- Adobe puts AI referred retail traffic at 42% better converting than other channels in March 2026, after the same traffic converted 38% worse a year earlier.
- The same Adobe analysis scores product pages at 66% for machine readability, the lowest of any page type on a retail site. Returns pages score 82%.
- One furniture retailer told Digital Commerce 360 that AI sources already account for about 5% of its inbound referrals, and expects 10% by 2027.
- There are two separate doors into an assistant: markup on your own pages, and a catalogue file you push. They want different fields.
- Holiday planning is happening in August this year, which makes the catalogue work a now job rather than a November job.
Open your analytics and filter for referrals from ChatGPT, Perplexity, Copilot and Google AI Mode. For most small shops the number is still small enough to ignore. It is also the only referral line that behaves differently from every other one, and the difference has flipped direction in the space of twelve months.
Adobe published its retail analysis of that traffic this summer, drawn from over a trillion visits to United States retail sites plus a survey of more than 5,000 people. In March 2025, visitors arriving from an AI assistant converted 38% worse than visitors from everything else. In March 2026 the same cohort converted 42% better, which Adobe recorded as a new high. They also stayed 48% longer and viewed 13% more pages per visit.
Worth noting what the 42% figure is and is not. It is a comparison between two groups of visitors on the same sites, not a promise that adding markup raises your conversion rate by 42%. People who arrive from an assistant have usually already described what they want and been told you match it, so they are further along than someone clicking a broad search result. The mechanism is selection, not magic. What follows from that is still useful: the traffic is pre qualified, so being absent from the answer costs more than being ranked fifth on a search page.
That is the part of the story a merchant would repeat to another merchant. The part worth acting on is buried further down the same report, and it is unflattering.
Which page on a shop is hardest for an assistant to read?
The product page. Adobe scored retail pages for what it calls AI content visibility, and product pages came last at 66%, below homepages at 75%, below customer service at 79%, below FAQ pages at 80%, and below returns and exchanges pages at 82%.
Sit with that ordering for a second. The page that explains how to send something back is easier for a machine to understand than the page that takes the money. The gap is not small either: Adobe put the top retailers at an average of 82.5% against 54.2% for the weakest.
There is a reasonable explanation. A returns page is mostly prose. Somebody wrote sentences about windows and conditions and refunds, and prose is exactly what a language model reads well. A product page is the opposite: the important facts live in a price widget, a stock badge that renders after the JavaScript runs, a variant selector, a star rating pulled from a third party script, and a delivery estimate that appears only once you enter a postcode. A person reads that layout instantly. A crawler reads a page that says almost nothing.
| Page type | Adobe AI content visibility score | Why it scores that way | The cheap fix |
|---|---|---|---|
| Returns and exchanges | 82% | Written as plain prose, few dynamic widgets | Nothing. This one already works |
| FAQ | 80% | Question and answer pairs in the HTML | Add the questions customers actually email you |
| Customer service | 79% | Policy text, contact details in text form | Put hours and response times in words, not icons |
| Homepage | 75% | Heavy on imagery, light on descriptive text | One paragraph saying what you sell and to whom |
| Category | 74% | Grid of tiles, little descriptive copy | A short intro paragraph above the grid |
| Product | 66% | Price, stock and variants rendered by script | Product markup in the initial HTML |
The right hand column is mine, not Adobe's. Adobe published the scores and the recommendation to improve machine readability; it did not say which page deserves your Saturday. The ranking says the product page does, because that is where the score is lowest and where the money is.
What does an assistant actually need to see?
Three things, in this order: what the item is, what it costs, and whether you can ship it. Everything else is refinement. The trouble is that there are two entirely separate ways to hand those facts over, they belong to different companies, and doing one does nothing for the other.
The first door is markup on your own pages. Google's ecommerce documentation lists the types that matter for a shop: Product and ProductGroup for the items themselves, BreadcrumbList for hierarchy, Organization for business details including return policy, Review for ratings, and LocalBusiness if you have a physical address. The instruction that most shops miss is about placement rather than content. Google says to put Product structured data in the initial HTML for best results across shopping surfaces. If your markup is injected by a script after load, some readers will see it and some will not.
The second door is a catalogue feed, and it is a different shape entirely. OpenAI's product feed specification asks for a file you deliver rather than a page it crawls. The required fields for a standard feed include item_id, title, description and url, plus brand, image_url, price with an ISO 4217 currency code, and an availability value drawn from a fixed list: in_stock, out_of_stock, pre_order, backorder or unknown. It also wants seller_name and seller_url, and geo fields for target_countries and store_country as ISO 3166-1 alpha-2 codes. Accepted formats are tab delimited or comma delimited text, optionally gzipped.
Read that field list as a merchant rather than a developer and something jumps out. Most small shops do not publish a machine readable brand anywhere. Many do not distinguish out of stock from backorder from pre order in any consistent way, because a human reading "back in two weeks" understands it fine. An enumerated availability value is not a nicety in that world, it is the difference between being listed and being skipped.
Does any of this replace ordinary search work?
No, and treating it as a replacement is the expensive mistake. AI referrals are still a minority channel for almost everybody. The Digital Commerce 360 piece on holiday preparation quotes Ayden Lin at the furniture retailer Povison saying AI associated traffic makes up about 5% of inbound referrals, with an expectation of 10% by 2027. Five percent is worth an afternoon. It is not worth rebuilding your site around.
What makes the work defensible is that almost none of it is specific to AI. Product markup in the initial HTML helps Google's shopping surfaces, which is where the other 95% comes from. A clean availability state helps your own customers. A description written in sentences rather than bullet fragments helps the person reading it on a phone. We went through the same reasoning when we wrote about what actually moves citations in AI answers, and the conclusion held: the techniques that survive are the ones that were good practice anyway.
