AI Ecommerce Glossary

Abandoned cart recovery is the practice of re-engaging shoppers who filled a cart and left without buying, traditionally via email sequences, now increasingly via real-time AI chat. There is a great deal of it to work with: Baymard Institute puts the average documented cart abandonment rate at 70.22%, averaged across 50 studies. Not all of that is a lost sale. The same research calls a large part of it largely unavoidable, because shoppers are window shopping, comparing prices or saving items for later.

The old approach is a three-email sequence starting an hour after abandonment. The AI approach is detecting exit intent while the shopper is still on the page and opening a conversation right then: asking what held them back, answering an objection, or surfacing the size variant they couldn't find. That is the job an AI sales agent does on a product page.

Timing is the part most stores underrate. A shopper still on the page has their intent intact, while an email the next day has to rebuild it from nothing. That is the argument for intervening on site rather than afterwards, and it is why the two approaches are not interchangeable.

The version of online shopping where AI acts on a customer's behalf: researching products, comparing prices, reading reviews, and placing orders, all with varying degrees of human supervision.

It sounds like science fiction. It isn't. Salesforce, Deloitte, and Stripe all published major reports on agentic commerce in 2025. Perplexity's Shopping Assistant, Google's Project Mariner, and early Shopify pilot programs are live examples. The shift is already underway.

The implication for store owners is practical: an AI agent browsing your store isn't going to respond to lifestyle photography and flash banners. It needs clean product specs, accurate inventory signals, and well-structured policy data. How your product catalogue is formatted today determines whether AI shoppers find and buy your products next year. A product page audit is usually where that gets checked first.

A step above a standard AI chatbot, an AI Concierge handles personalised, high-touch shopping guidance the way a knowledgeable salesperson on a shop floor would. It asks questions, learns preferences, and makes considered recommendations rather than just responding to what's directly asked.

The difference shows up in edge cases, which is where worked chatbot conversations are more use than a specification. Ask a basic chatbot "what's good for combination skin under $40?" and you'll get a list of products that match those keywords. A concierge asks a follow-up: "Do you have any sensitivities? Day or night use?" Then it narrows to one or two genuinely right options.

The best implementations remember context across a session and, where a customer consents, across repeat visits.

When an AI generates a confident, fluent answer that is factually wrong. In general use, this is annoying. In ecommerce, it's a commercial and legal liability.

An AI that hallucinates might tell a customer your return window is 60 days when it's 30, confirm a product is in stock when it isn't, or recommend a supplement as suitable for people with a specific condition, incorrectly. These aren't hypothetical failures. They've happened, publicly, to recognisable brands, and getting things wrong is the thing people say worries them most about AI, ahead of anything to do with capability.

An AI system that goes beyond answering FAQs to actively driving sales: handling product discovery, objection handling, personalised recommendations, upselling, and assisted checkout, autonomously, at any hour.

The distinction from a standard chatbot is important. A chatbot responds. A sales agent initiates, persists, and steers toward a commercial outcome. A chatbot tells you the return policy. A sales agent notices you've been comparing two jacket sizes for six minutes and asks which one you're on the fence about, then closes the sale.

Most tools marketed as "AI Sales Agents" are chatbots with a CTA button. An actual AI Sales Agent is trained on your specific product catalogue, understands your brand voice, tracks shopper behaviour in real time, and converts that behaviour into revenue, measurably. Telling the two apart before you buy either is most of the work, and what an AI sales agent actually does runs further than a definition can.

The average amount a customer spends per transaction. Simple to calculate: total revenue divided by total orders. Not simple to move.

AOV matters because customer acquisition is expensive. Every dollar you increase AOV is pure margin improvement without additional spend to get someone to the store. A customer who spends $85 instead of $60 is a 42% revenue lift from the same traffic.

AI moves AOV through real-time upselling and cross-selling: not with generic "customers also bought" carousels, but with contextually relevant suggestions surfaced at the moment the shopper is most receptive. The conversion gap between a generic product recommendation and a well-timed contextual one is substantial.

The real-time signals a shopper sends through their on-site behaviour, and the logic an AI uses to decide when and how to respond to them.

Examples: pausing on a product page for over 90 seconds (likely interest or confusion), switching between two similar products (indecision), viewing the returns policy before adding to cart (purchase hesitation), reaching checkout and stopping (friction). Each of these signals something specific about where the shopper is in their decision.

