How Multimodal AI Can Connect Products, Promotions and Content in Retail

Retailers have spent years building systems to understand what they sell.

Product databases contain SKUs, categories, prices, sizes, colors, descriptions, and availability. Customer data is stored across CRM, loyalty, commerce, and other customer-facing systems. Digital asset platforms store product photography and campaign creative. Merchandising teams work with spreadsheets and presentations. Suppliers send PDFs. Store teams use planograms. Marketing teams create banners, catalogs, and social content.

All of that information is valuable. Yet much of it still lives in separate systems, formats, and contexts, and a surprising amount of retail knowledge is visual.

• A product image can show a design detail that never appears in the product description.
• A promotional banner can contain a price that has changed in the commerce system.
• A planogram can communicate relationships between products that are difficult to capture as text.
• A supplier presentation can contain critical information in charts, tables, or images.

Traditional enterprise search and text-based RAG have made it easier to retrieve information from documents. The next step is making the AI layer capable of understanding the visual and structural context around that text.

Google Cloud’s Gemini Embedding 2 semantically maps text, images, document pages, rendered tables, and charts into a shared semantic space, enabling cross-modal retrieval and more context-aware understanding of enterprise content.

For retailers, this offers far more than merely adding conversational features to product discovery. It can provide an important foundation for a unified operational intelligence layer, one that seamlessly bridges product catalogs, active promotions, brand media, and merchandising insights.

Retail’s Knowledge Problem Isn’t Just About Data

Retailers often talk about having too much data. The deeper problem is that the data doesn’t necessarily understand itself.

Consider a single product launch:

• The product information management system might contain the SKU, product name, category, specifications, and launch date.
• The pricing system contains the current price.
• The marketing team has a campaign deck.
• The creative team has product images.
• The ecommerce team has web banners.
• The supplier has sent a PDF catalog.
• The merchandising team has created a planogram.

Every team has part of the story. An employee looking at just one system gets an incomplete picture.

A traditional search system can retrieve the relevant files. A conventional RAG system can retrieve passages from those files. But when meaning depends on the relationships among an image, a table, a layout, or information spread across several file types, simply converting everything to text can lose important context.

Google Cloud highlights this problem, noting that flattening complex tables or other structured content into text can separate values from the visual relationships that give them meaning. Multimodal embeddings are designed to preserve more of that visual and spatial context.

For retail, that matters because the enterprise knowledge base isn’t a text library but a mix of products, images, campaigns, catalogs, spreadsheets, presentations, layouts, and operational records.

From Product Data to Product Understanding

Retail search has already moved beyond simple keyword matching.

Google Cloud’s retail case with Target, for instance, describes a hybrid approach that combines traditional keyword matching with semantic search powered by vector embeddings to make product discovery more relevant to natural-language queries.

Multimodal search takes that idea further. Imagine a merchandising manager searching:

“Find all products that look like this, are part of the autumn collection, and are currently priced below €100.”

The answer could require understanding a reference image, product attributes, campaign metadata, and current pricing.

Or imagine a customer asking:

“Show me something similar to this jacket, but in a darker color.”

That request isn’t purely textual. The customer’s intent may depend on visual similarity as much as product metadata.

Multimodal search architectures combine product data with product images, using multimodal embeddings alongside search or vector search to support text-to-image and image-to-text retrieval. That opens the door to a broader idea. Retailers can start treating visual assets as part of the enterprise data model rather than as files that sit beside it.

The Operational Opportunity is Bigger Than Shopping

Much of the discussion around multimodal AI in retail naturally focuses on the customer, and that makes sense. Today’s retail strategy is heavily focused on agentic commerce, with AI agents helping shoppers discover products, receive recommendations, and navigate the purchasing journey. But there’s another opportunity that could be just as important.

What happens when the retailer’s employees can query the same multimodal knowledge layer?

• A merchandising manager doesn’t necessarily need another chatbot. They need to know whether the information they’re working with is accurate, current, and consistent.
• A marketing manager needs to know whether a campaign asset still matches the current offer.
• A category manager wants to find similar products across thousands of images.
• A supplier manager needs to compare a new supplier catalog against the retailer’s existing product records.
• A store operations team may want to find the right planogram for a specific store format.

Far from being basic consumer-facing issues, these represent core operational challenges where multimodal retrieval offers significant enterprise value.

1. Detect Discrepancies Between Product Images and Product Data

A retailer might update a product’s attributes in its product information system but leave an older image in the digital asset library. Perhaps the packaging or the color has changed, or a new product version has been introduced while older creative is still being reused.

