Help Articles – Page 5
What a customer tells a chat assistant is some of the clearest preference information a store will get, and it is usually lost when the conversation ends. Target showed how to keep it safely.
- Extract only signals that match a fixed list of categories, brands and attributes.
- Filter out anything sensitive or unrelated to shopping.
- Store each signal as a time-bound event with a confidence score and a record of where it came from, used alongside clicks and purchases.
Those rules allow for how conversations really work: a shopper may be buying for someone else, may change their mind next month, or may simply be wrong about what they want.
The assistants Target and Infobip showed at RecSys 2026 start from a goal, such as planning meals for the week or getting everything needed to make pizza. Infobip’s asks up to three questions to fill gaps like budget or dietary needs, then searches the catalogue and returns a grouped basket with prices.
Target’s assistant also draws on the shopper’s purchase history, translated into plain-language preferences such as “likes cooking from scratch” or “affinity for Thai cuisine,” and tells the shopper where those came from. Infobip’s assistant searches a catalogue index kept in sync with price and stock changes, so it does not recommend products that are unavailable.
Traditional recommendation picks items from a catalogue and ranks them. Generative recommendation uses AI to produce the recommendation: a short overview of a news story, a push notification written for one person, a five-day meal plan with a shopping list, or a grouped basket for a request like “everything I need to make pizza.”
The term also describes a technical approach in which a model predicts the next item piece by piece, the way a language model writes text. Spotify uses this approach in production.
RecSys is the annual ACM Conference on Recommender Systems. It brings together academic researchers and the industry teams who build recommendation and personalization at companies including Google, YouTube, Amazon, Netflix, Spotify, Meta and Target.
RecSys 2026 was the 20th edition, held in Minneapolis from September 27 to October 2, with the main conference running September 29 to October 1. The opening keynote came from Margaret Mitchell, Chief Ethics Scientist at Hugging Face, and the closing keynote from Thorsten Joachims of Cornell University.
A recommender system decides what to show each person: the products in a “you might also like” row, the order of search results, the stories in a feed, the next video, or which offer goes into an email. It learns from behavior such as clicks, views and purchases, from what people tell it directly, and from what similar customers did.
At RecSys 2026 the definition widened. Google, Target, Spotify and Infobip showed systems that write or assemble the recommendation itself, such as a news summary, a notification, a meal plan with a shopping list, or a full basket from a single request.
Three patterns recur. Unowned decisions: when two people sit in the accountable seat for the same milestone, the milestone stalls, and no amount of engineering capacity fixes it. Migrating legacy workarounds forward instead of categorising each one as replicate, refactor or retire, which faithfully rebuilds the problem you paid to escape. And no rollback plan with no named authority to trigger one, which turns cutover night into a judgement call made under pressure by whoever is still awake.
On budget specifically, research across 5,392 IT projects found a mean cost overrun of 1.8 times estimate, distributed with a long tail rather than normally. A ten percent contingency does not reflect that shape, and the projects that fail financially are usually the ones that treated the reserve as a formality.
At least twelve months. Google's guidance is to keep redirects in place for as long as possible, generally at least a year, because that timeframe allows all signals to transfer to the new URLs, including recrawling and reassigning links on other sites that still point at your old ones.
This is worth stating because most migration checklists treat redirects as a launch-week task that can be cleaned up once traffic looks stable. They are a twelve-month commitment. Removing them early forfeits exactly the link equity the project spent months protecting.
Very little, in common usage. Migration tends to describe moving data, content and URLs from one system to another. Replatforming tends to describe the broader programme: the new platform, the integrations, the team workflow and the operating model.
Rehosting is genuinely different and worth separating out. Moving the same platform to new infrastructure is a hosting change. It carries none of the data mapping or redirect risk that makes a replatform hard, so do not let a vendor price one as the other.
In most cases, yes. The pattern is to build and verify on the new platform while the current store keeps trading, lower your DNS TTL around 48 hours ahead, run a final delta import of orders and customers created since the main load, then switch.
What you cannot fully avoid is a short window where orders could land in two places. Plan for it explicitly rather than hoping: either pause order capture for a defined period, or reconcile the gap afterwards with a documented process.
The standard set: products and variants, categories and taxonomy, prices, inventory, customer accounts, order history, content pages, reviews, discount codes, tax classes, shipping rules and SEO metadata.
The set that causes production incidents when missed is different. Gift card and store credit balances are a live financial liability and need reconciling to the penny with Finance. Stored payment tokens are vendor-scoped, so confirm whether your processor supports vault-to-vault transfer or subscriptions will break at renewal. Live B2B state means open quotes, credit balances, approval chain configuration and contract price effective dates.
Expect temporary movement. Google's own guidance states that you may experience ranking fluctuations while it recrawls and reindexes, and that for medium-sized sites it can take a few weeks or more for new URLs to replace old ones in results. Larger sites take longer.
Permanent loss is an execution problem, not an inevitability: missing or chained redirects, bulk redirects to the homepage, staging environments left indexable, structured data dropped in the rebuild, or facet and pagination URLs nobody mapped.
For an enterprise migration, the cost comes from the scope. Four things move it most: how many systems the platform connects to, how much live commercial state has to carry across, how many storefronts, brands and markets are involved, and how long the old and new platforms run in parallel.
Platform licensing is usually the smaller part of the total. Implementation, integration, data migration and testing make up most of it, so a quote built from the licence fee alone understates the project.
For a mid-market store, plan three to six months end to end. Enterprise B2B builds involving ERP integration, contract pricing and corporate account hierarchies usually run eight to fourteen months. A full composable rebuild with a separate front end can run twelve to twenty-four months.
The build is rarely the long pole. Data reconciliation, redirect mapping, integration testing and stakeholder approvals are what stretch timelines. Teams that compress the schedule almost always compress testing, which is the phase that protects revenue.
AI product retrieval is an AI system finding, understanding and answering questions about your products, pricing, availability, compatibility and fulfilment. It is worth separating from agentic commerce, where an AI opens a website, navigates it and takes actions on a customer's behalf.
Retrieval is the one affecting almost every merchant today, because it happens whether or not you have done anything about it. A shopper asks for a recommendation and the system assembles an answer. Your products are either in it with accurate facts, in it with wrong facts, or absent.
No. A channel-specific catalogue recreates the problem every time a new interface appears, and it guarantees the versions drift apart.
The stronger architecture is to model the commercial truth once, covering identity, attributes, variants, price, availability, relationships and fulfilment, then expose the appropriate version through your website, structured data, merchant feeds and supported AI integrations. Each channel becomes an output rather than a separate source.