The value isn't in the signals themselves. It's in what the AI does with them. A shopper stuck on sizing needs a different message than one who's hesitating on shipping cost. The more granular the trigger logic, the more relevant the intervention, and the better the outcome.

The share of visitors who arrive on a page and leave without interacting with anything on it. On a product page a bounce usually means the shopper looked, decided the page did not answer their question, and went back to wherever they came from.

It is a blunt measure and it is often misread, which is why it belongs among the other numbers an ecommerce audit puts side by side rather than on a dashboard of its own. A high bounce rate on an article can be perfectly healthy: the reader got the answer and left satisfied. On a product or category page it rarely is, because the next step was right there and nobody took it.

What it cannot tell you is why. That is the gap real-time assistance fills: a shopper about to bounce because the sizing is unclear can be asked what they were looking for while they are still on the page.

A program that holds a conversation with a shopper in a chat window. The word covers two very different things: a scripted bot matching a question to a pre-written answer, and a generative one composing a reply from a model and from your own data.

The scripted kind is predictable and cheap, and it fails the moment a shopper phrases something it was not told about, which is most of the time and the usual reason an ecommerce chatbot disappoints. The generative kind handles the phrasing but will invent an answer unless it is grounded in a real source, which is what retrieval is for.

The word itself is now too broad to buy on. The questions worth asking are where the answers come from, what happens when the system does not know, and who the conversation reaches when it needs a person.

Selling through conversation: using chat, messaging, and AI to guide shoppers through product discovery and purchase rather than leaving them to navigate a catalogue alone.

The term was coined by Uber product lead Chris Messina in 2015, in the context of messaging apps. In ecommerce today, the most relevant form is AI-powered chat on your store's own website: a system that answers questions, surfaces products, handles objections, and closes sales through natural dialogue.

The business case is solid: shoppers who engage in a chat session convert at higher rates and spend more than those who don't. The challenge is making the conversation feel like genuine help rather than like filling out a form with extra steps. That gap, between a scripted FAQ bot and a real conversational experience, is where most implementations either win or lose.

The percentage of website visitors who complete a purchase. If 100 people visit your store and 2 buy, your conversion rate is 2%.

Published benchmarks disagree with each other, because they count sessions differently and sample different sectors, but they cluster around 2% to 3% across all sectors, with the strongest stores at roughly double that. The gap looks small in percentage terms and is enormous in revenue terms. It represents thousands of additional sales from exactly the same traffic volume.

Conversion rate is the lever most stores under-invest in relative to traffic acquisition. A 25% increase in traffic grows revenue 25%. A 25% improvement in conversion rate grows revenue 25%, but the second approach is almost always cheaper than the first. Buying more ads doesn't improve conversion rate. Better product pages, faster checkout, and real-time shopping assistance do. A free AI site audit is one way to see which of the three is costing you most.

The total revenue a single customer generates across all their purchases, for as long as they remain your customer. Sometimes written as LTV (lifetime value).

CLV matters for two distinct reasons. First, it tells you how much you can afford to spend acquiring a customer: someone worth $800 over three years justifies a much higher acquisition cost than someone who buys once for $40 and never returns. Second, it's the metric that post-purchase AI personalisation moves most directly.

Most stores optimise for the first purchase. CLV-focused stores design every subsequent touchpoint to extend the relationship: the follow-up chat, the next recommendation, the loyalty message six months later. AI makes this kind of personalised retention economically viable at scale, without requiring a customer service team that grows in proportion to the customer base.

A structured object a page publishes about itself: which product is being viewed, at what price, in what currency, what sits in the cart. Analytics tools, tag managers and AI systems read it instead of trying to scrape the same facts out of the markup.

Stores usually discover the data layer when something stops being measurable. A tag fires on the wrong page, revenue is counted twice, a product's category arrives empty. The cause is almost always the same: the page never published the fact, so every tool downstream is guessing.

It matters more as AI arrives, not less. An agent that can read what a page is actually about does not have to infer it, and a clean data layer is the cheapest way to say so.

Adjusting prices in real time based on demand, inventory levels, competitor pricing, and shopper behaviour, instead of setting a fixed price and leaving it. AI makes this practical at scale, recalculating across thousands of SKUs continuously rather than in occasional manual reviews.

