# Product Knowledge | Product Facts & Enrichment

> Connect every source, translate every asset, and build consistent, traceable product knowledge for AI discovery — with source evidence and merchant review.

*AI-readable version of [Product Knowledge | Product Facts & Enrichment](https://www.deeplumen.com/product-facts/) · generated by DeepLumen Agentic Page*

PRODUCT FACTS & ENRICHMENT

Connect every source. Translate every asset. Build product knowledge that is consistent, traceable and ready for AI.

“Fits every mattress. One piece in every pack.”

Mattress depth: 20–30 cm. Match the guide to the selected variant.

### Cotton Fitted Sheet

Bedding · C1

Mattress size

Packs · by SKU **1 or 2 sheets**

C1 records and source excerpts are illustrative. Photo: Secret Linen Store; it does not verify the example specifications.

## One product. A complete knowledge system.

The four-stage architecture connects source material, product facts, information formats and AI entry points. Evidence and outcomes return to the next version.

## How a product becomes understandable and recommendable to AI

Combine shopper questions, product data and commercial constraints into verifiable product facts, then compile them for different AI entry points.

- Mattress fit Which cotton fitted sheet fits a 180 × 200 cm mattress that is 25 cm deep?
- Pack contents Does the two-pack include pillowcases? How many sheets are in each SKU?
- Fit boundaries My mattress is 35 cm deep. Is this fitted sheet confirmed to fit?
- Care & claims Can I machine wash and tumble dry it? Are hypoallergenic and antibacterial claims supported?
- Material details What is the sheet made from, and is the fabric composition verified?
- Variant options Which sizes and colours are available, and how do I choose the right SKU?

### Product data sources

Shopify · WooCommerce · Product pages

Reviews · Q&A · Reddit · Support · Social

PIM · ERP · OMS · PLM · Brand policies

### Unified product facts

### One product, a set of verifiable facts

Interpret images, video, audio and documents alongside source product facts to create reviewable information. Prices, inventory and policies remain hard constraints.

### Information formats

Product facts · Contextual Q&A · Illustrated explanations · Video summaries · Evidence

Attributes · Features · FAQs · JSON-LD · metafield

Feed distribution · MCP tools · Skill packages

### AI entry points

THE DIFFERENCE, MADE VISIBLE

## Better context. Without changing the facts.

One fitted sheet. Three SKUs. See how fit, packs, care and unsupported claims become clearer, then see what AI enrichment adds.

### Match each fact to its SKU, source and version.

Fix missing compatibility fields and pack scope. Escalate conflicting care instructions, and keep unsupported claims out of approved copy.

#### Shared product page · S1

“Fits every standard mattress. One sheet per pack. Machine wash and tumble dry. Hypoallergenic and antibacterial.”

#### SKU records · S2 / S5

C1-150-1: 150 × 200 cm / 1 sheet. C1-180-1: 180 × 200 cm / 1 sheet. C1-180-2: 180 × 200 cm / 2 sheets.

#### Care guide · S4 · 2026 edition

Hand wash at ≤30°C. Do not tumble dry. Confirm which current batches and SKUs this version covers.

Fit depth All three SKUs: 20–30 cm. These are compatible mattress measurements, not finished-sheet dimensions.
Pack scope The selected SKU determines quantity. Pillowcases and props are excluded in packing list S5.
Care version After the merchant confirms S4 applies: hand wash at ≤30°C; do not tumble dry. Keep the conflict open until then.
Claims Hypoallergenic and antibacterial: evidence missing. Cotton composition does not establish either claim.

Confirm both the care version and its scope. The model cannot approve either for the merchant.

Care instructions conflict. Confirm the applicable version and SKU scope before publishing.

### Turn specifications into a useful buying decision.

This example uses reviewed fit and packing records S2/S5. Switching SKU changes its dimensions and pack count while retaining shared constraints.

See which facts change, and which constraints stay.

C1 cotton fitted sheet. Ivory. Sizes: 150 × 200 cm and 180 × 200 cm. Mattress depth: 20–30 cm. One- or two-sheet packs.

#### The shopper still needs answers.

- Which exact SKU matches my mattress and pack choice?
- Does the two-pack include pillowcases?
- What if my mattress is 35 cm deep?

