Skip to content

SYSTEM Cited by 2 sources

Instacart Carrot Ads

Definition

Carrot Ads is Instacart's omnichannel retail-media platform: a system that lets retailer partners build and operate their own advertising businesses on either their owned-and-operated (O&O) e-commerce sites and apps or on Instacart-hosted whitelabel Storefront properties.

Quote (Source: sources/2026-05-04-instacart-empowering-carrot-ads-with-domain-adaptive-learning):

"Carrot Ads is Instacart's omnichannel retail media solution that allows retailer partners to build and scale their own advertising businesses on either their owned-and-operated (O&O) websites and apps or their whitelabel Storefront hosted by Instacart. Carrot Ads empowers retailers and CPG brands to accelerate revenue, while improving the customer experience, engagement and Ads return on investment."

Carrot Ads draws demand from two pools: retailer-sourced advertisers (the partner's existing CPG / agency relationships) and Instacart-sourced demand from 7,500+ CPG brands. Partners can choose either pool or both.

What it does (system role)

  • Auction layer: real-time auctions over partner ad inventory, ranked by bid × predicted_CTR (the standard ads-ranking formula).
  • CTR prediction: a partner-specific pCTR model scores each candidate ad for each request, under real-time auction latency budgets.
  • Demand brokering: composes retailer-sourced and Instacart-sourced demand into the same auction.
  • Onboarding pipeline: brings each new partner online with a performant pCTR model from day one — the Domain Adaptive Learning stack on top of transfer learning is the load-bearing mechanism that closes the new-partner cold-start gap.

Domain Adaptive Learning as the cold-start solver

The structural problem Carrot Ads solves: "onboarding a new partner onto Carrot Ads introduces a key challenge: the 'cold start' problem, where limited historical interactions make it difficult to predict user behavior accurately." The two naive options both fail:

  • Train a partner-specific model from scratch — too data-hungry for a partner that has just launched.
  • Deploy Instacart's existing Marketplace model directly — "often fails to capture the nuances of the partner's specific inventory and user base."

Carrot Ads instead applies Domain Adaptive Learning (a subset of transfer learning) — Instacart Marketplace = source domain, partner site = target domain — and reuses representations and relationship signals from the source to warm-start the target model. Two adaptation layers are simultaneously active:

  1. Neural-network level ( patterns/cross-domain-warm-start-via-shared-embeddings): shared shopping-context-pre-trained embedding layers, feature transfer (wide ⇄ explicit; deep ⇄ pre-trained dense), selective fine-tuning of subsequent layers on partner data, reuse of shared representations for generalization.
  2. Training-data level: feature/taxonomy alignment across domains, source-domain selection (Marketplace), and per-partner feature trimming to honor real-time auction latency.

Reported outcomes

"By leveraging the 'source' knowledge of the Instacart Marketplace, we achieved higher CTR, total clicks per user and ads revenue across search ads and product category ads. This approach enables us to launch high-performing ad networks for partners immediately, eliminate the traditional data ramp-up period and converge to a better stable state."

Counter-intuitive property: Domain-adaptive training outperforms direct partner-data training even when the target partner has enough data to train a model on its own — because the Instacart first-party signal contributes information the partner site never sees on its own.

First-party data as the structural moat

Retail-media platforms (Instacart, Amazon Ads, Walmart Connect) have a structural advantage over standalone ad networks: years of proprietary purchase + behavior data. Carrot Ads makes that asymmetry concrete by reusing Instacart Marketplace data as source-domain signal across every partner deployment. The same embedding pre-training amortizes across all partners; each new partner gets to stand on top of it.

Operating constraints

  • Real-time auction latency — partner pCTR scoring is on the hot path of an ads auction; the model must produce scores within latency budgets the post does not numerically disclose.
  • Per-partner feature availability — different partners surface different features at request time. Carrot Ads applies per-partner feature trimming to "be flexible to various feature availability for the partners" and to keep the model lightweight per partner.
  • Negative-transfer risk — schema-mapping mistakes or distribution mismatches between source and target can make the transferred knowledge hurt target performance. "This process is not yet fully autonomous. The complexity of mapping data schemas and verifying model alignment currently requires human-in-the-loop verification to prevent negative transfer."

