The round sizes show which problem this category believes it can solve now. Profound sells generative engine optimization, the practice now challenging established SEO approaches, to brands that want to appear inside model answers [15]. Phia, the assistant that hunts deals on a consumer's behalf, has raised about $43 million in total [12]. One $96 million round against $43 million of cumulative funding is a ratio of roughly 2.2 to 1 [14]. The layer that gets a brand mentioned by a model sells whether or not agents are ever trusted with a card number; the layer that spends the customer's money does not. Personal agents are still raising well, with Instinct and Town both closing large rounds recently [7], but the heavier money sits where the trust question can be deferred.
Phia also supplies the cleanest reason for merchant finance teams to read this as an accounting matter first. It has been caught up in a cookie-stuffing episode, claiming referrals to partners it had not actually earned [12]. That is the same failure mode as an agent buying the wrong item, moved one layer up into the ledger: a commission paid on traffic that may never have been delivered. A retailer that opens a channel to agent intermediaries is opening a channel whose attribution it cannot independently verify.
The single documented consumer failure carries the whole liability question, because the shopper who reported it is the only party the account names as bearing the loss [8]. Newcomer treats errors of that kind as the thing to be resolved before ordinary consumers hand over card details [9]. What the reporting does not establish is who carries the cost as a rule [16]. We do not know yet whether this settles as a card-network dispute, a guarantee written into the agent vendor's terms, or a goodwill refund from the merchant that shipped in good faith. A retailer signing an agent integration this quarter is signing before that price exists.
The bearish case deserves its hearing. One investor told Newcomer that shopping is something people enjoy in their free time, and that e-commerce is still only a portion of global commerce, so offline buying will persist [11]. That argument caps the size of the prize. It says less about sequencing, which is the part a merchant has to decide. Daydream shows the version of the decision that does not require the liability question to be answered at all: since the end of July it has embedded its search into retailers' own sites, so the shopper queries that retailer's catalog on that retailer's domain [10]. The transaction never leaves the merchant's control.
One reading tracks power users, the other tracks the aggregate, and that gap is why the investor commentary conflicts without either side being wrong. M13's Mark Grace says consumer use accelerated significantly this summer and that early adopters can already find agents that will transact for them [5], and that is compatible with Forerunner's read that the aggregate data has not arrived [6]. Power users can be buying while the totals stay quiet. The distinction that matters is between work that is cheap and reversible, structured product data and a returns policy an agent can parse, and work that is neither, meaning agent-initiated checkout and commissions settled inside an attribution system someone else built. This quarter, the decision is a catalog project; next quarter, it becomes a contract.