A text-based system can compare product descriptions. A multimodal retrieval and AI system can examine the visual assets associated with the product and compare them with the corresponding product records. That creates the possibility of automated content-quality checks such as:

Product data says: Black running shoe

Campaign image shows: White running shoe

Or:

Product record: New packaging

Published catalog: Previous packaging

The value isn’t limited to catching mistakes. Over time, retailers could build automated content governance workflows that identify visual and textual inconsistencies before they reach customers.

2. Find Obsolete Promotional Material

Retailers produce enormous volumes of promotional content.

Creating marketing campaigns often results in a vast number of assets across channels such as email, e-commerce, apps, social platforms, physical store displays, and partner networks. However, these materials can rapidly become obsolete when promotional terms shift, prices are adjusted, campaigns conclude, or products are retired.

A multimodal AI agent could help retailers identify assets that reference outdated products, prices, promotional messages, or visual branding. As reflected in Google Cloud’s cross-document reconciliation pattern, multimodal models can evaluate information across diverse formats to detect outdated or inconsistent information.

For retailers, that could mean moving from manual content audits toward continuous content intelligence.

3. Check Whether Campaign Creative Matches Current Pricing

Imagine a retailer has 50,000 campaign assets spread across several repositories.

• The current pricing database says Product A costs €79.
• A banner published three months ago shows €99.
• A PDF catalog still lists €89.
• A product image includes an old promotional sticker.

While manual reviews can uncover these inconsistencies, they require significant time and effort. Conversely, a multimodal retrieval and agentic workflow can connect visual materials with underlying pricing and product records, automatically highlighting discrepancies. This highlights the critical shift from relying on a basic AI assistant to implementing an AI-driven operational workflow.

The system isn’t simply answering: “What is the current price?” It identifies multiple representations of the same product, compares them against an authoritative source, and surfaces an inconsistency. That is a much more operational use of AI.

4. Search Planograms Using Natural Language

Planograms contain highly structured spatial information. The position of products matters. Relationships between categories matter. Shelf placement matters.

A conventional text index can store the names of products in a planogram, but it can lose the spatial relationships that indicate where products appear on a shelf and how they are arranged relative to one another.

Multimodal retrieval creates another possibility. A store operations manager could ask: “Show me the summer beverage planograms for stores with this layout,” or “Find planograms where premium products are placed at eye level.”

The AI layer could retrieve visual documents based on their content and context, rather than requiring employees to know the file name, folder structure, or exact terminology used when the planogram was created. That can turn a large archive of visual merchandising assets into a searchable operational knowledge base.

5. Connect Supplier PDFs with SKU Data

Supplier information is another major source of fragmentation. A supplier might send a PDF catalog containing product images, tables, specifications, and promotional information. The retailer’s internal systems might contain SKUs, pricing, inventory, and category information.

The two sources rarely arrive in the same structure. A multimodal AI system could help bridge that gap.

Instead of treating the supplier PDF as a static document, an AI workflow could extract and connect relevant information with internal product records, helping teams identify:

• New products that aren’t yet in the catalog
• Differences in product specifications
• Missing SKUs
• Inconsistent product descriptions
• Pricing differences
• Duplicate products
• Visual differences between product versions

This is an example of the broader cross-file synthesis capability that multimodal embeddings can support, enabling retrieval of information across PDFs, spreadsheets, images, and other formats.

The embeddings provide the semantic retrieval layer, while document extraction, entity matching, and business-system integrations would be needed to turn those matches into reliable reconciliation workflows.

6. Give Merchandising Teams an AI Research Assistant

Merchandising decisions often require information from many places.

A category manager might ask: “Which products in this category have visual characteristics similar to our best-selling products, are currently in stock, and haven’t appeared in a campaign this quarter?”

Answering that manually means searching several systems. An AI research assistant could potentially combine product attributes, visual similarity, campaign history, and inventory information to assemble the answer. That doesn’t mean the AI should make the merchandising decision. It means the AI can compress the research process, bring together the relevant evidence, and give the human decision-maker a much better starting point.

This distinction becomes increasingly important as retailers move from basic AI assistants toward agentic systems that can operate across business workflows.

