When an AI Agent Shops, Who Is the Consumer?

Retailers spent two decades learning how to be found by Google, but the next audience might not be a person typing a query; instead, it could be software acting on a person’s behalf. Agentic commerce tools (AI “agents” that browse, compare, and increasingly transact for users) are quickly moving from demonstrations to products. When an agent does the shopping, the person who dispatched it may never see the listing it selects, the webpage it reads, or the source code behind either.

This shift raises interesting questions around branding and trademark law, where liability often turns on whether the public is likely to be confused. This inquiry was built around a human consumer who, for example, sees a storefront, reads product packaging, scans a domain name, or reviews a page of search results.

Canadian courts have grappled with what confusion means in the context of e-commerce, including where a website embeds content (metatags, keywords, and hashtags) aimed partly at machines. However, courts have not yet addressed what happens when the shopper itself is a machine reading the machine-oriented content that earlier case law assumed no consumer would ever see.

As such, this comment examines where descriptive-use and confusion case law stands today and where AI-driven commerce may change the calculus.

Background: Passing Off, Confusion, & “Use”

Subsection 7(b) of the Trademarks Act is the formal expression in statute of the common law tort of passing off (see MacDonald v Vapor Canada Ltd). To succeed, a plaintiff must establish the three elements confirmed in Ciba-Geigy Canada Ltd v Apotex Inc:

1. The existence of goodwill
2. Deception of the public due to a misrepresentation
3. Actual or potential damage

The analysis is conducted from the perspective of a particular hypothetical shopper. As Justice Binnie explained in Mattel Inc v 3894207 Canada Inc, the relevant person is neither the careful, diligent purchaser nor the “moron in a hurry,” but a mythical consumer in between (referred to in the case law as the “ordinary hurried purchasers”). In Veuve Clicquot Ponsardin v Boutiques Cliquot Ltée, the Supreme Court described the test as a matter of first impression in the mind of a “casual consumer somewhat in a hurry” who has only an “imperfect recollection” of the original or earlier-registered trademark and does not pause to scrutinize the differences.

Importantly, confusion is assessed as a first impression. In Masterpiece Inc v Alavida Lifestyles Inc, the Supreme Court held that the relevant moment is when the consumer encounters the mark, not after any later research or purchase.

However, not every appearance of another party’s mark is necessarily actionable. Using a mark merely to describe or to refer to the other party is sometimes approached differently from using it to distinguish one’s own goods. In fact, the Trademarks Act itself preserves the ability to make an “accurate description” of the character or quality of one’s wares or services (subsection 20(1)(b)(ii)).

As discussed below, courts have declined to find liability where a mark is used to identify or comment on the mark owner rather than to divert its customers. However, the underlying rationale for these decisions becomes less certain on the internet, where use of the mark may appear in code or metadata the actual consumer never sees.

When the Trademark is Invisible: The Metatag Cases

Over a decade ago, the Federal Court considered what happens when a competitor’s mark appears in page code consumers cannot see. In Red Label Vacations Inc v 411 Travel Buys Ltd, the defendant, a competing online travel agency, had copied the plaintiff’s metatags, including the plaintiff’s “red tag vacations” and “shop, compare & pay less” brands, into the source code of dozens of its webpages. Metatags are words embedded in a page’s source code that are not visible on the page itself and are used to influence how search engines rank the site. None of the plaintiff’s marks appeared anywhere a visitor could see.

The Court found no passing off. It held that “initial interest confusion” (the US concept of luring a customer with a competitor’s mark before any purchase) had not gained a foothold in Canada and, in any event, did not fit the facts. In brief, the metatags merely give a searcher a choice of links, and once the searcher arrives at the site, the site’s own identity dispels any confusion:

Applied to the facts: “the website is clearly identified as 411 Travel Buys’ website,” so there was “no likelihood of deception as to the source,” and the consumer remained free to redirect to the plaintiff. The decision turns in part on the fact that the offending text was invisible to the consumer, and that the decisive moment was when the consumer arrived at a clearly labelled site.

On appeal, the Federal Court of Appeal affirmed (see Red Label Vacations Inc v 411 Travel Buys Ltd), but it did so largely on deference, finding no overriding error in the trial judge’s findings of fact. Importantly, the Court declined to endorse the trial reasoning wholesale. The majority observed that “in some situations, inserting a registered trade-mark (or a trade-mark that is confusing with a registered trade-mark) in a metatag may constitute advertising of services that would give rise to a claim for infringement” but not on the factual record in issue. In concurring reasons, Justice Dawson added that the extent to which a mark may appear in metatags without infringing is “of necessity, fact specific,” and cautioned that dismissing the appeal did not mean the Court endorsed every alternate basis relied on below, including the trial judge’s remarks on initial interest confusion.

When Does Confusion Happen & Does Descriptive Use Save It?

Two further lines of authority frame the debate.

 Descriptive and Commentary Use 

In ICBC v Stainton Ventures Ltd, a lawyer operated “ICBCadvice.com,” a site offering (often adversarial) guidance on dealing with ICBC. The Court found no passing off and no breach of ICBC’s official marks (a category of government-related trademarks under Canadian law). It reasoned that using the acronym to refer to ICBC was what “thousands do every day” (identifying the subject of the site, not its source) and made a point that might resonate in any discussion of machines: “[t]he behaviour of search engines is not […] evidence of anything other than the operation of an algorithm, and search-engine marketing. It is certainly not evidence of confusion.”