#### Choose the fit. Know what is in the pack.

Match mattress width, length and depth to the selected SKU. Choose one or two fitted sheets according to the packing list; pillowcases are not included.

Selected SKU C1-180-2
Mattress fit 180 × 200 cm
Mattress depth 20–30 cm
Pack contents 2 sheets
Material 100% cotton
Not included Pillowcases not included

A spare-sheet rotation for a mattress inside the documented fit range.

The documented depth range is 20–30 cm. A 35 cm fit is not confirmed; ask for SKU-specific evidence instead of promising compatibility.

### Translate the asset. Keep what each source can actually prove.

Image labels, speech and document rows become different fields. Their source regions, timestamps and review states stay attached.

“Measure mattress width, length and depth before choosing a size.”

Composition: 100% cotton. This comes from the record, not the image appearance.

C1-180-2 · 180 × 200 cm · depth 20–30 cm

Measure the mattress before selecting the SKU.

material: cotton · composition_percent: 100

THE TRUST CONTRACT

## More context. Same facts.

Five rules govern the system, from the first source to the final expression.

### Keep the original.

Retain source values, product and variant relationships, dates and versions. Normalization never erases the input.

### Extraction is not verification.

A packaging claim or review remains a source statement until supported. Extraction confidence and truth are different questions.

### Enrich meaning, not missing specifications.

Use cases, intent, FAQs and selling points are reviewable candidates. Materials, dimensions, certifications and compatibility still need evidence.

### Read buying conditions from the source.

Price, inventory, delivery, returns and eligibility remain market-specific, time-sensitive facts from authoritative systems.

### Resolve conflicts before publishing.

Unknown and conflicting values keep their status. Merchant approval, visibility rules and channel versions control publication.

THE FIELD-LEVEL MODEL

## The knowledge behind each product.

Six groups cover what it is, how it is used and how it can be bought. Source facts, AI translations and enrichment remain clearly identified.

### Product identity & specifications

Taken from the store and authorized records. Missing and conflicting values keep their status.

- Product & variant IDs, SKU, brand
- Category, attributes, technical specifications
- Materials, dimensions, weight, care
- Evidence-backed compatible accessories

### Multimodal details

Extract what the asset contains. Extraction does not verify the claim.

- Image descriptions and text extraction
- Transcripts, key frames, instructions
- Document text, specifications, size tables
- Image regions, time ranges, document pages

### Shopping context

Evidence-based semantic suggestions, ready for merchant review.

- Use cases, occasions, audiences
- Buying intent and shopper questions
- Top features and reasons to choose
- Pros, cons and usage limits

### Answers & brand expression

Follow approved brand guidance. Store generated copy separately from facts.

- FAQs, comparisons, answer hints
- Brand copy under approved guidance
- Image alt text and visual explanations
- Video summaries and agent prompts

### Price & buying conditions

Refresh from authoritative systems under agreed rules. AI does not guess these values.

- Price, currency, market
- Inventory and purchasability
- Delivery, returns, warranty
- Eligibility, validity, last update

### Sources & control

Trace information to its source and control what can be published.

- Source, original value, version
- Extraction confidence, verification status
- Missing values, conflicts, review decisions
- Audience, permissions, publication channel

THE ARCHITECTURE IN DETAIL

## Four layers. Clear responsibilities.

The overview is the map. These 17 modules define where data enters, how knowledge is built, what gets published and who can use it.

### Collect & connect

Connect platform facts, enterprise data and multimodal assets—images, video, audio and documents—through one traceable pipeline.

#### Commerce platforms & storefronts

Read products, variants, prices, inventory, images and video through official Shopify, WooCommerce and SHOPLINE APIs and webhooks. Supplement them with product-page content and attachments.

Sources API · Webhook · Product page crawling
Input Platform authorization, catalogs, variants, inventory, prices, vendor fields, product pages, and their image, video and attachment links.
Processing Read structured fields and media references with platform connectors. Collect descriptions, FAQs, content blocks and attachments while preserving product and variant associations.
Output Product IDs, variant relationships, titles, descriptions, source images and video, prices, inventory, native metafields, attachment links and page snapshots.
Evidence boundary Preserve the platform, API version, fetch time, page URL and raw response. Read access does not grant permission to write back.

