Product pages have a new job: answering the buying question

Imagine a customer who needs a desk lamp before Friday. It has to fit a particular mount and work with the dimmer they already own. The customer might start by asking an AI assistant. A well-written product description helps only if it can also resolve those specific buying questions.

This is a hypothetical example, but the technology behind it became more concrete in week 36. Our thesis is that reliable product facts become part of the sales conversation when AI helps customers choose. For some businesses, the next improvement to inventory information and terms may therefore be worth more than another batch of AI-written advertising copy.

This week's English-language podcast draws on nine Hammer reports produced during August 31–September 6, 2026. It covers seven AI providers. Here we explore one commercial question from that research: what happens to selling when customers want to describe their problem rather than compare every product themselves?

Anthropic makes the buying conversation a product you can build

On September 2, Anthropic published reference implementations for commerce agents in retail, travel, telecom, and ticketing. A commerce agent is an AI assistant that uses a company's systems to help with buying or selling.

The shopping agent can search a catalog, compare alternatives, and assemble a cart for a need involving several products. It can also answer order and return questions in the same conversation. That is more specific than a chatbot that has read the website's frequently asked questions.

A separate merchant agent helps with sales analysis, inventory, and campaign proposals. The reference examples do not charge cards or place orders. The store's own system handles payment, and changes proposed by the merchant agent need approval before going live.

This is primarily a way to build an assistant inside a store or app. It does not prove that customers already let an independent AI buy freely across every retailer.

Product facts can determine whether an option can be recommended

In the lamp example, describing the product as flexible and perfect for a home office is not enough. The buyer needs to know which mount it fits, whether the dimmer works, and whether delivery will arrive in time. When a fact is missing, the assistant has to ask another question or leave the choice unresolved.

Anthropic's engineering guide therefore puts tools on top of the store's existing systems. Catalogs, ranking, inventory, and business rules should come from those systems. The model should interpret the results rather than invent its own version of the store's terms.

If our thesis holds, product information gets a clearer role in selling. Work that once looked like catalog maintenance may determine whether a customer gets a useful answer. The same applies to services: what is included, what needs to be in place, and when can delivery begin?

Images, brand, and good writing still matter. They can attract interest. The buyer still needs to know whether the offer fits.

Perplexity and Google show the value of specific information

This week's Perplexity research describes Computer gaining access to Chronograph's private-capital data and OpenSea's market data. Google's reports cover licensed books in Gemini Notebook and the new WeatherNext 3 forecasting service.

These are different markets, not evidence of identical buying behavior. The useful comparison is that providers are building services around information a general language model cannot replace with a plausible answer. Book ownership, a current portfolio record, and a weather forecast require different sources.

For a merchant, the equivalent is less spectacular: current stock, the correct spare part, and actual delivery terms. Investing in those facts makes sense even if customers continue shopping without AI.

Models are improving, but the whole week was not about commerce

The podcast's broader review also covers the following:

• OpenAI launched GPT-6 Astra with support for longer tasks, concurrent tool work, and steering during a response. Those features may help with complex comparisons, but they do not prove better sales.
• Google introduced Gemini 3.8 Flash. Lyria 3.5 music generation arrived in public preview. Anthropic lowered the price of reading reused context with Fable 5.1. The cost of a complete task still depends on how the service operates.
• Grok's X connector adds ways to search posts and track mentions. Mistral's HUMAIN collaboration concerns a planned regional initiative, not a verified new self-service region.
• Manus documented further recovery details. For anyone selling through an AI-built service, being able to recover customer history and the website matters.

The reports were produced during the completed week. Some Monday reports fill earlier research gaps with August events; we are not relabeling them as September launches.

The sales promise still needs testing

Anthropic reports larger carts and a greater likelihood of purchase among retailers using Claude agents. The announcement does not disclose enough about the sample or comparison groups to establish that your store would see the same effect. A working reference implementation is not a finished integration with your catalog, either.

A useful first step is to take a real, recurring buying question your support team already receives and see whether your product information supports an answer. If it does not, you have found a specific sales problem, whether the eventual solution is better web content or a commerce agent. That is a natural starting point for Tool Forge at Hammer Automation.

This episode is an AI-generated masterclass built from Hammer's deep, ongoing research into AI providers' updates and features, processed with NotebookLM. The article uses the same research and a NotebookLM synthesis, not a transcript. Product details and availability may change.