Roadmap: Domain Adaptation Platform

Instacart is "building an automated Domain Adaptation Platform that can detect domain shifts and fundamentally streamline the workflow." Goal: turn the per-partner DAL recipe into a self-service onboarding pipeline that detects schema drift / distribution shift and brings a new partner up with minimal human intervention. As of 2026-05-04 this is forward-looking work, not shipped.

This continues Instacart's platform-thinking arc visible in PIXEL (image generation), PARSE (attribute extraction), and Maple (batch LLM inference) — turn a specific ML capability into a self-service team-wide platform.

Composition with other Instacart systems

Capability System / pattern Role in Carrot Ads
pCTR model architecture systems/instacart-carrot-ads-pctr-model / concepts/wide-and-deep-architecture Wide-and-deep backbone with domain-adaptive training
Cross-domain warm start patterns/cross-domain-warm-start-via-shared-embeddings Reused shopping-context embeddings + fine-tuned partner layers
Auction-latency feature trim patterns/per-partner-feature-trimming-for-auction-latency Feature-importance-driven pruning per partner
First-party signal reuse Instacart Marketplace data Source-domain corpus across all partners

Caveats

  • No public partner roster, no partner count, no onboarding-time numbers. The post does not disclose how many partners run on Carrot Ads, who they are, or how long onboarding takes today versus the platform target.
  • No latency budgets disclosed. Real-time auction latency is named as a constraint without a numeric target.
  • No A/B-tested lift numbers. "Higher CTR, total clicks per user and ads revenue" is the claim; specifics are absent.
  • Auction stack itself is undocumented in this source. The post focuses on the pCTR model side of Carrot Ads; the bidding, budget pacing, fraud prevention, and brand-safety layers are out of scope.

Seen in

  • sources/2026-05-04-instacart-empowering-carrot-ads-with-domain-adaptive-learning — first wiki canonicalisation. Carrot Ads framed as the omnichannel retail-media platform; DAL as the cold-start solver for partner onboarding; Wide-and-Deep pCTR as the model backbone; HITL verification as the gating activity; Domain Adaptation Platform as forward roadmap.

  • sources/2026-06-02-instacart-from-scoring-to-spelling-rebuilding-ads-retrieval-at-instacart — canonicalises Carrot Ads' candidate-generation stage on browse surfaces (retailer home page + pre-checkout) as systems/instacart-generative-ads-retrieval, replacing the prior BERT-based CR scoring CG. The pCTR ranker disclosed in 2026-05-04 (systems/instacart-carrot-ads-pctr-model) runs unchanged over the new generative CG's candidate set. The two posts thus cover complementary axes of the Carrot Ads stack:

  • 2026-05-04 = ranking axis (Domain Adaptive Learning for pCTR cold-start across partners).
  • 2026-06-02 = retrieval axis (generative retrieval over Semantic IDs for catalog-coverage / brand-diversity / cold-start at the candidate-generation stage). Operational outcomes for the retrieval-axis change: ~2× more candidates per request at −10–17% mean latency; +5% CTR; +34% ATC; 2.7× more brands / 1.8× more sub-categories in retrieved candidates; +421% / +396% / +229% diversity in Alcohol / Beverages / Healthcare. Forward composition risk: the post explicitly acknowledges "if the subsequent ranking model was miscalibrated on these outlier products, these incoherent recommendations from the candidate set would eventually get bubbled up to the user" — which would have been a structural failure mode of the prior CR architecture; the generative retriever's autoregressive prefix conditioning is asserted to prevent the outlier leakage but pCTR-side calibration with the new candidate distribution is not separately addressed.
Last updated · 542 distilled / 1,571 read