The AI assistant grew revenue 620 percent. Ask what it grew from.

TL;DR

On its Q2 fiscal 2026 earnings call, Williams-Sonoma said revenue associated with its Olive shopping assistant is up 620 percent since the start of the year. On the same call, in the same quarter, total net revenue grew 6.7 percent to $1.96 billion. The company gave a base for the second number and no base at all for the first: no absolute figure, no share of revenue, no shopper count. A growth rate roughly 93 times the growth rate of the business around it is arithmetic about how small the starting point was, and reporting it without the starting point is a choice.

What Williams-Sonoma actually reported

For the fiscal quarter ended 2 August, Williams-Sonoma Inc. reported net revenue of $1.96 billion, up 6.7 percent from $1.84 billion a year earlier. Ecommerce comparable revenue rose 6.5 percent, stores rose 5.5 percent, and total comparable brand revenue rose 6.2 percent. Those are the numbers chief financial officer Jeff Howie gave investors, and every one of them arrives with a base you can check.

Then chief technology and digital officer Sameer Hassan gave the assistant numbers, as reported by Digital Commerce 360 and confirmed in the Q2 fiscal 2026 earnings call transcript. Since the start of the year, engagement with Olive, the assistant on the Williams Sonoma site, is up 700 percent. Revenue associated with Olive is up 620 percent. Customers who use Olive convert at three times the rate of other shoppers. On Otto, the assistant launched across the Pottery Barn portfolio in August, more than 70 percent of engagements are resolved without transferring the customer to an employee.

Every figure in that paragraph is an executive statement on an earnings call. None of it is independent measurement, and none of it is audited. That does not make it false. It makes it a claim, and claims are read differently from findings.

Put the two sets side by side and the pattern is hard to miss.

The disclosure standard changes at exactly the point where the number becomes flattering.

620 percent of what?

Divide 620 by 6.7. The assistant grew about 93 times faster than the business it lives inside, over overlapping periods, on the same storefront, sold to the same customers.

There is only one way a subset of a business grows two orders of magnitude faster than the business: it started very small. That is not a criticism of the assistant. It is what the arithmetic says, and the arithmetic is the whole of what the number can support.

The 4GPTs read: a growth rate with no denominator is not a performance metric, it is a launch metric. It tells you something happened. It cannot tell you whether the thing that happened matters yet, and the company is the only party that could settle that, in one sentence, by publishing assistant-attributed revenue as a number.

Be fair about what this does not prove. Olive may be working well. Williams-Sonoma sits at No. 22 in Digital Commerce 360's Top 1000 and at No. 306 in its AI Rankings, it took share in a flat home furnishings market, and it raised its full-year outlook. A company doing that is not obviously fooling itself. The point is narrower and it survives all of it: the reader cannot check the claim, because the one figure that would make it checkable was not given.

The one figure that is load bearing

Buried under the headline ratio is the number that actually carries weight.

Nine times, up from two times. That figure has a stated prior period, which means it is a comparison rather than a ratio off an unstated floor. You can see the direction and the size of the move. It is also the honest one to lead with, and it is about personalization across the sites rather than the assistant specifically. It went up by a factor of 4.5 in a year and nobody put it in a headline, because 4.5 reads smaller than 620 while saying considerably more.

Three times conversion is selection, not causation

The stickiest number in the set is that Olive users convert at three times the rate of other shoppers. It will be repeated in decks for a year. It is also the one that needs the most care, and the write-up passes it on without a caveat.

Think about who opens a product assistant. Somebody already on the site, already looking at a category, already close enough to a purchase to want help narrowing it down. The comparison group is everyone else, which includes people who bounced from the homepage. Comparing those two populations measures who chose to engage at least as much as it measures what the engagement did.

The 4GPTs read: this is not a Williams-Sonoma problem, it is the default shape of every assistant metric published this year, ours included when we get there. Any tool a motivated user opts into will out-convert the average of everyone. Until the cohort is matched on intent, a conversion multiple is a description of who clicked, not of what the software did.

What the company called it

The most interesting line on the call never made it into the trade coverage. Describing the Pottery Barn rollout, Hassan said Otto begins to bring "the agentic discovery experience we are building" to the furniture brands, and that "it is not just a chatbot".

A top-25 US online retailer is using the word agentic for its own storefront, in front of investors, unprompted. That is worth more than the 620 percent. The supply side of agentic commerce is arriving, and it is arriving built and owned by the merchant.

Which is the right move for that merchant, and it has a ceiling built into it. An assistant that belongs to a store holds that store's catalog. It can rank inside it, coordinate within it, and, in Hassan's description as reported by Digital Commerce 360, "coordinate items that go together, sofa to rug". What it structurally cannot do is tell you the right thing for you is not in this catalog. The shopper's side of the same transaction holds something different: the requirement. Size, use, budget, the date it has to arrive, the thing this person will not compromise on.

The 4GPTs read: those are two different jobs and they need two different agents. A store's agent answers for its catalog. A shopper's agent holds the spec and shops the market against it. Both sides being agents is what makes the transaction a conversation instead of a listing, and the missing piece is not another assistant on another storefront. It is the connection between the two. That is the layer 4GPTs builds, and this earnings call is the supply side of it showing up on its own.

For a wider view of how quickly that supply side is appearing, and why the count matters more than the leaderboard, see our read on why the AI 1000 count is the number to watch and the ranking is not.

What a merchant does on Monday

Four things, and none of them take a quarter.

1. Publish the base with the ratio. Report assistant-attributed revenue as an absolute number and as a share of total revenue. If the share is small, that is fine and it is early. A ratio without it invites exactly the question this article is.
2. Match the cohort before claiming lift. Intent-matched comparison, or the conversion multiple goes in an internal deck and not in a press release.
3. Name the prior period. The 9x figure is credible because it says what it was last year. Do the same for everything else.
4. Write down what your assistant can see. One catalog. That is not a flaw, it is a boundary, and knowing where it sits tells you which shopper questions you will never be the one to answer.

FAQ