- Epoch AI and Ipsos put the share of US adults using AI on at least six of the previous seven days at 19 percent in August 2026, against 8 percent in March.
- Epoch states plainly that the two waves asked the question in different shapes, so the jump is directional rather than exact. Most coverage drops that line.
- Visa's September 2026 economic insight puts the online share of US domestic payment volume at 58 percent, up from 48 percent in 2019.
- An assistant skips a product that fits when the attribute answering the shopper's constraint was never published. The product is not rejected, it is invisible.
- Google's product highlight attribute accepts 2 to 100 entries of 1 to 150 characters, recommends 4 to 6, and its six conversational attributes are all optional.
- Test with the words a customer would use and never your own brand name, or you only prove that an assistant can read your label back to you.
Two numbers landed within a week of each other, from people who were not talking to one another. Epoch AI, working with Ipsos, reported that the share of US adults using AI on at least six days out of seven had gone from 8 percent in March 2026 to 19 percent in August. Practical Ecommerce, writing for merchants, published a piece telling sellers to test whether AI systems can find their products at all. Neither mentions the other. Read together they describe one problem: the habit arrived faster than the product data did.
If you sell things, the useful question is not whether AI adoption is rising. It is whether the thing you sell can be matched to a request somebody types in a hurry. Those are different problems, and only the second one is yours to fix.
What did the survey actually measure?
It measured self reported frequency of use, not shopping. Epoch AI's data insight on near daily use reports that the share of US adults who used AI at least six days in the previous week rose from 8 percent to 19 percent between March and August 2026, while the share who used it on exactly one day fell from 17 percent to 10 percent. The March wave polled 2,017 adults, fielded 3 to 5 March. The August wave polled 1,016 adults, fielded 28 to 30 August.
The second movement is the more interesting one. People did not only start using these tools, they stopped using them occasionally. A weekly dabbler becoming a daily user is the behaviour that changes where a purchase decision starts, because a tool you open every day becomes the place you ask things.
Why does the methodology note matter to a seller?
Because it tells you how much weight the figure can carry. Epoch says the March wave asked one question with three response bands, while the August wave asked day counts for each service separately and then collapsed them into the same bands. Their own example: somebody who used one service on three days and a different service on three other days is counted at three days rather than six, so the August figure may understate frequency.
Epoch's conclusion is that comparisons of frequency across the two waves are approximate. That is a more honest framing than the round doubling, and it points the same direction. A seller does not need the number to be exact. A seller needs to know whether to treat assistant driven discovery as a fringe channel or a standing one, and both readings of this data say standing.
A survey about AI use is not a survey about AI shopping. Nothing in the Epoch data says these people are buying. It says the tool is open. Treat it as a measure of habit, and look to your own analytics for whether that habit reaches your checkout.
Where is the money actually being spent?
Increasingly from the sofa, and that trend is older and better measured than the AI one. Visa's Business and Economic Insights team published an analysis of what it calls the couch economy in September 2026, built on anonymised VisaNet transaction data across six markets. In the US, the share of domestic spending happening online went from 48 percent in 2019 to 58 percent in 2026. The United Arab Emirates moved from 35 percent to 55 percent. Poland moved from 10 percent to 24 percent. Brazil moved from 17 percent to 26 percent.
Visa does not mention AI once. That is what makes the pairing useful rather than circular. One dataset says the share of spending that happens on a screen keeps climbing. A separate dataset says the thing people do on that screen every day now includes asking a model. You do not need a causal story connecting them to accept that the screen is where your product has to be legible.
Can an assistant see your product at all?
Often it cannot, and the reason is dull rather than technical. Practical Ecommerce's guidance on testing products for AI discovery makes the point with a query worth reading twice: find waterproof hiking boots under 180 dollars for wide feet, suitable for rocky trails, that weigh less than three pounds.
Count the constraints. Waterproofing, price, width, terrain, weight. Five. A boot can satisfy all five and still never appear, because the width fitting and the weight were never published anywhere a machine could read them. The shopper's request was specific. Your catalogue was not.
This is where a seller's instinct misleads. The instinct says the listing needs better words. What it needs is more published facts. A paragraph calling the boot lightweight and comfortable answers nothing in that query. A weight in grams and a width designation answer two constraints outright.
| Constraint in the request | What answers it | Where it has to live | What happens when it is missing |
|---|---|---|---|
| Under 180 dollars | Price, and the same price at checkout | Page and feed, in agreement | Shown, then abandoned when the price differs |
| Wide feet | A width value on the variant | Variant attributes | Filtered out silently before comparison |
| Under three pounds | A stated weight with its unit | Structured attribute, not prose | Never considered, because weight is unknown |
| Waterproof | The membrane named, not the adjective | Description plus a highlight | Treated as a marketing claim and discounted |
| Rocky trails | Intended terrain and outsole detail | Product detail or a question and answer pair | Matched to the wrong use, then returned |
That table is the whole job. Every row is a fact you either publish or you do not, and the cost of not publishing it is a sale that went to somebody whose feed was more complete rather than whose product was better.
What does Google's feed specification actually ask for?
More than most small catalogues supply, and it says why in plain language. Google's Merchant Center page for the product highlight attribute describes it as short bulleted highlights that help customers discover product information across AI driven surfaces such as AI Mode in Google Search. The limits are specific: a minimum of 2 highlights, a maximum of 100, with 4 to 6 recommended, and each one between 1 and 150 characters.