#### Enterprise & supply chain systems

Connect PIM, ERP, OMS, PLM and supplier portals, plus CSV, XML, PDF, SFTP and APIs, to fill gaps in specifications, fulfillment and manufacturer facts.

Sources PIM · ERP · OMS · PLM · Supplier systems
Input Product master data, bills of materials, specification and sizing tables, packaging images, scans, installation manuals, supplier catalogs and fulfillment rules.
Processing Ingest data through APIs, SFTP or file imports; associate it with SKUs, suppliers and markets. Preserve original documents and versions, and pass scans, charts and manuals to the translation engine.
Output Source documents and their SKU index, plus sourced specifications, dimensions, weights, materials, packaging, manufacturers, supply relationships and lead times.
Evidence boundary Link each field to its source file or API, version and update time. Queue records that cannot be matched to a SKU for resolution.

#### Customer feedback & questions

Reviews, Q&A, support tickets, return reasons and presales questions reveal how customers describe experiences, concerns and suitable use cases.

Sources Reviews · Q&A · Support · Returns · Reddit
Input Text and image reviews from Shopify, Google, Amazon and Okendo; on-site Q&A, support records, authorized recordings and return reasons; and relevant Reddit, forum, YouTube, TikTok and Instagram content.
Processing Associate content with products and variants and preserve original text, image, video and audio references. Translate media, then group it by topic and sentiment while separating first-hand experiences, hearsay and marketing claims.
Output Shopper questions, image/video/audio experience evidence, common concerns, and candidates for pros and cons, use cases and comparison dimensions.
Evidence boundary Preserve original URLs, author or source type, publication date and cited excerpts. Social discussion is candidate evidence and must not automatically become a verified fact.

#### Brand assets & policies

Ingest product and packaging images, demo videos, audio and manuals from DAM systems, brand asset libraries, suppliers and policy documents, alongside brand communication rules.

Sources DAM · Brand assets · Supplier materials · Policies
Input Product and detail images, packaging images, demo videos, audio, manuals, brand guidelines, and shipping, returns, warranty and compliance documents.
Processing Register asset IDs, sources, usage scopes, languages and versions; link them to products and variants. Build reusable brand dictionaries and policy rules for translation, enrichment and publishing.
Output Traceable media and document inventories, brand rules, policy facts, permitted promises, prohibited claims and market conditions.
Evidence boundary Version policies by region, channel and effective date. Expired policies must not support public content.

### Understand & enrich

Build a unified product fact system with separate source facts, AI multimodal translations and semantic enrichment candidates, each linked to verifiable source evidence.

#### Unified product facts

Unify products, variants, categories, suppliers, markets and assets in one fact foundation. Associate each image, video segment and document with a specific product.

Sources Platform source facts + External records
Input Platform products and variants, PIM/ERP specifications, policies, review evidence, and source metadata for images, video, audio and documents.
Processing Normalize IDs, SKUs, GTINs, categories, units, markets and languages. Model product–variant–asset relationships and retain sources, versions, timestamps and conflicts.
Output A unified product fact foundation and asset index that distinguish source facts, AI translations and AI enrichment. Authoritative systems continue to govern prices and inventory.
Evidence boundary Trace every field to a source and version. Conflicts require review and must never be silently merged.

#### Multimodal understanding & translation

Convert product information in images, video, audio and documents into text, structured fields and locatable evidence for the fact enrichment engine.

Sources Images · Video · Audio · Documents → Product facts
Input Product, detail and packaging images; demo and user videos; support audio; manuals, scans and specification tables; and their product and variant associations.
Processing Recognize image content and text, interpret key video frames, transcribe speech, and parse document sections and tables. Align extracted results with SKUs, fields, image regions, audio/video time ranges and document pages.
Output Image descriptions, extracted text, packaging statements, speech transcripts, video summaries and chapters, demo steps, specification fields and citable evidence excerpts.
Evidence boundary Preserve source assets, evidence locations, extraction methods, confidence and review status. Assess extraction accuracy separately from claim validity; packaging statements and user claims are not automatically verified benefits.

#### Fact enrichment & semantic mapping

Combine source facts, multimodal translations and feedback evidence to propose reviewable attributes, use cases, intent, Q&A and selling points.

