AI Deal Matching, Explained

Keyword alerts are fragile. If you set an alert for "Sony WH-1000XM5" and a deal is posted as "Sony WH1000XM5 Wireless Headphones," you might miss it because of a missing hyphen. Most deal notification systems rely on exact text matching, which fails when merchants or community posters use different naming conventions, abbreviations, or common misspellings.

If you want a specific pair of noise-canceling headphones, you don't care if the title says "XM5," "1000XM5," or "WH-1000XM5." You want the deal. But traditional systems require you to anticipate every possible variation. If you miss one, the deal passes you by.

We built DropItLow to solve this through AI deal matching. Instead of looking for exact strings of characters, our system tries to understand the intent behind your alert and the reality of the deal being offered.

The Problem with Keywords

The core issue with keyword-based alerts is that language is fluid, but code is often rigid. A user's intent is a concept, not a string. When you say you want a "75-inch Samsung TV under $800," you're describing a set of parameters: a specific brand, a minimum screen size, and a maximum price.

A keyword system sees "75-inch Samsung TV" and looks for that exact phrase. If a deal is titled "Samsung 75" QLED 4K Smart TV," the keyword system might fail because "inch" was replaced by a double quote, or because the order of the words changed.

This leads to a frustrating user experience. You either set alerts that are too specific and miss out on deals, or you set alerts that are too broad and get buried in irrelevant notifications.

How DropItLow Matches Deals

We take a multi-layered approach to matching that combines semantic understanding with deterministic rules. The goal is to catch the deals you want while filtering out the noise you don't.

1. Normalizing the Noise

When a deal comes in from a deal feed or retailer listing, it's usually just a messy block of text. Our first step is normalization. We parse the title and description to extract structured signals:

• Merchant (e.g., Best Buy, Amazon, Newegg)
• Product Name
• Price
• Category

By turning a raw string into a structured object, we can apply specific logic to different parts of the deal. We know that "$249" is a price, not a model number.

2. Parsing User Intent

When you create an alert in plain English, we don't just save the text. We parse it into a structured intent. If you type "I want a MacBook Pro with 16GB RAM for less than $1500 from B&H," the system identifies your preferences for the product, specific specs, price cap, and preferred merchant.

This structured intent allows us to go beyond simple text search. We can compare your price cap of $1500 directly against the normalized price of an incoming deal.

3. Vector Similarity

This is where the AI deal matching happens. We represent both the deal and your alert as high-dimensional vectors. These vectors capture the semantic meaning of the words, not just the characters.

In this vector space, "Sony WH-1000XM5" and "Sony WH1000XM5" are very close together, even if they aren't an exact character match. "Noise-canceling headphones" and "Bose QuietComfort" would also have a degree of similarity because the system understands they belong to the same category of products.

We calculate the distance between your alert vector and the deal vector. If they are close enough, it's a potential match.

4. Keyword and Deterministic Scoring

Vector similarity is powerful, but it isn't perfect. It can sometimes be too broad. To fix this, we layer deterministic scoring on top.

We look for specific keywords that must be present or must be excluded. If you specifically said "no refurbished," we check the normalized deal data for any mention of "refurbished" or "renewed." If we find it, the match is rejected, regardless of how high the vector similarity score was.

We also apply strict filters for price and merchant. If a deal is $1 over your limit, it's not a match.

5. Final Combined Score

The system takes the semantic similarity score and adjusts it based on the deterministic checks. This combined score determines whether we send you a notification. This hybrid approach ensures we catch the semantic variations while maintaining the precision of traditional filters.

A Concrete Example

Imagine you have an alert for "Sony WH-1000XM5 under $280."

A deal is posted with the title: "Sony WH1000XM5 Wireless Headphones $249.99 at Best Buy."

A traditional keyword alert for "Sony WH-1000XM5" would likely miss this because of the missing hyphen. Even if it caught it, it wouldn't necessarily know that $249.99 is below your $280 limit without manual regex or brittle parsing.

Our system handles it differently:

• Normalization: It identifies the product as a Sony WH1000XM5 and the price as $249.99.
• Intent: It knows you want that specific model and have a $280 cap.
• Vector Matching: It sees that "WH-1000XM5" and "WH1000XM5" are semantically identical.
• Verification: It confirms $249.99 is less than $280.

You get a notification instantly, while others using basic keyword tools are still waiting for a hyphenated version to appear.

The Limits of AI Matching

AI deal matching is an improvement over keywords, but it isn't magic. There are still challenges. Sometimes a deal for a "Laptop Case" might have high similarity to an alert for a "Laptop" because the words are so similar.

This is why we don't rely on AI alone. By layering structured data and deterministic rules over the top, we can catch those false positives. We're constantly tuning the balance between catching every possible deal and keeping your inbox clean.

Our engineering philosophy is that the system should work for you, not the other way around. You should be able to describe what you want in your own words, and we should be the ones figuring out the complex logic required to find it.