What does a shopping agent read before it puts a brand on the shortlist? Not the campaign calendar, and not the hero image. It reads fields: price against a timestamp, stock, delivery window, return terms. Salesforce projects AI to drive 75.8 billion dollars of Cyber Week spending, 13 percent above 2025, and as of November 13, 2026 that single figure is what every merchandising argument now has to answer to.
Referral quality is not referral volume
The volume story is the easy one, and it is already priced in. Salesforce puts AI-referred visitors at more than twice the conversion rate of typical ecommerce traffic, with an 8 percent higher average order value and 23 percent more items per order when a shopper engages a brand's own agent. That is a strong number and a fragile one, because most shoppers still abandon an assistant after two or three mistakes.
Volume without quality is a discount on attention. A referral that arrives because an agent listed the brand next to a cheaper alternative converts at the price of a comparison, not at the price of a recommendation. The distinction only shows up when referral sessions are read by category and by assistant, which most analytics stacks still flatten into a single line called AI.
What the benchmark actually tested
Visa's Agentic Commerce Benchmark, released on November 5, ran 22 models from 12 labs through 15 shopping challenges and passed 57.2 percent of runs. The number worth carrying into a merchandising meeting is 34.2 percent, the share of episodes where all three attempts succeeded. A system that works two times out of three is not reliable enough to build a checkout flow on.
The same study locates the failure. Models that surfaced the most hidden requirements averaged 70.2 percent, against 44.4 percent for the five that surfaced the fewest, and 54 percent of failed runs still ended in a purchase, often of the wrong product. An agent can only surface a requirement it can read, which turns catalogue structure into a conversion variable.
Five fields an agent can verify
The fields that decide agent outcomes are not the ones most merchandising teams optimise. Imagery, copy and campaign calendars move human shoppers; an agent filters on constraints first and ranks on evidence second. Five fields carry almost all of that weight.
| Field | Why it decides the outcome | Failure when missing |
|---|---|---|
| Hard constraints | Size, compatibility and usage limits act as filters | Product excluded before any brand consideration |
| Price with a timestamp | Agents compare live values and drop stale entries | Wrong price quoted, order cancelled later |
| Unit of measure | Comparisons across pack sizes need a common denominator | Unit price misread, product judged uncompetitive |
| Return terms | Policy is checked before a purchase is proposed | Agent routes the shopper to a competitor |
| Verified review signal | Aggregated feedback ranks options | Ranked below a weaker but better documented rival |
None of these is new, which is the uncomfortable part. Most catalogues already hold the information somewhere, just not in a form a retrieval system can use: a specification sheet in a PDF, a compatibility list in an image, a price that refreshes on a different cadence than the listing.
Consistency beats richness
A single excellent product page does little if the same item carries different specifications, prices or claims across three marketplaces. The agent reads whichever version it retrieves, and the retrieved version is not always the one the brand maintains most carefully. Aligning titles, attributes and units across channels therefore comes before adding descriptive depth.
This is where promotional policy becomes a data problem. A discount that exists in a campaign tool but not in the feed is invisible, and an expired promotion that lingers in a feed is worse than invisible, because it produces a promise the checkout will not honour.
Where promotional depth breaks
Attest's finding that 62 percent of AI-assisted shoppers excluded a product on an assistant's advice without checking it separately should reframe how promotional spend is judged. The exclusion happens before a brand is ever seen, so it generates no session, no bounce and no remarketing audience. A price monitoring programme that only watches competitor pages will report a stable market while the shortlist quietly narrows.
Gartner's survey adds the institutional half: 42 percent of CMOs already monitor and correct AI output misinformation, and 61 percent have adjusted content strategy because of it. Both numbers imply a workflow, not a campaign. Someone has to own the correction, and the correction has to reach the feed the agent reads.
The measurement gap
Ahrefs data from 963 French domains shows a median 5.7 percent drop in click-through rate after AI overviews arrived, with 8.8 percent fewer clicks overall and a 5.4 percent decline in ecommerce. Clicks are the metric that disappears first, and they are the metric most teams still report on. Replacing them means measuring whether a brand appears in an answer at all, and with which facts attached.
A monthly routine for the feed
Most of this work is unglamorous maintenance, and that is the point. The brands that gain from agent referrals will not be the ones that wrote the cleverest copy. They will be the ones whose price, stock and policy were already readable when the question was asked.
FAQ
Do AI referrals really convert better?
Salesforce reports that shoppers arriving through AI referrals are more than twice as likely to convert, and that those engaging a brand's own agent spend 8 percent more per order. The caveat is that most shoppers abandon an assistant after two or three mistakes, so the gain depends on accuracy rather than reach.
Which listing field matters most?
Hard constraints, because they act as filters. If compatibility, size or capacity is missing or ambiguous, the product is excluded before price is even compared.
Why does promotional data need a timestamp?
Agents compare live values and drop stale entries. A price without an update time reads as uncertain, and an expired discount that stays in the feed produces an order the checkout cannot honour.
How should a team measure AI visibility?
By asking the assistants the questions shoppers ask, then recording whether the brand appears and which facts the answer repeats. Site analytics cannot see an exclusion that never became a session.