Sources Fact foundation + Multimodal translation + Category templates
Input Source product facts; translated images and video; extracted audio and document information; reviews and shopper questions; brand rules, category templates and merchant preferences.
Processing Align entities and meaning across assets, then use rules and models to propose use cases, Q&A, selling points and comparisons. Link evidence and assess confidence without automatically overwriting source values.
Output Attributes, FAQs, use cases, occasions, user intent, pros and cons, key features, compatible accessories, brand guidance, image alt text and agent prompts.
Evidence boundary Attach the generation method, evidence, confidence, review status and version to each candidate. Mark unsupported candidates as inferred or awaiting confirmation.

#### Live commercial constraints & evidence

Model prices, inventory, shipping, returns, eligibility and variant compatibility as constraints that can expire, be verified and trigger fallback behavior.

Sources Live platform APIs · Policies · Fulfillment systems
Input Prices, inventory, shipping, returns, warranty, purchase eligibility, markets, locales, currencies and variant relationships.
Processing Evaluate sellability by market, currency, time and channel. Detect stale, conflicting or missing data and decide whether publication is allowed.
Output Constraint objects for pages, feeds, MCP and agent queries, with update times, expiry and fallback policies.
Evidence boundary Prioritize live source checks. Models must never guess prices, inventory or delivery promises.

#### Provenance, versions & quality status

Trace each field to an API, image region, audio/video excerpt or document page, with its update time, reviewer and intended audience.

Sources All ingestion & enrichment nodes
Input Source values, AI translation and enrichment candidates, asset IDs, source links, evidence locations, confidence, conflicts, review comments and channel requirements.
Processing Record provenance, asset versions, image regions, audio/video start and end times, document pages and source excerpts. Manage generation types, approval, visibility, freshness and quality status.
Output Field–asset–evidence links and product fact versions that support review, publishing, rollback and withdrawal.
Evidence boundary Preserve original assets and extraction records. Label source statements, AI observations and verified facts separately; high extraction confidence does not establish claim validity.

### Review & express

Compile reviewed product facts into formats that shoppers, crawlers, search engines and agents can consume.

#### Review & publication control

Review AI candidates by field, product, market and channel. Choose whether to show them to shoppers, agents or both.

Sources Review workspace · Version system
Input Source facts, AI translation and enrichment candidates, original assets, locatable evidence, brand rules, quality alerts and channel requirements.
Processing Review field changes against original image regions, audio/video excerpts and document pages. Support bulk actions, drafts, approval, publishing, withdrawal and rollback; choose shopper, agent or shared visibility.
Output Approved product versions with reviewer, timestamp, visibility scope, version number and change history.
Evidence boundary Generated or revised content candidates require approval under review rules before publication. Dynamic fields such as price and stock refresh under authorization and freshness rules.

#### Existing product pages & Agentic Page

Improve product detail content on the existing storefront and provide AI-readable expression through Agentic Page.

Sources Existing merchant site · Shopify themes and app blocks
Input Approved titles, descriptions, attributes, FAQs, use cases, selling points, policies and constraints, plus image descriptions, video summaries, instructions and source evidence.
Processing Compose illustrated explanations, contextual Q&A, comparisons, video summaries and chapters, evidence and actions. Keep content readable by shoppers, parsable by machines and linked to its assets.
Output Optimized product pages, Agentic Pages, image alt text, illustrated content blocks, video explanations, citable summaries and their page versions.
Evidence boundary Inspect page source, structured data, versions and publication receipts. A page visit does not establish an AI citation or an order.

#### Structured data & Shopify writeback

Compile facts into standard structured markup and Shopify-readable fields to reduce parsing effort across entry points.

Sources JSON-LD · Shopify metafields
Input Products, variants, prices, inventory, FAQs, brands and policies, plus approved media links, image alt text, video summaries and evidence references.
Processing Generate applicable Product, FAQ, ImageObject and VideoObject structured markup. Write metafields and product-page content blocks within the platform and theme capabilities.
Output Structured product, Q&A and media data, Shopify metafields, illustrated and video content blocks, and verifiable page source.
Evidence boundary Validate schemas, read platform values back and record page versions. Complete structured data does not guarantee a search rich result.

