Short answer. Brands get recommended in AI shopping answers when three things line up: assistants can read the product data, the merchant supplies a structured and frequently refreshed feed of price, availability and fulfilment facts, and third-party reviews, forums and comparison pages agree with what the brand says about itself. The shortlist now forms before any merchant site is opened.
Key takeaways
- AI assistants now form the product shortlist inside the chat, before a shopper reaches any merchant site, whether or not the purchase completes there.
- Adobe recorded AI-sourced US retail traffic up 393% year over year in Q1 2026, with AI-referred conversion swinging from 38% worse than other traffic in March 2025 to 42% better in March 2026.
- Retail product pages averaged a machine-readability score of 66 out of 100 in Adobe's April 2026 benchmark, the lowest of any page type measured.
- A product feed makes an item eligible for AI shopping answers; third-party evidence, explicit attributes and consistent facts decide whether it is actually recommended.
What has actually changed about product discovery?
Product discovery used to end on a results page; it now often ends inside an AI assistant's answer. A shopper describes a budget, a use case and a constraint in a chat window and receives a short list of named products with reasons attached.
Agentic commerce is the use of an AI agent that acts on a shopper's behalf, up to and including completing a purchase. For twenty years a shopper typed a query, scanned ten blue links, clicked two or three and compared, and the merchant's job was to rank and then convert. The comparison that used to happen across browser tabs now happens inside the assistant.
The infrastructure caught up quickly. In late 2025 OpenAI launched Instant Checkout alongside the Agentic Commerce Protocol, an open standard built with Stripe that lets agents, shoppers and merchants complete a purchase without the agent becoming the merchant of record. In March 2026 OpenAI narrowed the in-chat purchase flow and redirected effort towards product discovery, letting merchants keep their own checkout, a change payment providers read as structural rather than a retreat: the assistant handles discovery and intent, the merchant keeps the transaction.
Whether or not the buy button lives in the chat, the shortlist now forms there. A brand that is not legible at that moment is not in the consideration set, and no on-site conversion work reaches a shopper who never arrives. How assistants choose among candidates is covered in how ChatGPT picks sources.
How large is the AI shopping channel right now?
Large enough to plan around, and growing on a curve that has already reversed one of its early assumptions. AI-referred shoppers now convert better than other visitors, even though the channel is still a small share of visits.
Adobe Analytics, which bases its retail figures on more than a trillion visits to US retail sites, reported that AI-sourced traffic to US retail sites grew 393% year over year in the first quarter of 2026, following a 693% jump during the 2025 holiday window. By July 2026 growth had moderated to 62% year over year, a normal shape for a channel maturing off a small base.
Quality is the more interesting number. In March 2025 AI-referred visitors converted 38% worse than non-AI traffic. In March 2026 they converted 42% better, a swing of roughly eighty points in twelve months, and they spent 48% longer on site and viewed 13% more pages. By July 2026 Adobe put AI-referred revenue per visit 53% above other traffic, with conversion 60% higher, the eleventh consecutive month of outperformance.
| Period | AI vs non-AI conversion | Revenue per visit |
|---|---|---|
| March 2025 | 38% lower | Below non-AI |
| March 2026 | 42% higher | 37% higher |
| May 2026 | 54% higher | 53% higher |
| July 2026 | 60% higher | 53% higher |
Two cautions apply. The figures come from one analytics vendor with a specific customer base of large US retailers, so they describe that population rather than the global web, and Adobe's reported basis shifts between months. High conversion on a small channel is also partly a selection effect: the shopper arrives pre-qualified because the comparison already happened, which is why the shortlist stage matters more than the landing page. A merchant that treats AI traffic only as a conversion-rate bonus misses the point: the same shoppers form their view of the category in the assistant before any click happens.
Why are so many product pages invisible to AI models?
Because the page a human sees and the page a crawler receives are frequently different documents. A product page built from client-side components can reach a retrieval system as empty containers.
A machine-readable page is one whose key content, such as price, specifications and availability, is present as text in the HTML a crawler receives. Adobe's AI Content Visibility Checker scores how much of a page a large language model can read, out of 100. Across US retail in April 2026, homepages averaged 75 and product pages averaged 66, so roughly a third of the content on the page closest to purchase is not machine-readable. Homepage scores alone ran from 82.5 to 54.2 between best and worst performers.
| Page type | Score out of 100 |
|---|---|
| Product pages | 66 |
| Store locator | 73 |
| Category pages | 74 |
| Homepages | 75 |
| Loyalty or membership | 78 |
| Customer service or help | 79 |
| FAQ | 80 |
| Contact us | 81 |
| Returns or exchanges | 82 |
The pages scoring best are the plainest: returns policies, contact details and FAQs, which are static text in ordinary HTML. The pages scoring worst are the most heavily engineered, with specifications inside tabs, prices injected by script and key attributes rendered as images. This is among the most common causes of a brand being absent from an AI shopping answer, and it has nothing to do with content quality. A page can be well written, well photographed and well reviewed, yet still contribute nothing to an answer because the words never reach the model. The fixes are covered in the technical AEO checklist.
