There is a metric quietly deciding your revenue that never shows up on a dashboard, and it is the completeness of your product data. For years the numbers a store obsessed over lived at the bottom of the funnel, conversion rate, average order value, cart abandonment, the moments where a visitor already on your site either bought or did not. Those numbers still matter, but a new and equally consequential metric now sits at the very top, before a shopper ever reaches you, and it is invisible precisely because failure at this stage produces no visitor to measure. When an assistant cannot read your product data properly, you are not converting poorly. You are not being considered at all.

To understand why feed hygiene became this important, you have to see the catalog the way an AI agent sees it, which is nothing like how a person sees it. A shopper looks at a product page and forgives its gaps. If the material is not listed, they assume cotton from the photo. If the fit is unclear, they read a review. If a field is blank, their brain fills it in from context, because humans are relentless inference machines who tolerate ambiguity. An agent assembling a recommendation does none of this. It works from the structured facts you provided, and where a fact is missing, the product simply fails to match the query. There is no assumption, no benefit of the doubt, no filling in. An empty attribute is not a minor imperfection. It is a query you have chosen not to answer.

This changes the nature of the fields you used to treat as optional. In the keyword era you could leave pattern blank, skip occasion, ignore the compatibility list, and nothing bad happened immediately, because those fields were nice-to-have metadata that a human would never miss. In the agent era those same blanks are fatal in a specific and literal sense. If a shopper asks for a formal striped shirt and your striped formal shirt has neither pattern nor occasion filled in, the agent cannot see that your product matches, and it recommends a competitor whose data was complete. The industry has taken to calling these empty optional fields ghost attributes, because the product might as well be a ghost for that query. It is there, it is exactly what the shopper wanted, and it is invisible because the entity was never defined.

The governing principle is that explicit beats implicit, every time, without exception. Do not leave the assistant to infer that your jacket is warm because it looks like a winter coat. State the insulation. Do not hope it deduces the style from the photo. Hard-code that the fit is formal. Every fact you make explicit is a hook the assistant can grab to pull you into an answer, and every fact you leave implicit is a hook that does not exist. This is a genuine inversion of how good marketing copy used to work. The old craft was to evoke, to suggest, to let the customer imagine. The new requirement, for the data layer at least, is to spell everything out in plain structured terms, because the reader is a machine matching facts, not a person being seduced.

Titles deserve particular attention because the old habits actively hurt you now. The keyword-stuffing instinct, cramming a title with every term you hoped to rank for, made a kind of sense when you were feeding a search engine that rewarded keyword presence. An assistant reads a title to understand what the product is, and a stuffed title is noise that obscures the entity rather than clarifying it. A clear, accurate, descriptive title that says exactly what the product is beats a keyword salad, because the assistant is trying to comprehend, not to index. The same logic applies to descriptions. The ones that perform contain the specific facts a query might hinge on, the exact material, the true dimensions, the intended use, the care instructions, rather than atmospheric prose about lifestyle and aspiration.

Then there is the requirement that trips up more stores than any single missing field, which is consistency across surfaces. Your data does not live in one place. It lives on your product page, in your merchant feed, and often on third-party sites and marketplaces, and an assistant may check these against each other. When they disagree, when your site says one price and your feed says another, when your return policy differs between your page and your feed, when a product is in stock on your site and out of stock in the data an agent reads, the agent does not know which version to trust, and the safe move for the agent is to distrust you and recommend something else. A human shopper would shrug at a small mismatch. An agent treats it as a reliability problem, and reliability is the thing it optimizes for when it is about to stake a recommendation on your data.

Accuracy over time is a related discipline that static thinking underrates. A feed you generated cleanly on Monday can be wrong by Wednesday if your inventory moved and your data did not. Price changes, stock changes, and variant availability changes are exactly the facts an assistant most needs to be current, because recommending an out-of-stock or mispriced item is the fastest way for the assistant to erode its own trust with the shopper, so it learns to avoid sources whose data goes stale. This is part of why the newer protocols added the ability for agents to pull live product details on demand rather than relying only on a periodic push, a shift covered elsewhere in this series. Whether your data is pushed or pulled, its value is a direct function of how closely it tracks reality at the moment of the query.

It is worth being blunt about the return on this work, because it is unusually high and unusually cheap. Feed hygiene is not a redesign, a rebrand, or a large engineering project. It is the deliberate work of filling in fields, correcting inconsistencies, and keeping data current, and yet it sits upstream of an entire emerging channel of high-intent demand. A store that does this well becomes eligible to appear in answers it was previously invisible in, and because the assistants weigh data quality rather than ad spend, the payoff accrues on the strength of information you already possess about your own products. Few investments in retail offer this combination of low cost and structural upside, and even fewer are as widely neglected, which is exactly why acting on it early is an advantage rather than table stakes.

There is also a compounding dynamic that rewards the diligent. When an assistant repeatedly finds your data complete, accurate, and current, it has more reason to surface you, those recommendations generate engagement, and the engagement reinforces the surfacing. Conversely, a store whose data is thin or unreliable teaches the assistant to skip it, and that lesson is sticky. The gap between a clean catalog and a messy one is therefore not static. It widens, because good data earns visibility that generates signals that earn more visibility, while bad data earns an absence that produces no signals at all. The store that treats feed hygiene as ongoing operational discipline pulls steadily ahead of the store that treats it as a one-time chore.

The practical program is not complicated, which is part of what makes neglecting it inexcusable. Audit your catalog as a table of facts and find every empty attribute, because each one is a query you are forfeiting. Fill the ghost fields, the pattern, the occasion, the material, the compatibility, the care instructions, because an undefined entity is an invisible one. Rewrite titles to describe rather than to stuff. Load descriptions with specific facts rather than mood. Reconcile every point of difference between your site, your feed, and your marketplace listings, because inconsistency reads as unreliability. And put a process in place to keep price, stock, and variant data current, because accuracy decays and stale data is worse than useless. Do this, and you have moved the metric that now matters most, the one that decides whether an assistant considers you at all. In a market where being unreadable means being uncounted, feed hygiene is not back-office maintenance. It is the new conversion rate, and it is being decided before a shopper ever arrives.

Because the work is finite, the smart move is to sequence it by impact rather than trying to perfect everything at once. Start with your bestsellers and highest-margin products, because these are the items where an appearance in an assistant's answer is worth the most, and where completeness therefore returns the most. Fix their attributes first, reconcile their data across every surface, and make their descriptions dense with the specific facts a query might hinge on. Then work outward to the long tail. Within each product, prioritize the attributes most likely to appear in a shopper's phrasing, the material, the size, the use case, the compatibility, before the truly marginal fields. This triage turns an intimidating catalog-wide chore into a focused sequence that produces visible gains early. You do not need every product perfect on day one. You need your most valuable products legible first, and the rest brought up steadily behind them, so the return begins immediately rather than waiting for a total cleanup.