As of November 7, 2026, shopping agents have moved from pilot to production. Salesforce said its Shopper Agent, Buyer Agent and Merchant Agent are generally available with native integration into ChatGPT and Google Search, including AI Mode, with Gemini app support to follow. The commercial case it cites is large: AI influenced 20% of global online holiday sales in 2025, about USD 262 billion, retailers running their own shopper agents grew sales 59% faster than those that did not, and AI-referred traffic converted at eight times the rate of social traffic. For ecommerce teams, the operative question is narrower than the headline: if an agent decides what to surface, catalog accuracy stops being a merchandising nicety and becomes a margin line.
One false start: the agent is not shopping, it is filtering
An agent asked for a product does not browse a category page. It reads structured fields, filters by constraints, and returns a short list. The filtering happens before any human sees a brand, which is why catalog quality now behaves like a paid placement: the products that survive the filter get the impression, and the ones that fail never enter the consideration set at all. That shift is invisible in traffic reports, because a product excluded by an agent generates no session to measure.
What the benchmark says about catalog quality
A Visa benchmark released on November 5, 2026 tested shopping agents across episodes and repeated runs, and the results are less flattering than vendor demos suggest. Agents failed 42.8% of runs, and only 34.2% of episodes passed all three attempts. The spread between a single successful run and three consecutive successes is the number that matters operationally, because a merchant cannot build a checkout flow on a system that works two times out of three and calls it reliable.
| Measure | Result | What it implies for merchants |
|---|---|---|
| Runs passing the benchmark | 57.2% | Roughly two in five agent sessions still fail a basic task |
| Episodes passing all three runs | 34.2% | One good run proves little; consistency is the real gap |
| Best model pass rate | 82.2% | Model choice matters, but no model clearly leads |
| Cost per run, top versus cheaper model | USD 0.54 vs USD 0.02 | Cost per correct outcome is the number to compare |
The benchmark's most useful finding for catalog owners is the discovery gap: the five models that surfaced the most hidden constraints scored 70.2% against 44.4% for the five that surfaced the fewest. Better questions, not bigger budgets, separated the leaders. Translated to catalog work, an agent only finds a constraint if the catalog states it as a field. Attributes buried in marketing copy, or expressed differently across marketplaces, are effectively invisible.
Consistency beats one perfect listing
A single well-written 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. That is why the practical fix is consistency first and richness second: align titles, attributes and units across channels before adding new descriptive depth, because consistency determines which version wins.
Three catalog fields agents actually read
The fields that decide agent outcomes are not the ones most merchandising teams optimize. They are the ones that allow a machine to eliminate a product without asking a human. Merchandising tends to invest in imagery, copy and campaign calendars, because those move human shoppers. Agents read none of that in the first pass: they filter on constraints, then rank on evidence, which means the fields below decide whether a product is ever considered at all.
| Field | Why it decides the outcome | Failure mode when missing |
|---|---|---|
| Hard constraints | Size, compatibility, capacity and usage limits are filter conditions | Product excluded before any brand consideration |
| Price and stock as of a timestamp | Agents compare live values and drop stale entries | Wrong price quoted, order cancelled later |
| Unit of measure | Comparisons across pack sizes require a common denominator | Unit price misread, product judged uncompetitive |
| Verified review signal | Agents weigh aggregated feedback when ranking options | Product ranked below weaker but better-documented rivals |
None of these fields is new, and that is the uncomfortable part: most catalogs already contain them 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 updates on a different cadence than the listing, all read as absent. The work is less about writing new content and more about moving existing facts into fields that are unambiguous, machine-readable and consistent across every surface where the product appears.
Zero-click changes the measurement, not the value
Search behavior has already shifted underneath this. AI Overviews appear on 86.8% of commercial queries, organic click-through on affected queries fell from 1.41% to 0.64%, and only 0.3% of AI Overviews currently cite ecommerce sources. The last figure is the one to sit with: for now, most AI answers about products do not link to a store at all. Brands cited in AI Overviews do earn more clicks when clicks happen, but the strategic point is that being the source of a correct fact is becoming separable from being the destination of a click.
Four steps before the holiday peak
Catalog remediation is boring work with a hard deadline, so it needs a sequence rather than a project plan. The temptation is to start with the products that matter most commercially, which is usually the wrong first move: the products that matter most are already the best documented, while the ones silently excluded by an agent are the ones nobody has checked in months.
The fourth step matters more than it looks. Comparing ranking positions invites arguments about noise, because positions move between runs even when nothing changes. Comparing the surfaced set is more stable: a product either enters the answer set or it does not, and that binary is far easier to attribute to a catalog fix than a shift from fourth to third.
Where the margin leaks
Two leaks follow from inaccurate agent answers. The first is discount-driven: an agent quotes a lower price than the checkout shows, the shopper abandons, and the brand pays for the click without the sale. The second is return-driven: an agent overstates a capability, the product arrives and underdelivers, and the return costs more than the margin. Both leaks sit outside the catalog budget and inside the P&L, which is why catalog accuracy tends to get funded only after a bad quarter rather than before a good one. The underlying shift is in consumer demand rather than in technology: shoppers now delegate the shortlist to an agent, so the catalog has to satisfy a reader that cannot see a photograph and will not ask a clarifying question.
FAQ: agents, catalogs and margin
Do shopping agents replace search ads?
Not yet. They change the entry point for discovery, but paid search still captures intent that agents route to a shortlist. Budget should follow attribution, not fashion.
Why not optimize for the highest-scoring model?
Because no model clearly leads, and pass rates vary widely within a single vendor. Optimizing for the fields all agents read is more durable than chasing one leaderboard.
How much catalog work is enough?
Enough that hard constraints, price with a timestamp, unit of measure and review signal are consistent across every marketplace. Beyond that, returns diminish quickly.
Should brands build their own shopper agent?
Only if the catalog is already clean. An agent built on inconsistent data will surface the inconsistency faster than it surfaces products.