Those numbers are worth holding onto because they tell you the shape of the work. Four to six short factual lines per product is an afternoon for a catalogue of thirty items and a real project for a catalogue of three thousand. Knowing which you are looking at changes whether you start with your best sellers or with the whole feed.
There is a second and less known group. Google's page on conversational attributes lists six of them: question and answer, document link, related product, item group title, variant option, and popularity rank. Google's stated purpose is that they help AI systems and conversational agents understand a product's specific nuances. All six are optional, which in practice means almost nothing outside large feeds carries them.
The question and answer attribute deserves particular attention from a small seller, because it is the one place where the thing customers keep emailing you about becomes machine readable. If four people a month ask whether the jacket fits over a jumper, that answer belongs in the feed and not only in your sent folder. The same logic runs through turning support tickets into a help centre: the questions already arrived, and the work is publishing the answers where something can read them.
Which identifiers does a machine need first?
The plain ones, published consistently: product name, brand, category, your own SKU, and a global identifier such as a GTIN or UPC where the product carries one. Practical Ecommerce puts this first for a reason. An identifier is how a system knows that the boot on your page and the boot in three other listings are the same object, which is what lets it compare them at all.
Variants are where small catalogues leak most. A size, a colour, a model or a configuration has to be stated as its own value rather than implied by the title. A listing called Trail Boot in several sizes reads to a person as obvious and to a machine as one item with an unspecified size, which loses every request that names a size or a width.
This is unglamorous work and it is also the cheapest work available, because none of it requires a decision. You are not choosing positioning or rewriting a brand voice. You are writing down numbers you already have on a supplier sheet.
Why does price and stock consistency belong in a discovery article?
Because an assistant that recommends you and then sends somebody to a different price has cost you more than one that ignored you. The request was qualified, the shopper arrived intending to buy, and the mismatch arrived at the worst moment. Practical Ecommerce lists verification across the page, the feed, the cart and the checkout, covering price, availability, shipping, delivery timing, promotions and terms.
Promotions are the usual culprit for a small shop. A discount applied in the storefront but never reflected in the feed produces exactly this failure, and it appears only when somebody follows a recommendation. Nothing in your own analytics flags it, because from your side the order simply did not happen.
Worth noting that this cuts against a common assumption. Sellers often treat feed accuracy as an advertising chore, relevant only if they are running shopping campaigns. Once assistants read the same data to answer questions, the feed became a description of your shop that strangers act on whether or not you are paying for placement.
How do you test this without guessing?
You write the request a customer would write, and you never use your brand name. Practical Ecommerce calls this the shop test, and the brand name rule is the part that carries the weight. Searching your own brand proves an assistant can retrieve a label it was handed. Searching the need proves whether it can find you among people solving the same need.
A workable routine for one person with an hour. Pick your five best selling products. For each, write two requests in a customer's words, including at least one constraint you are not certain you publish. Run all ten through the assistants your customers actually use. Record three things per run: did the product appear, was what the assistant said about it true, and which constraint it got wrong or ignored.
The third column is the one that pays. An assistant that describes your product accurately but omits it from a constrained request is telling you exactly which attribute is missing. An assistant that includes it and states the wrong price is telling you your feed and your page disagree, which is a different and more urgent problem.
Two failures that look identical from your side
From a merchant's chair, not appearing in an AI answer feels like one problem. It is two, and they need opposite responses.
The first is invisibility. The assistant never had your product in the set it was choosing from, because an identifier, a category or a feed was absent or broken. Nothing about your copy matters here. The fix is plumbing.
The second is rejection. The assistant had your product, compared it, and set it aside because a constraint went unanswered or a competing item answered more of them. Here the copy and the attributes matter enormously, and the fix is publishing facts you already know.
Telling them apart takes one question: does the product appear for an unconstrained request in its own category? If yes, you have a rejection problem and the constraint table above is your work list. If no, you have an invisibility problem, and getting product categories right across a catalogue is the earlier step. Doing copy work on an invisibility problem is the most common way sellers waste a month.
What does this change about measurement?
Less than sellers hope, and the honesty is worth more than the optimism. Assistant referrals are poorly attributed, frequently arrive without a referrer, and are easy to over read or under read depending on which report you believe. We went through what is and is not measurable in which assistants send traffic you can actually see, and nothing in this week's data changes that picture.
What does change is the value of a manual test. When a channel resists measurement, a repeatable manual probe is not a poor substitute for analytics, it is the instrument. Ten prompts run monthly against your five best sellers, with the answers written down, will tell you more about your discoverability than any dashboard currently available to a small shop.
Where a one person shop should start
Start with your best sellers and the constraints you know customers care about, because that is where a missing attribute costs real money today. Publish the boring facts first: weight with its unit, dimensions, materials, fit or size guidance, compatibility. Then make your page and your feed agree on price and availability, since a mismatch there can lose a sale that discovery already won.
After that, add the four to six highlights Google recommends, and put your three most asked customer questions into the question and answer attribute. If you are choosing what to fix on the page itself, how AI shoppers read a product page covers what tends to be skipped. If you are rebuilding the catalogue structure rather than patching it, the shape of your product data is worth deciding deliberately, and it is one of the things you control completely when you build a store where you own the product data and the code.
None of this is an optimisation trick, which is also Practical Ecommerce's conclusion: what works is complete, specific and trustworthy product information. The reason to do it this quarter rather than next is the 19 percent. Whether the true figure is 19 or somewhat lower, the direction has been consistent across two waves of the same survey, and the people who already made their catalogue legible will keep being the ones who get recommended.