AI in e-commerce glossary: LLM, feed, PMax, RAG, GA4, GDPR and governance. Definitions for marketers and product owners, linked to the free course.
A language model trained on vast text corpora; generates and reasons over natural language.
In e-commerce, LLMs support copy (descriptions, RSA headlines, FAQ), report summaries, and category mapping. They do not replace ERP or pricing rules - human-in-the-loop and product-fact verification are required.
Models from Google, Meta and other platforms that optimize bids and budget allocation on conversion signals.
Smart Bidding in Google Ads and Performance Max campaign optimization are platform ML - a "black box" trained on advertising ecosystem data. You supply the feed, conversions and segments; the platform picks channels and bids.
Explicit if/then logic (e.g. margin < 10% → exclude from PMax) - predictable, cheap, easy to audit.
Rules are the first choice where conditions are known and stable: feed validation, SKU exclusions, margin thresholds. LLMs should not replace rules for price, GTIN or availability.
A fact invented by AI - a product attribute, promotion or promise not present in the source of truth.
Typical example: a shoe description with a Gore-Tex membrane that is not in the PIM. In campaigns - a non-existent price guarantee. Prevention: brief with facts from ERP, QA checklist, rejection log.
Change in AI model behavior over time - the same prompt yields different output after a provider update.
After an API version change, description or category quality can drop without any change on your side. Version prompts and retest a control sample after every model update.
Instruction for an AI model: role, context, constraints, output format and examples (few-shot).
A good e-commerce prompt includes tone of voice, character limits (RSA, meta title), banned phrases and input data exclusively from the PIM. Store versions in a repo, not only in a tool UI.
LLM answers from an approved knowledge base (FAQ, terms of service), not from model “memory”.
Critical for store chatbots: the bot cites current shipping and return policies. Without RAG the bot invents promotions - a typical process failure in the CX module.
A process where a person approves or rejects AI output before publication or a business decision.
Minimum in e-commerce: log SKU + date + author + QA status, content acceptance owner, gate before export to feed. Without this, overnight mass automation ends in incidents.
Structured file (XML, CSV, JSON) with product offers for Google Merchant Center, Meta Catalog and marketplaces.
Bridge between ERP/PIM and sales channels. Minimum: id, title, price, availability, link, image_link. Feed quality determines Shopping, PMax and dynamic remarketing results.
Source-of-truth system for product attributes - descriptions, specs, media, categories.
AI enrichment works on a draft layer with PIM as input. Without PIM the model deepens chaos: duplicates, conflicting specs, wrong GTINs.
Company operational system: prices, stock levels, costs - one-way export to the feed.
Price and availability never come from an LLM. Two sources (spreadsheet + ERP) are the most common cause of stale prices in dynamic creatives and Merchant Center disapprovals.
Enhancing attributes (description, category, custom_label) - manually, by rules or with AI assistance.
Safe pilot: 50 SKUs in a copy, manual control sample, acceptance log. Risky: google_product_category and titles without QA; forbidden: GTIN, price, legally regulated dimensions.
Global trade item number (e.g. EAN-13); in the feed exclusively from the source of truth.
AI must not "fill in" GTINs. Wrong code = Merchant Center disapproval and product matching issues. See also the full identifier dictionary.
Feed fields for campaign segmentation: margin, season, bestseller, clearance.
Example: custom_label_0 = high_margin → separate ROAS strategy in PMax. Empty labels on a "high margin" campaign mean no segmentation and budget spread on low margin.
Google hub for product feed management, error diagnostics and Shopping / PMax campaigns.
Diagnostics report shows disapprovals (missing GTIN, wrong price, broken link). AI does not replace MC audit - every data quality error from enrichment shows up there.
Official Google taxonomy assigned to a product - affects matching and reports.
An LLM can propose a category from the description, but a control sample is required. Wrong category = poor grouping in campaigns and lower ROAS on the segment.
Google campaign combining feed, conversion signals and creative assets; optimized by platform ML.
You supply feed quality, primary conversions and custom_label segments. Good ROAS at 3% margin is a signal to exclude low_margin, not to double the budget.
Automated Google Ads bid strategies (Target CPA, Target ROAS) powered by platform ML.
Requires sufficient conversions and a correct purchase tag. Signal quality drops after strict Consent Mode - interpret reports cautiously.
Google text ads with multiple headlines and descriptions - the platform tests combinations.
AI speeds up variant generation (≤30 / ≤90 characters), but requires a brand brief and A/B testing. Promises like "lowest price" without proof = disapproval or legal risk.
ROAS - return on ad spend; CPA - cost per acquisition.
Platform metrics may differ from incrementality and ERP. Budget decisions should combine Ads reports, GA4 and lift tests - do not scale on Meta ROAS alone.
Ads built automatically from the product feed and graphic templates.
Stale prices on a banner = first audit price sync with ERP, not a new prompt. image_link and title quality affect CTR and Google policy compliance.
Customer segmentation: when they last bought, how often, how much they spent.
AI helps describe segments in business language and propose tests. VIP segment in CRM without on-site identification = consent or technical issue, not missing LLM.
Tailoring content, banners or recommendations to a user segment.
Remarketing and ad personalization require consent where the law requires it. Compare weeks before/after CMP with a note on data quality.
Store assistant - safest as LLM + RAG on approved FAQ with escalation to a human.
Root cause of promises outside terms: missing current FAQ in the RAG base. Measure escalations per 100 conversations and off-policy promise incidents.
"Frequently bought together" widget - association rules or CDP platform ML, not LLM as the SKU engine.
LLM can write widget copy. Pots recommended with smartphones = wrong category rule, not a bad prompt. Measure widget CTR and AOV of sessions with a click.
Funnel view_item → add_to_cart → begin_checkout → purchase with items parameters (item_id, price).
Purchase revenue −40% with stable ERP = audit the tag, not cut budget. item_id must match id in the feed.
Product identifier in GA4 events - must match id in the feed and PIM.
item_id mismatch breaks product reports and campaign optimization. Test in DebugView: one test order with full items parameter.
Measuring additional campaign effect above baseline (geo test, holdout, lift study).
Meta +30% ROAS with incrementality +2% = cautious scaling. Platform report and GA4 attribution are analytical perspectives, not the only business truth.
Assigning conversions to channels and touchpoints in the customer journey.
The model in GA4 changes the view, it does not create incremental lift. Combine with incrementality tests for million-dollar budget decisions.
Google mechanism adjusting tags to cookie consent; modeled conversions on refusal.
Drop in GA4 conversions after CMP rollout does not always = drop in sales. Compare purchase with ERP week over week.
Document: allowed use cases, forbidden data in prompts, acceptance owner, quarterly review.
Without a policy after a 90-day pilot - everyone uses their own ChatGPT. Link to the company Data Governance hub.
Data processing agreement with AI provider; no pasting full customer databases into public APIs.
Allowed: aggregates without PII, public product descriptions. Stop an export of 10,000 emails to ChatGPT - that is an incident, not a "quick campaign".
Flow: PIM trigger → LLM (draft) → ERP validation → human acceptance → feed.
Two gates minimum before overnight publication. Version prompts in a repo. Measure records stopped at the gate vs errors after publication.