How do you measure AI referrals in your own analytics?
By referrer hostname, and the answer will be lower than reality. Assistants send visitors with a referrer such as chatgpt.com, perplexity.ai or copilot.microsoft.com, so a segment filtering on those hosts gives you a first number in about five minutes. Google AI Mode is the awkward one, because traffic from it can arrive looking like ordinary Google traffic.
The undercount matters more than the count. A shopper who asks an assistant for a recommendation, reads the answer, and then types your brand name into a browser shows up as direct traffic. Nothing in your analytics connects that session to the conversation that caused it. The same blind spot existed with word of mouth for twenty years and nobody solved it then either. What you can do is watch the trend on the referrals you can see, and treat it as a floor rather than a measurement.
Before you conclude the channel is worthless because it is 2% of sessions, check the conversion rate on that segment separately. That is the whole point of the Adobe finding. A channel that is small and converts at twice your site average deserves different treatment from a channel that is small and converts at half. Adobe also reported that consumer trust moved underneath these numbers, with 66% of survey respondents saying they believe AI tools give accurate results, which is the sort of shift that shows up in behaviour before it shows up in traffic share.
What else did the holiday research turn up?
One finding that has nothing to do with AI and is probably worth more to a shop with a physical location. Digital Commerce 360 reported that retail chains displaying in store stock availability converted at 3.1% against a 2.9% average across the Top 1000 retailers in 2025. Two tenths of a point sounds trivial until you apply it to a year of sessions.
It belongs in this article because it is the same underlying idea in a different costume. Both findings say that publishing a fact you already know, in a place a reader can find it, converts better than leaving that fact implicit. The stock level exists in your system either way. Whether it appears on the page, in words, is a choice. Shops that make that choice do better with humans and with machines, for the same reason.
Why is everyone doing this in August?
Because the holiday calendar moved. Digital Commerce 360's reporting on the 2026 season describes retailers launching campaigns well before the Cyber 5 window, with Garrett McMartin of Draco Diamond framing early promotion planning around delivering orders on time rather than around discounting harder. A made to order jewellery business has an obvious reason to start early. The interesting thing is that the reasoning now applies to catalogue data too.
The advertising side tells the same story from a different angle. Modern Retail reported on a Tatari survey in which 58% of advertisers said they are focusing on their own direct site for the holidays, with 39% taking an omnichannel approach that still puts digital first, and not one naming in store as the priority. Everybody is pointing at the same destination, which means the traffic arriving there is more contested than last year.
If your catalogue is going to be read by an assistant during the peak weeks, the reading happens on whatever data exists at that moment. Crawlers revisit on their own schedule and feeds have approval steps. Fixing a product template on 20 November is a fix that lands in December.
The two doors have different entry costs. Markup you can ship yourself today. A feed to a specific assistant generally requires applying as a merchant and being approved first, so the lead time is outside your control. Start with the markup.
What should a one person shop do first?
Work down the list in this order, and stop when you run out of Saturday. Each step is useful on its own, which matters when you cannot finish.
One, check what a machine sees. Fetch one of your product pages with JavaScript disabled, or view the raw source. If the price is absent from that HTML, that is your finding. You do not need a tool for this and it takes two minutes.
Two, get Product markup into the initial HTML. If you are on a platform with a structured data extension, this is a settings change. If you own your storefront code, it is a block of JSON-LD rendered on the server. Google's guidance is explicit about the initial HTML part, and it is the part that platform plugins most often get wrong by injecting late.
Three, write availability as an enumerated state. Pick one vocabulary and use it everywhere: in stock, out of stock, pre order, backorder. Prose like "ships soon" is unreadable to a feed validator and ambiguous to a customer.
Four, write one real paragraph per product. Not a spec dump. A sentence about what it is, who it suits, and what it does not do. This is the text an assistant quotes when it recommends you, and it is also the text that reduces returns. If you are generating that copy with a model, the practical constraints are the same ones we covered in producing product imagery and copy without ending up with the same asset as everyone else.
Five, only then look at feeds. Apply where it makes sense for your category, and treat the required field list as a catalogue audit rather than a form to fill.
What this does not fix
Being readable is a precondition, not a promise. An assistant that can parse your page will still recommend a competitor whose price is lower, whose delivery is faster, or whose reviews are better. Machine readability gets you considered. It does not get you chosen.
It also does nothing about the underlying volatility. Assistants change how they source products, and a channel that grew several hundred percent in a year can reshape itself in a quarter. Anyone building a business on one referral source is taking the same risk merchants took with a single social platform, and it ended the same way each time. Our earlier look at what shopping agents do when they reach a storefront goes through the mechanics of that dependency in more detail.
The honest framing is narrow. A measurable minority of your traffic now arrives from assistants, it converts better than the rest, and the page it lands on is the one those assistants read worst. The work to fix that is a template change and a vocabulary decision, both of which pay off in ordinary search too. That is a good ratio, and it is available in August rather than in the week before Black Friday.
Is a feed worth it for a very small catalogue?
Usually not as a first move. The approval process and the delivery mechanics cost more than the markup route, and the markup route serves every reader rather than one. A shop with thirty products and no engineer gets more from clean server rendered Product data than from a catalogue file that needs a scheduled upload. The calculation changes if you carry hundreds of items with fast moving stock, because then the feed is the only way to keep an assistant's copy of your prices current.
One more thing worth saying plainly, because it decides whether any of this is even possible for you: all of it assumes you can change your product template. On a rented platform that means waiting for a plugin or a theme update. If you own the storefront code, it is an afternoon. That difference is the whole argument for building a shop whose code you actually control, and it shows up on exactly these occasions, when a format changes and everyone else is waiting in a queue.