Where the Feed Ends and the Model Begins
OpticOdds' Matt Restivo on accountability, freshness discipline and what changes when institutional betting data reaches readers who cannot interrogate a number.
A consumer asks Perplexity Computer where a line has moved. The product calls OpticOdds, and a model turns the feed into a sentence. OpticOdds owns the price. Perplexity owns the wording. The join between them is where accountability is least defined, and it is now exposed to readers with no habit of checking whether a number is still live.
Matt Restivo runs Sports Data Services at Grandstand, the division OpticOdds sits in. He calls the split a clean division of labour rather than a gap. We asked him to show his working: how suspended markets and fast price moves are flagged, why divergence between books is not averaged away, when the correct answer is to decline, and where market information ends and something a user reads as advice begins.
There is a commercial question underneath it too. If price comparison becomes conversational, operators either look interchangeable or gain a new discovery channel. Restivo argues the second.
Gaming Eminence: You have described AI as a new distribution layer for sports-market data. What changes when consumers can access live odds, player props and line movements through a conversational interface rather than a sportsbook, comparison site or professional trading platform?
Matt Restivo: "The data has always lived in three places: inside a sportsbook's own product, on a professional trading terminal, or on a comparison site built for people who know what they're looking for. All three assume you speak the language. You know what a line move means, how to read a prop, which book is sharp on a given market.
A conversational interface drops that assumption. You ask in plain English and get an answer without learning the tool first. That's the real shift. The data isn't new. What's new is that millions of people who never had access to institutional-grade infrastructure and know-how can now read it correctly."
Gaming Eminence: When someone asks Perplexity Computer a sports-betting question, the product calls the OpticOdds API and then uses AI to formulate the response. Where does OpticOdds’ responsibility for the underlying data end, and how do you prevent the AI’s interpretation from changing or overstating what that data shows?
Matt Restivo: "Our job is the data layer: accurate, timestamped, delivered exactly as it exists in the market. What Perplexity's model does with it, how it phrases the answer, sits downstream of us. That's a clean division of labor, not a gap. It's also why we build the data to be unambiguous in the first place. Structured fields and precise numbers, not anything a model has to interpret or round."
Gaming Eminence: Betting markets can move within seconds. How does the integration handle timestamps, suspended markets, rapid price changes and differences between sportsbooks and when should it decline to answer rather than return information that may already be stale?
Matt Restivo: "Every data point carries a timestamp down to the specific market and book it came from. "The odds" isn't one number. It's a specific price at a specific moment. That precision is what tells the system when to hold back.
If a market is suspended, the feed reflects that instead of surfacing an old price as if it were live. And if the most recent price is old enough that showing it as current would mislead, the right answer is to say so, not return a number that looks authoritative but isn't. A stale number with no flag on it is worse than no number at all.
Books often genuinely disagree with each other, and we don't collapse that into a single averaged figure. The disagreement is itself information worth noting. None of this is a new standard we invented for a consumer product. It's the same discipline we've held for years with sportsbook and trading customers, where getting freshness wrong has real financial consequences. We're making it available to a new audience."
Gaming Eminence: Natural-language questions can be ambiguous, particularly around player names, market definitions and betting terminology. What technical work is required to convert an informal consumer question into the correct structured data request?
Matt Restivo: "This is exactly where clean infrastructure underneath shines. A player might show up by a nickname or a spelling variant in a casual question. Because our data model resolves every player, market, and book to one canonical entity, that ambiguity gets caught and corrected before it becomes a bad query.
Terminology works the same way. "The line," "the spread," "a prop" all map to distinct, well-defined structured fields, so a model asking a loose question has something precise to land on instead of guessing.
We designed OpticOdds to be machine-readable from the ground up, long before AI models were the ones calling it, and that discipline is what makes reliable conversion possible now. The edge cases, nicknames, regional terms, market variants, still take real engineering work. They're solvable because the underlying data was built with this kind of clarity in mind."
Gaming Eminence: How are OpticOdds and Perplexity evaluating the quality of the answers being produced? Beyond latency and data accuracy, are you testing factors such as completeness, consistency and whether users can understand how an answer was reached?
Matt Restivo: "Latency and accuracy are hard requirements for us. Fast and correct is necessary but not sufficient here, because a technically correct answer a consumer misreads is still a bad outcome.
So, here we need to look at completeness. Does the answer address what was asked, or quietly leave out a relevant market or caveat? Does the same underlying question produce the same answer no matter how it's phrased? Can a user tell where the number came from and when it was current, rather than just getting a confident-sounding sentence?
That last one matters more here than it did for us historically. Our professional trading customers know how to interrogate a number when something looks off. A consumer asking Perplexity doesn't have that instinct, and shouldn't need it. We want to be able to help someone without that background look at an answer and understand not just what it says, but why it's trustworthy."
Gaming Eminence: This is OpticOdds’ first consumer-facing deployment. What regulatory and responsible-gambling considerations arise when professional-grade betting data becomes available through a general AI product, and how do you distinguish objective market information from something a user may interpret as betting advice?
Matt Restivo: "OpticOdds doesn't predict outcomes or make recommendations. We provide the data the market has already generated. That distinction has to hold even more carefully in a consumer product than it did in a B2B one, where our customers already knew how to read a number in context.
In practice, what shows up is market information: a price, a line move, a probability the market implies. Not a suggestion about what to do with it. A line move tells you what the market believes. It doesn't tell you what to do.
This is one of our first consumer-facing deployments, and we're approaching it with real humility. The infrastructure is proven. The audience is new. We're treating that as something to keep getting right as the launch matures, not a box we've already checked."
Gaming Eminence: Easier price comparison could change how consumers discover and assess sportsbooks. Does this risk making operators appear interchangeable except for price, or could conversational AI become a meaningful new acquisition and engagement channel for them?
Matt Restivo: "The sportsbooks that win aren't winning on price alone. They win on product, trust, and experience. Price transparency raises the bar on all three rather than flattening it.
The more likely outcome is that conversational AI will become a new discovery layer. Nikesh Arora brilliantly laid out the disintermediation of analytical SaaS companies at Liquidity 2026. I think this is an example of people now being able to get answers far faster than they ever could have before. It's similar to what happened when odds-comparison sites first started showing up. Some operators worried about commoditisation. The ones who leaned in built real acquisition channels from it probably learned the most instead of complaining and losing ground."
Gaming Eminence: OpticOdds now has two different routes into AI: a direct integration within Perplexity Computer and an MCP connector that customers can use with Claude. Which model do you expect to become more important commercially, embedded partnerships with major AI products or a broader ecosystem in which businesses connect OpticOdds data to their own agents and workflows?
Matt Restivo: "I don't think it's either/or, and I'd push back on the framing that we have to choose. The MCP work is about being the infrastructure any AI agent can call. That's a platform bet, and it compounds the more the ecosystem adopts open standards. The Perplexity integration is a deep, productised partnership that reaches a huge existing user base immediately, with nobody needing to set anything up.
Commercially, the embedded partnerships prove the model works and get the data in front of people right now. The open ecosystem is the longer-duration bet. It's what makes us infrastructure for the whole category rather than a feature of any one product. Partnerships like Perplexity are the near-term proof points."