What does an AI agent actually need from a merchant?
An agent needs structured, current, complete product facts it can read without interpretation. OpenAI's Agentic Commerce Protocol documentation is unusually explicit about this, and works as a specification even for brands that never join the programme.
A product feed is a regularly refreshed file, in CSV or JSON, listing each product's identifiers, descriptions, pricing, inventory, media and fulfilment options. OpenAI's guidance asks merchants for a secure, regularly refreshed feed, with required fields ensuring price and availability display correctly and recommended attributes such as rich media, reviews and performance signals improving relevance and ranking. Integration runs through a validated sample feed followed by daily snapshots. Read as an AEO brief, that says four things:
- Facts must be atomic. An agent needs the price as a value with a currency, not as a phrase inside a paragraph, and availability as a state, not a badge graphic.
- Facts must be current. A daily snapshot cadence implies stale data is worse than absent data, because a recommendation that leads to an out-of-stock product damages the assistant's output.
- Coverage beats polish. Required fields exist so answers do not break; recommended fields exist so answers improve. Most catalogues fail on completeness across thousands of SKUs long before they fail on copy quality.
- Post-purchase behaviour is observable. With a third party in the order loop, fulfilment delays and refund friction become visible outside the merchant's systems. OpenAI names quality as a ranking consideration without fully defining it, so operational reliability may feed that signal over time.
Structured product data is one case of the wider discipline in structured data and entity identity for AEO.
Which signals decide the recommendation beyond the feed?
A feed makes a product eligible but does not make it recommended. The assistant still chooses between eligible options using the wider record: independent evidence, explicit attributes, consistency and category structure.
- Independent evidence. Assistants lean on third-party sources when judging whether a product is good, because a merchant describing its own product is a weak signal. Review platforms, editorial testing, forum threads and comparison pages carry disproportionate weight in category questions, as explained in why third-party brand mentions matter.
- Attribute matching. Most shopping prompts contain constraints such as budget, size, compatibility or use case. Products that state these explicitly get matched; products that imply them get skipped. "Suitable for small kitchens" is retrievable, while a photograph of a small kitchen is not.
- Consistency. When a feed, a product page and a retailer listing disagree on price or specification, the assistant averages, hedges or picks another product. Inconsistency reads as unreliability, so one source of truth for price and specification, published everywhere, is worth more than another round of copy polish.
- Category structure. Third-party comparison and "best for" content tends to be the scaffolding on which recommendation answers are built, so brands in credible category round-ups appear in assistant answers more often than their own site alone would justify.
How does this connect to Lifewood's work?
Most of the fixes above are data problems dressed as marketing problems. Lifewood provides AEO and GEO services, so it has a commercial interest in this subject.
The work includes catalogue attribution, structured product data, multilingual product copy that says the same thing in every market, and the ongoing human checking that keeps a feed truthful across thousands of SKUs. That overlaps with the AI data work Lifewood's delivery teams already run, which is why AEO and GEO sit alongside annotation and multilingual collection in its service lines. The AEO service and GEO service pages describe the offer.
Other organisations approach the same problem from different directions, and the right choice depends on the bottleneck. Adobe has built the visibility diagnostics and LLM optimisation tooling referenced here. OpenAI and Stripe maintain the protocol layer, and payment providers such as Worldpay and Checkout.com have published merchant-side integration paths. A brand whose gap is catalogue quality needs a different partner from one whose gap is payment plumbing; best AEO and GEO agencies compared lays out how providers differ.
What should you do if your category is being answered without you?
Start by testing what assistants say today, then fix the worst-performing template and the third-party record. The first diagnostic costs nothing: fetch your own product page as a bot and read what comes back.
- Ask the question your buyer asks. Run ten realistic shopping prompts across ChatGPT, Gemini, Perplexity and Copilot. Record which products are named, which sources are cited and whether your prices are right.
- Fetch your own product page. Request the raw HTML with JavaScript disabled, using a command-line tool such as curl. If price, specifications and availability are not in that response, no crawler sees them either, and the page needs server-rendered or static text for those fields.
- Audit by page type, not sitewide. Adobe's benchmark shows product pages lagging every other template, so fix the worst template first rather than averaging the problem away.
- Complete the boring fields. Dimensions, materials, compatibility, warranty and returns window are the attributes prompts filter on.
- Date inventory and price. Freshness is a retrieval signal, not a courtesy.
- Fix the third-party record. Correct wrong specifications on retailer listings and review platforms, which are cited more often than your own site.
- Re-measure three times. Answer engines are volatile run to run, so a single check will mislead, and the same prompt can name different products on different days; see measuring share of answer.