The risk is trust. Shoppers who notice a price move between visits can feel manipulated. The stores that do it well keep the swings modest, always honour the price a shopper already has in their cart, and use it to protect margin on high-demand items rather than to squeeze every visitor. Whether it suits your catalogue at all is a strategy question before it is a technical one, which is what ecommerce consulting is for.

A list of numbers that represents the meaning of a piece of text or an image, arranged so that things with similar meaning sit close together. It is what lets a machine treat “waterproof jacket” and “rain coat” as near neighbours without anyone writing a rule that says so.

Every product in your catalogue can be turned into one, and so can every question a shopper types. Finding the right product then becomes a matter of finding the nearest neighbours, which is a calculation rather than a keyword match.

This is the machinery under semantic search and under the retrieval step in an AI assistant. It is also why the quality of your product copy matters more than it used to: an embedding of a thin, templated description carries thin, templated meaning.

A behavioural signal detected when a shopper is about to leave a page, typically through mouse movement toward the browser tab or address bar, a sudden scroll to the top, or idle time crossing a threshold.

Exit intent is valuable because the moment just before someone leaves is often the highest-leverage intervention point. They've already invested time in your store. Something stopped them from converting. A well-timed, relevant message at that moment converts at meaningful rates.

What stores usually get wrong: a generic 10%-off popup triggered on exit intent performs poorly because it ignores what the shopper was actually doing. "Still thinking about the Merino pullover in size M?" performs better, because it signals that the store noticed. Relevance is the variable that separates a useful intervention from an irritating one, and getting it right is as much UX and UI design as it is targeting.

The filters down the side of a category page: size, colour, price, brand, availability. Each facet narrows the set, and the counts beside them tell a shopper what is left before they commit to the click.

Facets work when the catalogue data behind them is clean and fail quietly when it is not. A colour filled in as “navy” on some products and “Navy Blue” on others splits one facet into two, and a shopper filtering for one never sees the other.

They also only help someone who already knows what to narrow by. A shopper who cannot name what they want is better served by a conversation than by a longer list of checkboxes.

Behavioural and transactional data collected directly from your own customers on your own properties: pages viewed, products clicked, purchase history, time spent, search queries.

First-party data matters more than it used to because the tracking infrastructure that let brands follow shoppers around the internet for years is being dismantled. Browser privacy changes, iOS updates and GDPR enforcement have steadily eroded third-party cookies. Brands that built their targeting strategy on borrowed data are feeling it.

What replaces it is data you own and control. AI systems use first-party data to personalise recommendations in real time, predict which customers are at risk of churning, and trigger the right message at the right moment, without needing to reach outside your own ecosystem. The richer your first-party data, the more precise the personalisation.

AI that creates new content (text, images, code, product descriptions, chat responses) rather than simply retrieving or classifying existing information.

GPT-4, Claude, Gemini, and Llama are all generative AI models. In ecommerce, generative AI powers: chat responses written fresh for each individual shopper's specific question; product description generation at scale; dynamic email copy personalised to individual customers; and image generation for product variations or marketing creative. Which of those repay the effort, and which only look like progress, is the difference between AI in ecommerce and the hype around it.

The distinction that matters operationally: a rule-based chatbot retrieves pre-written answers from a script. A generative AI chatbot writes a response for the specific question being asked, in the moment. The second approach handles edge cases the first can't. But it also requires solid grounding in accurate, up-to-date store data to avoid hallucinations.

The work of making a store legible to AI answer engines and shopping agents rather than only to a page of search results. Where SEO competes for a position in a list, GEO competes to be the source an answer is built from.

Much of it is familiar: clean structured data, accurate product attributes, pages that state facts plainly instead of burying them in marketing copy. Some of it is not. An answer engine quotes and attributes, so the question becomes whether your page is the one worth quoting on a given claim.

The practical test is simple. Ask an assistant a question a customer would ask about your category and see whose pages it leans on. If your catalogue is missing from the answer, the gap is usually structural rather than editorial.

An architecture that separates the storefront a shopper sees from the commerce engine holding products, prices and orders, with the two talking over an API. The front end can then be rebuilt, or replaced, without touching the system of record.