#### Feeds, MCP, APIs & Skills

Compile separate outputs for AI platforms and agents using each channel's schema, taxonomy, policies and receipts.

Sources ACP · UCP · Feed · MCP · API · Skill
Input Approved product facts, translated content, original media links, evidence excerpts, channel rules, field mappings, merchant authorization and market settings.
Processing Compile product fields, media links, translated summaries and evidence references against the target channel’s supported schema. Generate feeds, expose MCP/API queries or package Skill knowledge, then validate publishing receipts.
Output Channel-adapted feeds, MCP tools, APIs and Skills with supported media references, product facts, summaries, evidence and synchronization receipts.
Evidence boundary Channels accept different image, video, audio and extension fields. Validate format support, receipt, indexing, citation, tool use and orders separately; a live endpoint does not establish distribution.

### Discover & use

AI entry points retrieve and cite product knowledge. Subsequent shopping actions depend on platform permissions and existing commerce interfaces.

#### AI crawlers & search engines

Crawl public pages and structured markup for indexing, summaries and citation candidates. Feed ingestion and tool calls use separate access paths.

Sources AI crawlers · Search retrieval
Input Public PDPs, Agentic Page and product, FAQ and media markup within those pages.
Processing Crawl, parse, index and retrieve information. Platform rules determine whether summaries or citations are displayed.
Output Discovery, indexing, summaries, citations and product candidates.
Evidence boundary Measure crawling, indexing and citations separately. A request is not shopper traffic, and a citation is not an order.

#### LLMs & AI search surfaces

ChatGPT, Claude, Perplexity and Google AI Overview use web pages, search results or their supported integrations in response to shopper questions.

Sources ChatGPT · Claude · Perplexity · Google AI Overview
Input Shopper questions, retrieved results, product facts, policy constraints, citations and tool responses.
Processing Understand intent, filter candidates, compare products, explain recommendations and propose next actions.
Output Product recommendations, comparisons, citations and purchase links; cart or checkout entry points depend on platform support and merchant interfaces.
Evidence boundary Reading, citing and recommending depend on platform retrieval, ranking, region, conversation context and user preferences.

#### Task-based agent

Other task agents may use browsers, feeds, MCP or APIs to search, compare and complete multistep tasks, depending on integration and permissions.

Sources Other task and shopping agents
Input Task objectives, structured product facts, constraints, tool permissions and merchant action APIs.
Processing Plan tasks, call tools, verify prices and inventory, filter products and execute the next action.
Output Completed tasks, product candidates, purchase links, carts or handoff to checkout systems.
Evidence boundary A tool call does not establish user acceptance. Record the agent, task, SKU, tool response and final action.

#### Personal agents & shopping systems

Personal assistants such as Muse and Dots, alongside shopping systems, combine product facts with lasting preferences, lists and execution workflows.

Sources Muse · Dots · Shopping systems
Input Personal preferences, budgets, historical behavior, product facts, inventory and shipping constraints.
Processing Use long-term user context to filter, remind, compare, repurchase and carry out shopping actions.
Output Personalized recommendations, reminders, purchase decisions, orders and aftersales tasks.
Evidence boundary Authorize and record personal context, platform permissions and order status separately. Inferred preferences must not be treated as merchant facts.

EXPRESSION, NOT DUPLICATION

## One knowledge base. Purpose-built expressions.

Approved content is published by version. Price, stock and other changing conditions refresh from authoritative systems under authorized rules.

### Agentic Page

Product facts, scenario answers, image explanations, video summaries and traceable sources.

### Product detail pages

Approved titles, attributes, FAQs, selling points and media copy, adapted to the storefront.

Check the published page as well as the metafield. A theme block may still display old hard-coded content.

### Feeds, MCP & Skills

Package facts, assets and evidence into feeds, tools and reusable agent knowledge.

Product feeds, ACP/UCP adaptations, MCP tools and Skills each follow their own specifications and agreed scope. Checkout remains with the connected commerce platform.

Current starting point: Agentic Page for Shopify. Other connectors, storefront changes and machine integrations are confirmed separately. AI-platform access, indexing and adoption follow each platform’s rules.

A LOOP THAT LEARNS

## Publish. Observe. Improve the next version.

Visibility, human traffic and paid orders answer different questions. Keep their evidence separate.

