DoorDash's 90.7% Query Accuracy: A 2026 Business Case for AI Adoption in Marketplaces

A useful new AI business case surfaced on June 4, 2026, when DoorDash researchers published a paper describing how the company uses large language models to improve recommendations in newer grocery and retail categories. By itself, that would be interesting but incomplete. What makes the case stronger is that it sits on top of an earlier DoorDash production paper from March 2, 2026 showing how the company already deployed an agentic search-intent system across more than 95% of daily search impressions. In that system, long-tail query accuracy reached 90.7%, or 13 percentage points better than baseline.

Taken together, the two papers describe something many companies still have not achieved: AI that is not floating on the edge of the business as a chatbot, but wired directly into the marketplace engine. DoorDash is using LLMs to understand ambiguous customer intent, map behavior across verticals, and improve how products are ranked for people who may never have ordered groceries or retail items before. That is a commercially meaningful problem because the company's growth increasingly depends on getting users to expand beyond restaurants.

The latest paper adds another important detail: cost discipline. DoorDash says it found GPT-4o-mini delivered similar output quality for its recommendation feature-generation task at lower cost, and that prompt caching plus just-in-time updates cut overall compute costs by roughly 80%. That matters because it turns the case from a pure quality story into an operating model story. Better relevance is useful. Better relevance with explicit cost controls is a business case.

What DoorDash Actually Built

The March paper focused on a marketplace search problem. Some DoorDash queries are ambiguous. A term like "Wildflower" might mean a restaurant, a retail brand, or a floral item. Traditional classifiers can force one answer too early, while generic LLMs can hallucinate things the marketplace does not actually carry. DoorDash's answer was to ground the model in two sources: its own staged catalog-entity retrieval pipeline and an agentic web-search layer for cold-start queries. Instead of forcing one label immediately, the system emits an ordered multi-intent set and lets business rules resolve the ambiguity.

The June paper tackled a different but related problem: cross-vertical personalization. DoorDash already has rich restaurant-order data, but newer verticals like grocery and retail have a cold-start problem. Many users simply have not generated enough non-restaurant history for the ranker to know what to show them. DoorDash used a hierarchical retrieval-augmented generation pipeline to infer product-category affinities from restaurant orders and search behavior, then injected those features into its production multi-task ranking model.

In practical terms, the company did not replace its marketplace stack with an LLM. It kept the production ranking architecture and used LLMs to manufacture better inputs for it. That is one reason the case is more credible than most enterprise AI theater. The model is there to improve the decision system, not to impersonate the decision system.

The strongest marketplace AI wins usually come from grounding the model inside the catalog and ranking loop, not from asking a generic model to freestyle the answer.

Why This Looks Like a Real Business Case

There are four reasons this case deserves attention.

First, the metrics are attached to live production surfaces. Serving more than 95% of daily search impressions is not a pilot. Neither is a 90.7% long-tail query accuracy result against ambiguous marketplace intent. This is high-volume workflow infrastructure, not a showcase demo.

Second, the company is solving a growth problem that matters commercially. Moving a restaurant user into grocery, convenience, or retail ordering changes order frequency, basket composition, and category reach. Better search resolution and better cold-start recommendations are exactly the kind of invisible improvements that compound into a stronger marketplace over time.

Third, DoorDash is showing some economic discipline around the AI layer. The latest paper did not just say the model improved personalization. It also described choosing a smaller model for comparable quality and cutting compute costs by around 80% with prompt caching and update discipline. That is what a serious AI operating model looks like: quality gains that are constrained by cost logic.

Fourth, the architecture reflects a mature view of where LLMs help. DoorDash uses them to infer affinities and resolve ambiguity, but it still relies on deterministic business rules, feature stores, and production ranking systems to do the final operational work. That hybrid model is far more transferable than the common idea that an LLM should simply sit in front of the customer and improvise.

What Other Companies Should Copy

Most businesses are not multi-vertical marketplaces, but the design lessons transfer well.

• Ground the model in proprietary context. Catalog structure, taxonomy, and policy rules are often worth more than a bigger model.
• Use AI to improve the ranking system, not replace it. The highest leverage often comes from better features and better retrieval, not from throwing away the production stack.
• Separate inference from resolution. Let the model propose likely intents, but keep deterministic business logic for final decisions.
• Measure live workflow quality. Search accuracy, ranking lift, conversion quality, and impression coverage beat generic productivity claims.
• Treat cost as part of the product. Smaller models, caching, and just-in-time updates can matter as much as raw output quality.

This is especially relevant for retailers, travel marketplaces, financial platforms, logistics networks, and B2B software companies with large catalogs or complex discovery journeys. In those environments, AI often creates the most value when it makes the matching system smarter, not when it adds another conversational layer on top.

The Caveats

The case is still incomplete in one important way: DoorDash has not published a clean revenue or margin bridge tied specifically to these systems. The evidence comes from technical papers, not audited financial disclosures. That means the business value has to be inferred from better discovery quality, better cold-start personalization, broader category expansion, and lower model cost. Those are reasonable signals, but they are still indirect.

There is also a transferability limit. DoorDash has a dense catalog, strong data infrastructure, and enough traffic for small improvements in relevance to matter quickly. A smaller company with weak taxonomy, sparse behavioral data, or poor product instrumentation will not reproduce these results by adding an LLM alone. The surrounding retrieval, ranking, and measurement systems still matter.

Even so, this is a better AI adoption story than most because it clears a practical bar. The company used AI on a real marketplace problem, tied it to production systems, measured quality, controlled cost, and aimed the work at category expansion rather than novelty. That is much closer to business reality than a generic promise that AI makes teams "more productive."

The Business Takeaway

DoorDash's latest case suggests that successful AI adoption in marketplaces does not start with a flashy assistant. It starts deeper in the decision stack: understanding ambiguous intent, translating sparse behavior into usable signals, and feeding those signals into systems that already decide what customers see and buy.

If you are building an AI business case inside your own company, start there. Find the revenue-critical workflow where discovery is weak, personalization is thin, or the system lacks enough context to make a confident decision. Then ask whether AI can improve the quality of the inputs while staying grounded in your data, your rules, and your cost envelope. That is when AI stops being a side experiment and starts becoming operating leverage.

← Back to all articles