The Court of Appeal for British Columbia agreed (see 2014 BCCA 296), crediting the ordinary internet user with understanding that a domain name containing a business’s name or acronym is not necessarily affiliated with it and may instead signal the site’s subject matter. The “most that person would conclude,” Justice Frankel held, is that the site “likely had something to do with the corporation.” Descriptive, referential, and critical uses of a mark, on this view, do not amount to passing off.

The Timing of Confusion

In Vancouver Community College v Vancouver Career College (Burnaby) Inc, the Court of Appeal found passing off where a private college used the acronym “VCC” and the domain “VCCollege.ca” in its search advertising. Importantly, the Court held that the moment for assessing confusion is the first encounter (when the search results appear) not when the searcher reaches the landing page. The Court distinguished ICBC on the basis that “ICBCadvice” itself disclaimed affiliation, whereas “VCCollege.ca” was equally descriptive of the appellant and carried its acronym with no disclaiming words. Of note, the Court found Red Label unhelpful on the temporal question and considered itself bound by the Supreme Court’s decision in Masterpiece: confusion is judged at the point of first impression, before any subsequent research or purchase.

The Tension Between These Lines of Authority

Although these cases turn highly on their facts, there is arguably a tension. Red Label discounts confusion by looking to the moment the consumer arrives at a clearly identified site; Vancouver Community College locates the relevant impression earlier, when the results first appear. The Federal Court in Wanakome Inc v Martin recently acknowledged this tension in the context of defendants arguing that hashtags should be treated like search-engine metatags to avoid a finding of infringement.

The Federal Court walked through both Red Label Vacations and Vancouver Community College, acknowledged the tension, and distilled the governing principle: “It is the association that would be made from the use of the hashtag that is key, and whether this would create confusion in the mind of the consumer as to source.” On the facts of that case (short-term hashtag use with no evidence of confusion or damage), there was no passing off. The Federal Court of Appeal subsequently upheld this finding on an absence of palpable-and-overriding-error standard (see Wanakome Inc v Martin).

What (If Anything) Changes When the Consumer Is an AI Agent?

Machine-Readable Content Becomes the Front Line

To a human consumer, metatags, alt text, structured product data, and off-screen text are arguably functional and largely unseen, which is why the defendant in Red Label Vacations could argue no consumer would be deceived by them. An AI agent, however, reads exactly that layer. Product feeds, schema markup, hidden descriptions, and even instructions addressed to automated tools are all content an agent parses to decide what to surface or buy.

This raises a direct question: how should a court approach a competitor that embeds a rival’s mark in machine-readable text to influence an agent’s selection? In such a scenario, the safe harbour for machine-oriented indicia (such as metatags) may no longer hold. Whether an agent “seeing” text qualifies as the kind of encounter the confusion inquiry contemplates remains to be seen, and we expect parties will argue both sides of this issue on the facts.

The “They’ll See Our Real Site” Defence May Fail

Red Label Vacations discounted confusion because “once that person reaches the website,” it was plainly the defendant’s site and associated services. However, an AI agent may compare options and complete a transaction without the human consumer ever visiting a site. If the relevant impression forms when results (or a shortlist) first appear, Vancouver Community College’s framing may prove more persuasive to a court than Red Label Vacations’.

This refines the threshold question courts have been circling: from whose perspective, and at what moment, is confusion judged? If the relevant first impression is the agent’s shortlist, the analysis resembles Vancouver Community College; if the relevant perspective remains the human consumer who dispatched the agent, the question becomes whether that person, relying on an agent diverted by a rival’s mark, was misled as to source.

Could Courts Remove the AI Agent from the Confusion Test?

Courts mindful of the AI agent’s role might decline to let the agent drive the test at all. An agent that ingests and synthesizes all content on a webpage arguably does not form the same kind of impression as the “casual consumer somewhat in a hurry” from Veuve Clicquot. A court might approach the first impression at the point where a human consumer enters or re-enters the loop (typically the landing page, cart, or order-confirmation screen), because that is the first moment a human being encounters the mark.

The difficulty with this approach is that, in a fully automated purchase, a human consumer may never be in the loop. This could leave trademark owners without a remedy where a rival uses machine-readable content to compete unfairly.

Advertising, Initial Interest Confusion, & Agent Diversion

Embedding a mark in machine-readable text to steer an agent’s choice might be characterized as “advertising of services” of the kind the Federal Court of Appeal in Red Label Vacations said may ground a claim. “Initial interest confusion,” expressly not adopted in Canada but not necessarily foreclosed, may also bear on such agent diversion, where the relevant diversion happens before the transaction.

An AI agent, like the “ordinary hurried” human consumer, might be taken to “understand” that a competitor’s mark in a comparison is not a claim of affiliation. But whether an agent reads text more literally, or with more surrounding context, than a hurried person will turn on the specifics. We expect defendants to frame agent-facing uses as descriptive or referential, and plaintiffs to frame them as diversionary advertising.

Key Takeaways

1. Monitor machine-readable content. Brand owners should monitor how competitors use their marks in metatags, structured and product data, alt text, agent-facing feeds, and instructions addressed to automated tools, not only on visible pages.
2. Optimizing for AI agents carries risk. For businesses treating agent-facing content as the next SEO, using a competitor’s mark to capture agent-driven traffic carries more risk than the metatag cases might suggest. The Federal Court of Appeal left the door open to infringement, and the “the human consumer will still see our real site” defence may ring hollow when the human consumer never actually visits.
3. Facts and evidence will be extremely important. Both Red Label Vacations and Wanakome were decided on their records and upheld because the appeal court saw no clear error in the trial judge’s findings. Evidence of how an agent encounters and acts on the text, and of actual diversion or damage, will likely matter. Parties should consider whether expert evidence would assist the Court in understanding the relevant technologies and code.