The appeal is freedom: a storefront in whatever framework the team prefers, several of them against one catalogue, and a change to the look that does not wait on a platform release. The cost is that everything a monolithic platform gave you for free is now something someone owns and maintains.

It is a real answer to a real constraint and a poor answer to boredom with a theme. The question worth asking first is which specific thing the current front end stops you doing, and whether anything short of a replatform would unblock it.

A setup where AI handles the routine work but a person stays in the loop for the decisions that need judgement: reviewing, approving, or correcting the AI's output before it reaches the customer.

In ecommerce it's the difference between "the AI answers every question unsupervised" and "the AI drafts, a human approves the edge cases." It's how teams get the speed of automation without handing brand voice, refunds, or sensitive product claims over entirely.

The AI's ability to understand what a shopper actually wants, the intent behind their message, rather than just the literal words they typed.

This matters because people don't talk to search boxes the way they talk to each other. "Something for date night under $100" is a clear shopping intent with an occasion, a budget, and a purchase goal. Keyword search can't handle it. A well-trained intent recognition system can: it identifies the context, the category, the constraint, and the intent, then surfaces relevant options.

Intent recognition also distinguishes between messages that look similar on the surface but mean different things. "How long does shipping take?" from a first-time visitor is a pre-purchase question. The same message from someone who ordered three days ago is a support request. The right response is completely different, and a system without good intent recognition will get this wrong frequently.

The search box on your own store, and the results it returns. Shoppers who use it are further along than those who browse: they have named what they want, which makes the quality of the answer disproportionately valuable.

It is also the most honest feedback channel a store has. The queries people type are the words they actually use for your products, and the ones that return nothing are a list of things they expected you to sell, or expected you to call something else.

Most stores never read it. The log is sitting in analytics they already pay for, and it usually takes an afternoon to find the ten queries costing the most.

Whether the content of a page exists in the HTML the server sends, or appears only after the browser has run the page's JavaScript. For a reader the difference is invisible. For a crawler, an answer engine or a shopping agent, it decides whether the content exists at all.

Search engines do render JavaScript, but on their own schedule and not always completely, and the simpler agents that read pages on a shopper's behalf often do not render it at all. A price, a stock state or a description that appears only after a script has run is a fact you have chosen not to publish, which is why rendering belongs on an ecommerce SEO checklist rather than in a backlog of its own.

The check takes a minute: open the page source rather than the inspector, and look for the text you care about. If it is not there, it is not guaranteed to be seen.

A lightweight format for describing your products, prices, reviews, and policies in a way machines can read directly, embedded invisibly in a page's code. It's the difference between a search engine or AI guessing what's on a page and being told, explicitly.

As AI shoppers and answer engines grow, structured data stops being a nice-to-have for rich search snippets and becomes how your catalogue gets understood at all. Clean JSON-LD is one of the highest-leverage, lowest-glamour investments a store can make, and it is one of the first things a product page audit looks at.

The curated set of information an AI draws on to answer accurately: your product specs, policies, FAQs, shipping rules, and sizing guides. It's the "source of truth" the model is grounded in, rather than its general training data.

The quality of an AI sales agent is capped by the quality of its knowledge base. A well-maintained one means correct, on-brand answers; a stale or thin one is exactly where hallucinations and wrong policy quotes come from.

A model of the things in your business and how they relate: this product belongs to this category, is made by this brand, fits this model, replaces this discontinued line. Where a knowledge base holds documents, a knowledge graph holds the relationships between facts.

It is the difference between an assistant that can find the right page and one that can answer “will this fit my car”. The second needs to know that a part relates to a vehicle, which relates to a year range, and that a near match is not a match.

Most stores already hold the pieces, scattered across attributes, categories and a spreadsheet someone maintains by hand. Joining them up is usually a data exercise rather than an AI one.

The type of AI model behind most modern chatbots and AI assistants, trained on vast amounts of text to understand language, generate human-like responses, and reason through complex questions.

GPT-4 (OpenAI), Claude (Anthropic), Gemini (Google), and Llama (Meta) are the major LLMs. Most AI products in ecommerce are built on top of one of these foundations. The vendor's product sits above, adding domain-specific logic, product data grounding, and guardrails.