A Visual Knowledge Layer for Retail

The bigger architectural opportunity is to stop thinking about product images, documents, and business data as separate categories. A retailer’s product knowledge might look something like this:

Product

→ SKU

→ Category

→ Price

→ Inventory

→ Product description

→ Product images

→ Supplier documents

→ Campaign assets

→ Planograms

→ Customer interactions

Rather than consolidating these assets into a unified repository, the primary objective is to establish a semantic framework that models their interrelationships. This approach enables direct connections between a product image and its corresponding SKU, current pricing, active promotional campaigns, associated planograms, and supplier documentation.

The result begins to resemble a visual retail knowledge graph, even when the underlying architecture combines multiple technologies rather than relying on a traditional graph database.

Multimodal embeddings can provide one important foundation for this by allowing different modalities to be represented in a common semantic space. Gemini Embedding 2 supports text, images, document pages, rendered spreadsheet tables, and charts in a unified representation, with cross-modal retrieval across formats. The same architecture can also extend beyond static retail assets to video, audio, and other multimodal content. From there, retailers can build search, recommendation, analytics, and agentic workflows on top of it.

From Search to Action

Here lies a far more compelling opportunity: moving beyond simple information retrieval, a multimodal knowledge layer can serve as one foundation for retail agents.

• An agent could detect a stale promotional asset, identify the associated SKU, check the latest pricing, notify the responsible team, and create a review task.
• Another could compare supplier content with the retailer’s product catalog and prepare a list of discrepancies.
• Another could help a merchandiser research a category by combining visual product similarity with sales and inventory information.

The human remains accountable for decisions that require judgment or business approval. The agent handles the repetitive work of finding, comparing, reconciling, and preparing information. That distinction is important as retailers bring AI deeper into operational workflows.

The Architecture Behind Multimodal Retail AI

Getting this right requires more than choosing a multimodal model. The underlying data architecture matters.

Retailers need a way to connect structured business data with unstructured and multimodal content. They also need reliable metadata, identity and access controls, data freshness, observability and clear ownership of authoritative sources.

A practical architecture could bring together:

• Data sources: Product information, pricing, inventory, campaign assets, product images, supplier documents, planograms and merchandising data.
• Data and storage layer: Cloud storage and analytical platforms such as BigQuery, with appropriate metadata and governance.
• Semantic layer: Multimodal embeddings and vector retrieval to connect products, images, documents, tables and other content based on meaning.
• AI layer: Gemini-powered applications and agents for search, research, reconciliation, content analysis and workflow assistance.
• Business systems: Commerce, PIM, ERP, CRM, loyalty and campaign platforms.
• Governance and security: Identity-aware access, permissions, auditability, source attribution and human approval for high-impact actions.

This last layer is critical:

• An AI agent that can identify a stale banner doesn’t need permission to change the price.
• An agent recommending a product doesn’t necessarily need access to customers’ financial information.
• An agent researching supplier documents shouldn’t automatically gain access to every internal commercial file.

The architecture needs to carry the retailer’s existing governance model into the agentic world.

The Next Retail Data Advantage May Be Visual

Retailers have invested heavily in first-party data, product data, and digital commerce infrastructure. The next opportunity is to make their visual assets equally useful.

Product photography, promotional creative, catalogs, planograms, presentations, and scanned supplier documents contain business knowledge. For years, much of that knowledge has remained difficult for enterprise systems to search and reason across.

Multimodal AI is changing how this information can be indexed, retrieved, and connected. As retail strategy increasingly shifts toward intelligent agents that can assist both customers and internal teams across the entire value chain, multimodal embeddings provide a structural layer. This enables AI architectures to bridge and synthesize information across diverse data formats and modalities.

The real opportunity for retailers is to build a retail knowledge layer that can see the whole picture.

When a product image, promotion, price, supplier document, and merchandising plan can be understood as connected pieces of the same business context, AI agents can move from answering questions to helping retailers keep that context accurate, current, and useful. Long before a shopper queries a digital assistant, multimodal AI can create new operational advantages across the retail value chain.

How Kartaca Can Help

For retail businesses, implementing a new AI model is only part of the hurdle; the real key lies in seamlessly integrating multimodal content into existing operational workflows, core applications, and enterprise data systems.

Kartaca helps retail organizations build cloud, data, AI, and analytics architectures that turn fragmented enterprise data into a foundation for intelligent applications and agentic workflows. From data platform modernization and BigQuery architectures to AI assistants, analytics, and customer-centric solutions, we help retailers connect their data foundation with the next generation of AI.

The next retail experience may be agentic. The data foundation behind it needs to be ready first.

Contact us today to start building your retail data and AI foundation.