Google’s Open Knowledge Format Is a Reminder That AI Agents Need Structured Source Data, Not Better Prompts

What Google actually announced, and why ecommerce teams should care

Google Cloud has introduced the Open Knowledge Format, a way to package knowledge so AI agents can read it directly. The shift is straightforward: the useful unit is moving away from a single page or a clever prompt and toward a structured knowledge package that can be retrieved, compared and reused.

For ecommerce teams, that lands in the middle of everyday work. Product data sits in one system, help articles live in another, policy pages are tucked elsewhere, and internal notes often hide in shared docs that half the team trusts and half the team tolerates. The result is familiar, search surfaces the wrong thing, support gives mixed answers, and merchandising teams spend time cleaning up the same inconsistency again and again.

That is why this announcement matters even if nobody on your team ever touches an AI agent directly. Better prompts do very little if the source material is scattered or stale, especially if it is contradictory. Structured knowledge gives the system something stable to work with, and reliable behaviour starts with that foundation.

The same source mess affects external content and internal operations. Shoppers see it as a confusing product page or a returns answer that changes depending on where they look. Teams feel it as slower search and decisions based on partial facts, with support becoming noisier.

Why prompt quality hits a ceiling when the source data is messy

A prompt can only work with what the model can already access or retrieve. If the underlying facts conflict, the prompt just repackages the same mess in nicer language. You can ask for a clean answer all day, but the system still has to choose between bad inputs.

Ecommerce teams already deal with these failure modes. One product page says a shirt is machine washable, while another says hand wash only. A policy page says returns are accepted within 30 days, a PDF says 45, and an old help article still talks about store credit. A support macro copied six months ago tells agents to quote a shipping rule that no longer exists.

That is where agentic systems become unforgiving. They fetch facts and compare sources, so they need reliable material more than clever wording. McKinsey has estimated that generative AI could add hundreds of billions of dollars in value to retail and consumer packaged goods, which makes source quality a commercial issue.

Take a simple returns example. A shopper reads one FAQ entry that says final sale items are excluded, a PDF that lists exceptions for faulty goods, and a checkout note that mentions hygiene items. If those sources disagree on wording or scope, an agent cannot safely answer the question or trigger the right workflow.

That is the real signal in Google’s announcement. The market is moving towards packaged knowledge, where structure matters as much as wording. For ecommerce teams, that means the quality of the source set matters more than the polish of any single prompt.

What structured source data looks like in an ecommerce business

Structured source data means information stored in fields and linked records rather than prose spread across pages, with clear ownership. A product’s material and care instructions sit in defined attributes, along with variant sizes. A shipping rule points to a zone and a carrier, while exceptions are handled separately.

A support article references the policy it depends on. That setup gives an agent facts it can use without guessing.

Most most ecommerce stores already have pieces of this. Product attributes live in a catalogue, inventory rules sit in an operations system, and size guidance often exists in a spreadsheet or fit note. Support teams also keep macros, escalation rules and warranty wording, which are knowledge assets when they are maintained properly.

A clean product feed is only one part of the picture. Agents also need policy logic and exception handling, plus context around edge cases, such as what happens when a bulky item ships to a remote postcode or when a gift card cannot be combined with a discount code. Without that layer, automation can describe a product but still fail when it matters most.

This is where many stores break down. PDFs hold warranty terms nobody updates. Image text in banners carries sizing notes that search cannot read.

Copy-pasted CMS blocks repeat the same shipping promise across collection pages, and one edit misses three others. The data looks present, but it is fragmented in ways that agents cannot trust.

That is also why structured knowledge is built for reuse across channels and systems from the start. A single source of truth can feed product pages, support replies, search snippets and internal workflows without each team rewriting the same fact in a different tone. Once the structure is in place, the wording can vary safely.

The parts of brand knowledge that agents need most

Google’s Open Knowledge Format is a reminder that AI agents need source material they can sort and trust, then connect into reliable answers. For an ecommerce brand, that starts with breaking knowledge into practical buckets, because a single long page mixes facts, rules and sales copy in a way that helps no one.

The first bucket is product facts. That includes materials, dimensions, ingredients, compatibility, care instructions and variant details, the hard facts a shopper needs before they buy. If a running shoe line uses different foams across men’s and women’s fits, the system needs that distinction stored clearly alongside the sizes and colourways it applies to.

The second bucket is operational rules. Shipping thresholds, return windows, bundle exclusions, preorder conditions and warranty limits all belong here. Agents need the rule itself and the exception that overrides it, because a blanket policy page leaves them guessing when a gift card, final sale item, or customised product changes the outcome.

The third bucket is customer-facing explanation. Installation guides, ingredient notes, material care and troubleshooting steps do real work because they answer the questions people type after they land on a product page or contact support. A mattress brand’s setup guide, for example, can help a shopper decide whether the item fits a narrow staircase before checkout.

The fourth bucket is internal decision logic. That covers which policy applies first, which source wins when two pages disagree, and when a human needs to step in. Agents do far better when that logic is written down than when it lives in someone’s head or in a thread nobody can find on a busy Monday.

Content teams should also think about reuse from the start. The same answer may need to appear in search results, on support pages, in product detail copy, and in merchant-facing documentation, with each surface using the same facts in a different shape. When the source stays consistent, humans get cleaner pages and agents stop remixing old copy into fresh mistakes.

That consistency matters more than stylistic polish. An agent doesn’t care whether a material note sounds elegant, it cares whether the note matches the care instructions, the size chart, and the return policy everywhere it appears. One source feeding many surfaces keeps the whole operation aligned.

A practical audit for scattered pages, PDFs, and hidden knowledge

Most stores already have the knowledge an agent needs, but it is spread across product pages, PDFs, order policy pages, help articles and old campaign decks. The audit checks whether that material functions as a knowledge base or as a disorganized collection of useful material that was never labelled properly.

Start with the first test, which is simple enough to do in an afternoon. Pick ten common ecommerce questions, such as “does this jacket run small”, “how long is delivery to my area”, or “can I return an opened skincare product”, then check whether one authoritative answer exists for each one. If your team has to compare three pages and a Slack memory before answering, the knowledge is scattered.

The second test checks whether the answer can be pulled into fields cleanly. A decent source can separate the policy title, the rule, the exception, the eligible products, and the last review date without a human having to interpret a paragraph of prose. If the only usable version lives inside a beautifully written block of copy, an agent will struggle to extract it without mangling the meaning.

The third test is about exceptions and update rules. Every store has edge cases, a clearance item with a shorter return window, a subscription box with different shipping terms, or a bundle that follows its own refund policy. If those exceptions live only in memory or scattered messages, the system will answer confidently and wrongly the moment a weird order appears.

The fourth test looks at ownership and review. Someone should own each knowledge area, each piece should have a version, and the review cycle should be visible enough that stale pages get caught before they spread. A policy page that gets edited ad hoc by whoever noticed a typo yesterday is a recipe for drift, especially once the same wording appears on a dozen other pages.

You can turn those four tests into a quick audit sheet. For each important question, note where the answer lives, whether it can be parsed into fields, what overrides exist, and who checks it next. That shows whether the knowledge is ready for agents or trapped in disconnected assets.

The useful part of this audit is what it reveals about the business and the content. In the content audits we run for ecommerce clients, the most common gap is a policy that exists only inside a PDF nobody formally owns and that hasn’t been reviewed since the original launch. When the same shipping rule appears in a help article, a checkout FAQ, and a wholesale PDF with different wording, the team has already created three versions of truth. Agents expose that problem faster than search ever did.