The delay between asking something and getting an answer. For an assistant on a storefront it is measured in the seconds a shopper spends watching a blinking cursor, and it is the difference between a feature that feels alive and one that feels broken.

The number people quote is usually the model's. The number the shopper feels includes everything around it: the lookup into your catalogue, the retrieval step, the network, the rendering. A model that answers in under a second can still take six seconds to say anything useful.

Test it on the connection your customers actually have rather than on office wifi, and decide in advance what the assistant shows while it is thinking. Silence reads as failure.

Deciding what a shopper sees first and in what order: which products lead a category, what fills the recommendation slots, what gets pushed when stock is heavy or margin is thin. Online it is a set of rules and sorts rather than a shop floor, but the job is the same one.

The tension is always between the rule and the shopper. A hand-picked order expresses what the business wants to sell. An automated one responds to what this particular person is likely to buy. Neither is right on its own, which is why most stores pin a few positions and let the rest sort itself.

AI changes the granularity rather than the principle. The same category can be ordered differently for someone who has been browsing waterproofs than for someone who arrived on a sale link, without anyone maintaining two category pages.

AI that works across more than one type of input (text, images, and sometimes audio) in a single model. A shopper can upload a photo and ask "do you have this in blue?" and get a real answer, not a shrug.

For ecommerce this unlocks visual search, image-based product matching, and richer product understanding. A multimodal agent can effectively "see" your product photos, not just read the text around them.

Picking the meaningful things out of a sentence and labelling them: in “do you have the merino crew in navy, medium”, a product type, a colour and a size. It is the step that turns a sentence into something a catalogue query can use.

It is unglamorous, and it is where most conversational search quietly fails. A system that recognises “navy” as a colour but not “storm blue”, or that reads “medium” as a clothing size when the shopper meant a dog harness, gives a confidently wrong answer rather than no answer.

How well it works depends on your own attribute data more than on the model. An entity that is never named in your catalogue is an entity it cannot match.

The branch of AI that allows computers to understand, interpret, and respond to human language: not just keyword matching, but actual comprehension of meaning, context, and nuance.

NLP is why a modern AI can handle "do you have those boots in a darker colour and half a size bigger?" and give a useful answer. Without it, that question fails because no keyword matches "darker colour" to a product attribute or "half a size bigger" to an inventory filter.

NLP has existed for decades. What changed recently is accuracy and scale. Modern NLP handles context across multiple messages in a conversation, sarcasm, ambiguous phrasing, and the difference between "I love this product" said sincerely and said ironically. That quality gap is what separates a genuinely useful shopping assistant from one that frustrates every third customer.

Selling across multiple channels (website, mobile, social media, marketplaces, physical store) in a way that feels consistent and connected to the customer, regardless of where they engage.

The operative word is connected. Multichannel means selling in multiple places. Omnichannel means those channels share data, so the customer experience doesn't fragment across them. A customer who adds to cart on mobile and completes checkout on desktop doesn't start over. A customer who asks a question in Instagram DMs gets the same service quality as one chatting on your website.

AI is increasingly the connective tissue in omnichannel setups: maintaining customer context across touchpoints, routing inquiries appropriately, and personalising messages based on the full cross-channel picture rather than a partial one. Getting the channels to share that picture at all is integration work, and connecting those systems is a project in its own right.

The system holding what was ordered, what has been paid for, what has shipped, what came back and what is owed. It sits between the storefront, the warehouse, the accounts and the carriers, and it usually decides whether a promise made on the product page can be kept.

Stores outgrow a platform's built-in order handling at predictable moments: a second sales channel, a second warehouse, partial shipments, exchanges. The symptom is almost always a spreadsheet someone reconciles on Mondays, and the fix is usually connecting the systems properly.

It matters for AI because an assistant can only answer “where is my order” as well as the system it asks. Grounding a chat in a stale order feed produces a confident wrong answer, which is worse than no answer.

AI-generated product suggestions tailored to each shopper based on their browsing behaviour, purchase history, stated preferences, and real-time signals.

The figure everyone quotes, that 35% of what people buy on Amazon comes from its recommendations, is a McKinsey estimate from 2013 and it has not held up well. A randomised experiment on a fashion retailer's site, published in Information Systems Research in 2019, measured the net lift at about 11%: the recommended products sold 9% more, while the product being viewed sold 1.9% less. Eleven per cent is well worth having. It is also a different order of claim from a third of revenue.