AI visibility

### Can AI retrieve and cite it?

Crawler access, parsing, available indexing and citations.

AI traffic

### Do shoppers reach your store?

Human referral visits, product views and shopping actions, separated from agent requests.

AI Orders

### What does the order evidence show?

Paid orders with available AI-attribution evidence, reconciled with refunds.

A crawl is not a recommendation. A referral is not an order. Attributed orders do not establish causal lift; that requires an experiment.

THE PRACTICAL DETAILS

## Built for clarity.

No. The system connects the knowledge behind your existing storefront. Agentic Page is an AI-readable layer on your site; the store and checkout remain in place.

Translation extracts what an asset contains, preserving its image region, time range or page. Enrichment adds reviewable meaning such as use cases, FAQs and selling points. Both keep evidence.

Commercial conditions come from authoritative systems. Hard product claims need supporting evidence. Missing and conflicting values remain visible for resolution; approval never rewrites the original source.

The page explains the target operating architecture. Agentic Page for Shopify is the current starting point. Additional source connectors, PDP writes, feeds, MCP and Skills are scoped with the team.

No guarantee. AI platforms control access, indexing and recommendations. Measure visibility, traffic and order outcomes separately, and use experiments when testing incremental impact.

## Your product knowledge.Ready for what comes next.

See how your catalog, assets and buying conditions can work as one system.

## Structured data (JSON-LD)

```json
{
  "@context": "https://schema.org",
  "@graph": [
    {
      "@context": "https://schema.org",
      "@type": "Organization",
      "@id": "https://www.deeplumen.com/#organization",
      "name": "DeepLumen",
      "url": "https://www.deeplumen.com/",
      "inLanguage": "en",
      "logo": {
        "@type": "ImageObject",
        "url": "https://www.deeplumen.com/logo.png",
        "width": 200,
        "height": 200
      },
      "image": "https://www.deeplumen.com/og-image.png",
      "slogan": "The merchant-side layer for agentic commerce",
      "description": "DeepLumen is an agentic commerce platform that helps brands become discoverable, recommendable, and transactable by AI agents.",
      "sameAs": [
        "https://www.linkedin.com/company/deeplumen/",
        "https://x.com/Deeplumen0922"
      ],
      "hasMerchantReturnPolicy": {
        "@type": "MerchantReturnPolicy",
        "@id": "https://www.deeplumen.com/#merchant-return-policy",
        "applicableCountry": "US",
        "merchantReturnLink": "https://www.deeplumen.com/legal/terms/"
      },
      "hasShippingService": {
        "@type": "ShippingService",
        "@id": "https://www.deeplumen.com/#digital-delivery-shipping-service",
        "name": "DeepLumen digital delivery",
        "description": "DeepLumen offers software, SDK, and consultation services that are delivered digitally and do not require physical shipping.",
        "fulfillmentType": "https://schema.org/FulfillmentTypeDelivery",
        "shippingConditions": {
          "@type": "ShippingConditions",
          "shippingDestination": {
            "@type": "DefinedRegion",
            "addressCountry": "US"
          },
          "shippingRate": {
            "@type": "MonetaryAmount",
            "value": 0,
            "currency": "USD"
          }
        }
      },
      "knowsAbout": [
        "Agentic commerce",
        "AI shopping agents",
        "Generative engine optimization",
        "AI search optimization",
        "Product schema",
        "Structured data",
        "llms.txt",
        "AI referral traffic",
        "M2AI"
      ]
    },
    {
      "@context": "https://schema.org",
      "@type": "WebSite",
      "@id": "https://www.deeplumen.com/#website",
      "name": "DeepLumen",
      "url": "https://www.deeplumen.com/",
      "inLanguage": "en"
    },
    {
      "@context": "https://schema.org",
      "@type": "BreadcrumbList",
      "itemListElement": [
        {
          "@type": "ListItem",
          "position": 1,
          "name": "Home",
          "item": "https://www.deeplumen.com/"
        },
        {
          "@type": "ListItem",
          "position": 2,
          "name": "Product Knowledge",
          "item": "https://www.deeplumen.com/product-facts/"
        }
      ]
    }
  ]
}
```