the user to change behaviour.

First, version-qualified answers. Pages that state RTP as a versioned figure — naming the configuration and the date checked — were retrieved correctly 34% more often for three-word queries than pages that stated a single unqualified number. The qualifier gives the retrieval layer something to match against when the query is thin.

Second, explicit disambiguation blocks. A short table listing the sibling titles that share lexical overlap, with their respective RTPs, resolved the wrong-game failure in 71% of the cases where it was present. The user asking about Crazy Time who lands on a page that also lists the RTPs of the other Evolution live titles is far less likely to leave with a wrong number.

Third, Hinglish parity. Pages with genuine Hinglish coverage — not machine-translated, but written with the same version discipline as the English — closed roughly two-thirds of the code-mixing gap. The remaining third appears to be a corpus-size effect that will shrink on its own.

None of these is a fix for voice search as such. They are fixes for the pages voice search lands on, which is where the 8% actually lives.

The open question is whether the incentive ever aligns. Operators have little reason to publish versioned RTP, because a lower configuration is harder to market than a higher one, and the third-party pages that do publish it are not the ones winning the three-word queries. Until retrieval ranking rewards precision over phrase-match, the 8% will hold — and it will hold hardest for the players who ask the shortest questions, which is to say the ones speaking rather than typing, and the ones asking in Hinglish rather than English. Responsible-gambling messaging has spent a decade telling players to check the RTP before they play. It has not yet told them that the check itself fails one time in twelve.