The quality gap between implementations is large. "Customers who bought X also bought Y" is blunt and frequently irrelevant. You've already seen it be wrong. Recommendations that account for what someone has already purchased, what price range they've been browsing, and what they're looking at right now are meaningfully more useful. And they convert at meaningfully higher rates. An AI sales agent can ask the question a carousel has to guess at.

An AI system that initiates actions (starting conversations, surfacing recommendations, sending messages) based on predicted intent, rather than waiting passively for a user to do something first.

Most chatbots are reactive: they sit on the page until someone clicks the chat bubble. Proactive AI watches what shoppers are doing and decides when to step in, based on a combination of behavioural signals and predicted intent models.

Published figures put proactive engagement rates several times higher than passive ones, but they come from the companies selling chat software rather than from independent research, so read the multiple as a direction and not as a measurement. There is also a quality threshold. A popup that fires five seconds after anyone lands on a page isn't proactive intelligence. It's just an annoying popup with extra steps. The trigger logic is what separates useful proactivity from irritating interruption.

A chat interaction initiated by the AI, based on real-time shopper behaviour, rather than opened by the customer clicking a chat button.

Proactive chat is one of the highest-ROI applications of AI in ecommerce. A visitor who's spent two minutes on a product page and scrolled back to the top has an unanswered question, and an unanswered question is the quietest way a store loses revenue. A proactive message like "Trying to decide between sizes? I can help." catches them at exactly the right moment.

Timing and relevance determine whether this feels helpful or invasive. Proactive chat that fires on page load with "Hi there! How can I help today?" performs poorly. It's a generic interruption. Proactive chat triggered by specific signals, with a message relevant to what the shopper is actually looking at, performs substantially better. The content of the trigger matters as much as the fact of it.

The process by which shoppers find products they want to buy, and how AI changes that process from a keyword search into a guided conversation.

Traditional product discovery relies on the shopper knowing what they're looking for and being able to describe it in terms that match your product taxonomy. This works for someone searching "black running shoes size 10." It fails for gift-buyers ("something for my dad who likes cooking"), shoppers with an aesthetic preference they can't describe in keywords ("minimal, earthy tones, nothing too trendy"), and anyone who doesn't know the technical name for what they need.

AI-powered product discovery handles natural language queries, guided "tell me about yourself" flows, and follow-up clarifying questions, routes that keyword search simply can't serve. For stores with large or complex catalogues, the impact on conversion is especially significant. An AI sales agent is one way to put that conversation on the page.

Interpreting what a shopper actually means, not just the words they typed. "Something warm for a toddler under £30" is a sentence a keyword search chokes on but a query-understanding system parses into intent, constraints, and budget.

It's the front door of any good AI shopping experience. Get it wrong and even a great catalogue returns zero results; get it right and vague, natural questions map cleanly to the right products.

A panel that opens over a category page showing a product's images, options and add-to-cart without leaving the grid. It exists so a shopper can check three candidates quickly instead of making three round trips to product pages and back.

It helps most where shoppers compare within a set, and gets in the way where they need the detail a full page carries. A quick view that hides the size guide, on a product people only buy once they have checked the size guide, saves a click and loses the sale.

Whatever it shows, it has to be reachable by keyboard and announced when it opens, because it is a dialogue over the page rather than a new page.

A technique that connects a language model to a specific knowledge base (your product catalogue, return policy, shipping information, FAQs) so the AI gives accurate, grounded answers instead of generating responses from general training data alone.

Without RAG, an AI answers questions based on what it learned during training. That's fine for general knowledge questions. It's a problem when someone asks about your specific return window, or whether a particular product comes in a specific variant, or what your current shipping times are. Without RAG, the model either says it doesn't know or, worse, fabricates a plausible-sounding answer.

With RAG, the model retrieves the relevant, current information from your own data before generating a response. The answer is grounded in reality. This is the core technical difference between an AI that's safe to deploy on your store and one that will create customer service problems within a week.

The share of orders that come back. It is counted by units, by orders or by revenue depending on who is asking, which is the first thing to pin down, because the three numbers are not close to each other.

