Where shopping is going

More and more product searches now start with a question to an AI agent rather than a search box. The agents answer with a short list, compare the options and, increasingly, help the shopper buy.

• ChatGPT. In March 2026 OpenAI refocused shopping in ChatGPT on discovery and side by side comparisons, with retailers like Target, Sephora and Best Buy supplying product data through its Agentic Commerce Protocol.
• Google. In May 2026 Google extended its Universal Commerce Protocol across Search and Gemini, and gave retailers a tool to compare their share of voice on AI surfaces with similar brands.
• Amazon. Alexa for Shopping now sits in the main Amazon search bar and answers with a handful of picks.

Most of these journeys still end on a website. The agent finds and compares, then sends the shopper to the brand or the retailer to buy. So the product page an agent quotes has to agree with the one the shopper lands on.

Agents check your story

An agent builds its view of a product from everything it can read. The marketplace listing, the brand site, retailer pages, product feeds, reviews and forums. When those sources agree, it can recommend the product with confidence. When they don't, it hedges, or picks a rival whose story is simpler to trust.

In our audits the same problems come up again and again.

• Facts that don't match. Different sizes, counts or ingredients on Amazon, the brand site and retailer pages.
• Claims that drift. A claim worded one way on pack, another on the site and a third on a retailer page, sometimes with one version that was never approved.
• Gaps. No answer to who the product is for, how to use it or how often, so the agent fills the gap from reviews and forums.
• Markets out of step. Translated pages that were accurate at launch but missed every update since.

In one recent audit of a brand's 40 best selling products, 62% had at least one fact or claim that differed between Amazon and the brand's own site.

The brand site is often the weakest link, which is odd, because it's the one channel a brand fully controls. Product pages are built at launch, then left alone, while the detail that agents want sits in claims documents, packaging artwork and email threads.

Why optimising copy isn't enough

Many content tools are built to optimise copy that already exists. They are good at it. But rewording a thin product page gives you a better worded thin product page.

What agents need is the missing content. Who the product is for, how it works, how to use it, the answers to the questions shoppers ask and claims the brand can stand behind. Producing that means finding the source material, checking every claim against it and delivering it in the format the website and the feeds need. That's the work most brands are short of.

How we build product content agents can trust

We use a staged process. Each stage is run by its own AI agent with a narrow job, and our team checks the work at the points where judgement matters. The stages overlap, so a finding at one stage can send work back to an earlier one.

1. Collect

A shopping agent gathers every live version of the product. The Amazon listings, retailer pages and brand site in each market, plus the pack copy.

2. Extract

The facts go into one product record. Sizes, counts, ingredients, directions, warnings and every claim, with where each one came from.

3. Source

Every claim is matched to an approved source, such as a signed off claims document, the packaging artwork or a regulatory approval. Brands usually hold far more of this than anyone has time to search, across shared drives and old email threads. No source, no claim.

4. Draft

Each field in the site's product template is written from the record. We draft against the shopper contexts AI systems map products to, such as who a product is for, what it's used for and when, the approach behind Amazon's COSMO research. We add answers to the questions shoppers actually ask.

5. Check

Every claim gets a rating. Green claims match an approved source and are ready to use. Amber claims are close, such as a reworded approved claim, and are logged for the brand to see. Red claims have no source and go to the brand's approvers. Nothing red goes live.

6. Transcreate

Our native speaking creatives adapt the content for each market, with local search terms and local claims rules, rather than translating it word for word.

7. Publish

The content is delivered in the format the site and feeds take, from content management fields and HTML to feed attributes, so developers can import it without retyping.

8. Keep it current

When a pack, claim or price changes, the record changes and every channel follows. We track how the products show up on AI surfaces, so we can see where the story still isn't landing.

One set of answers, used everywhere

The questions shoppers ask don't change from channel to channel, so the answers shouldn't either. We pick the most asked questions for each product once, by search and review volume, and use the same answers in the brand site's FAQs, the Q&A modules in Amazon's Premium A+ content and the product feeds. Agents see the same answer wherever they look, and the brand only has to approve it once.

Feliway and Douxo

Ceva's animal health brands show why one record matters. Feliway and Douxo are sold in dozens of countries, through vets, pet retailers and Amazon, and every claim has to stay within what's approved in each market.

For both brands we build one master for each product, with every claim taken from Ceva's labels, claims documents and study references and signed off by its regulatory team. That master is then localised for each market, so a Feliway diffuser says the same thing in Warsaw as it does in Madrid.

Keeping it consistent is ongoing work. When Amazon introduced a new title format and Product Highlights in 2026, we moved around 250 Ceva listings across to it, every one kept to the approved claims. Amazon can treat an upload as a suggestion, so we check what actually went live and raise a case wherever it differs. A product story is only consistent once you've checked it.

Across 64 Feliway products in 25 countries, ordered revenue is up 48%. Read the Feliway case study.

We built the staged process itself for a global health brand's websites, where it filled the product pages for 16 markets in the format the developers import, with every claim checked before it went live.

What brands should do now

1. Audit what the agents can see

Take your top products and put the Amazon listing, the brand site page and your biggest retailer's page side by side. Note every fact and claim that differs.

2. Find your approved sources

Gather the claims documents, packaging artwork and approvals in one place. They are the foundation for everything else.

3. Fill the gaps, not just the words

Give every product page the sections agents look for. Who it's for, how it works, how to use it and the answers to common questions.

4. Make the brand site the reference

It's the channel you control. Keep it the most complete and current version of every product, and make sure your feeds carry the same content.

5. Keep one record up to date

When anything changes, update the record first and let every channel and market follow from it.

The brands agents will recommend

Agents reward the brands that are easiest to trust. That has less to do with clever copy and more to do with consistent, complete, approved product information, wherever an agent looks and in whatever language it's asked.

If you want to know how your products look to an agent, our team can audit them across Amazon, your site and your retailers, and Freak Lab can rebuild the content from your approved sources, in every market you sell in.

Sources

1. OpenAI reveals updates to its agentic commerce experience for ChatGPT, Digital Commerce 360
2. Agentic Commerce Protocol, key concepts, OpenAI
3. New Universal Commerce Protocol features and AI tools on Google, Google
4. Building commonsense knowledge graphs to aid product recommendation, Amazon Science