In some categories a high rate is the business model working as intended. In others it is a sign that the product page is writing cheques the product cannot cash. The useful split is between returns caused by the item and returns caused by the description of it.

The second kind is the one a store can fix: sizing stated rather than implied, photographs that show scale, materials and measurements where a shopper looks for them, and an assistant that can answer the question before the order rather than after it. What happens afterwards is its own discipline, and the post-purchase stretch decides whether a return becomes an exchange or a lost customer.

Search that matches on meaning rather than exact keywords. "Gift for someone who loves cooking" surfaces relevant products even when none of them contain those words, because the system compares concepts, not strings.

It's powered by embeddings stored in a vector database, and it's a core reason AI-native stores convert better on search: a near-miss query returns useful results instead of an empty page.

Reading a piece of text for the feeling in it rather than the facts: whether a review is happy, whether a support message is angry, whether a conversation is going well or badly. It produces a label or a score, not an understanding.

On a store the useful applications are narrow and real. Routing a furious message to a person instead of a bot. Noticing that reviews for one product turned negative after a supplier change. Flagging a conversation that is about to end badly while it is still open.

The failure mode is sarcasm and understatement, which are common in exactly the messages you most want to catch. It is a reasonable trigger for a human to look, and a poor basis for a decision taken without one.

The percentage of incoming customer support requests resolved automatically by AI, without a human agent getting involved.

Most support queues are dominated by a small set of repeating questions: order status, return policy, shipping estimates, size charts. These questions have correct answers. An AI that's properly grounded in your store's data handles them accurately, at any hour, in seconds. The ticket deflection rate tells you what fraction of that volume you've automated.

Vendors publish deflection rates in the 60% to 70% range. Those are their own numbers rather than independent ones, so read them as the top of the range rather than the norm. The arithmetic underneath is what matters: a team handling 500 support tickets a month at 60% deflection answers 200, not 500. Same staffing level, far less volume, and faster replies on the complex cases that actually need a person. Where that volume is already being paid for per seat or per ticket, it is also the clearest case for replacing the SaaS tool with something you own.

The unit a language model reads and writes in, roughly a short word or part of one. Models are priced by it, limited by it, and billed for both what you send and what comes back.

It becomes a practical concern the moment an assistant is grounded in your own data. Every product description, policy and past message you put in front of the model is tokens, and a system that pushes the whole catalogue into every answer is both slower and dearer than one that retrieves the few things that matter.

The context limit is the other half. A model has a ceiling on how much it can hold at once, and a long conversation eventually pushes the earliest part of it out of view, which is why an assistant can forget what a shopper said ten messages ago.

Upselling is persuading a customer to buy a higher-value version of what they're already considering. Cross-selling is recommending complementary products alongside their intended purchase. Both drive average order value.

In AI ecommerce, neither looks like "would you like fries with that?" AI upselling is specific: noticing that a shopper has been comparing two camera models and spending disproportionate time reading video spec sections, then surfacing the next tier with a focused explanation of what the upgrade actually gets them: not a list of features, but the relevant one.

Cross-selling done well feels like service. Someone buying a camera should probably know they'll need a memory card and a case, not because the store wants to sell more things, but because arriving without those accessories is a bad experience. The difference between an AI that cross-sells effectively and one that annoys customers is whether the suggestions feel relevant or random.

Anything your customers made that ends up on your store: reviews, ratings, questions and answers, photographs of the product in use. It is content you did not write and cannot fully control, which is exactly where its value comes from.

It does two jobs at once. A shopper reads it because it is the nearest thing to asking someone who already bought the thing. A search engine and an answer engine read it because it is the part of the page written in the words customers actually use, which is rarely the words in the product title.

It is also a support queue in disguise. The questions people ask under a product are the questions your product page failed to answer, and the list of them is free.

A database built to store and search "embeddings": numerical representations of meaning. It's what lets a system find the most relevant products, documents, or answers based on similarity rather than exact matches.

It's the engine behind semantic search and RAG. When an AI agent grounds an answer in your catalogue, a vector database is usually what retrieves the right pieces of information to ground it in.

A shopper asking out loud instead of typing, through a phone, a speaker or a car. The queries that result are longer, phrased as whole questions, and almost never match the keyword a category page was built around.

The practical consequence is not a separate voice strategy. It is that the gap between how people ask and how catalogues are organised keeps widening, and the stores that answer questions plainly on the page are the ones that get picked up. In practice that means writing product pages for the way people ask out loud, which is mostly a copy job rather than a technical one.

It overlaps heavily with what answer engines and shopping agents need, which is the useful part: the same work serves all three.

When an AI agent passes a conversation to a human without making the customer start over, carrying the full context, the question, and what's already been tried across to the person taking over.

It's what separates a helpful AI layer from a frustrating one. A cold handoff ("please hold, someone will contact you") loses the moment; a warm one keeps the momentum and the intent intact.

A list a shopper saves for later, on purpose. Unlike a cart it carries no intent to buy today, which is what makes it useful: it is a statement about taste and budget made without pressure.

Most stores treat it as storage. It is better read as a signal. A product sitting on wishlists and not selling is usually a price objection or a stock problem, and a shopper whose saved item comes back into stock has told you exactly what to say to them.

It only works if it survives. A wishlist that demands an account before it will save anything mostly measures who was willing to make an account.

A structured file that lists your products and their attributes (title, price, availability, images, identifiers) in a format shopping channels and AI systems can ingest wholesale. It's how your catalogue reaches Google Shopping, marketplaces, and increasingly AI shopping agents.

The feed is only as good as its data. Missing attributes, stale prices, or vague titles quietly suppress reach. As agentic commerce grows, a clean, complete feed becomes a direct sales channel, not just a marketing one. Keeping it accurate as the catalogue changes is an integration job, not a one-off export.

A file listing the URLs on your store that you want crawled, with a note of when each last changed. It does not make a page rank and it does not stop a page being found; it tells a crawler what exists and where to look first.

On a large catalogue it earns its keep. Products come and go, variants generate URLs, filters generate more, and a crawler left to discover all of it by following links will spend its time in the wrong places. A sitemap listing the pages you actually want indexed, and nothing else, is a budget decision.

The common faults are quiet ones: URLs that redirect, pages blocked in robots.txt listed anyway, a file nobody has regenerated since a replatform. None of them produce an error anyone sees.

A comparison of a metric against the same period a year earlier, this June versus last June, rather than against last month. It strips out seasonality, so you're comparing like with like.

For ecommerce it's the honest way to judge growth: monthly numbers swing with holidays and promotions, but YoY tells you whether the business is actually bigger than it was. Watch it for revenue, conversion rate, and average order value.

Setting price and availability together to get the most out of a fixed amount of stock over a fixed period. It came out of airlines and hotels, where a seat unsold at departure is worth nothing, and it transfers to retail wherever the stock has a deadline.

Seasonal lines, perishables, end-of-life models and anything tied to an event behave the same way. A coat that has not sold by February is not worth what it was in October, and the decision is when to move the price rather than whether to.

Where it does not transfer is everything a store can simply reorder. Perpetual stock has no deadline, so the discipline becomes ordinary pricing, and treating it as yield management is a good way to teach customers to wait for the discount.

Information a shopper explicitly and voluntarily shares with a brand (quiz answers, stated preferences, wishlist items, gift intent) as distinct from behavioural data you infer from tracking.

The value of zero-party data is two-fold: it's accurate (the customer told you directly, not inferred from proxy behaviour), and it's privacy-compliant by definition (they chose to share it). As third-party tracking continues to erode, zero-party data becomes the premium input for personalisation.

AI sales agents are a natural collection mechanism. A conversational flow that asks "Who are you shopping for? What's the occasion? What's your budget?" collects high-quality preference data in a way that feels like helpful service rather than a data grab. That data then feeds better recommendations, more relevant email campaigns, and a personalisation engine that improves with every conversation. Done well, both sides benefit: the store gets richer data, the shopper gets a better experience.

A search on your own store that comes back with nothing. Every one of them is a shopper who told you what they wanted, in their own words, and was shown an empty page.

Very few are a real absence of stock. Most are a vocabulary gap: a synonym the catalogue does not carry, a brand spelled the way customers spell it, a category the industry calls one thing and everyone else calls another. Some are a typo the search was too literal to forgive.

The list is short, it is specific to your store, and it is usually the fastest thing on this page to act on. Fixing the top ten queries costs an afternoon of synonyms and redirects, and the page they land on afterwards is one you chose.