I gave my AI agent a two-week free trial to a subscription-based site. Is that the new business model, or an existential threat to the business?

This is exemplary of a debate playing out about agentic marketing. If you're a retailer, publisher, or brand, to what extent do you let agents in? Can you afford to, or can you afford not to?

Amazon, for instance, is notoriously against letting AI agents shop its site. When I ask the agent Instinct for product recommendations, it doesn't send me any Amazon listings. It’s like it doesn't have a valid visa for Amazon, so the agent won't even try to visit there. Still, thanks to the strength of Amazon's brand, the convenience of it, and the reliability of its shipping windows, I often then wind up buying Instinct's recommended products on Amazon anyway. Few retailers have such advantages.

LinkedIn also is tougher for bots to crack, mostly in a good way. Some bots still have their way with the site, but you risk your account getting banned if you push it too far, and agentic apps like Instinct limit what they’ll do. It’s like they follow Isaac Asimov’s laws of robotics and won’t let humans be harmed.

The mission I gave Instinct was a practical one. I set up a two-week free trial of Ancestry.com after giving Instinct access to a folder of mine with files about my family tree. Instinct had this window to keep searching Ancestry on a daily basis. As it presented its discoveries, including several helpful documents I never saw in two decades as an on-again, off-again customer of Ancestry, I had it refine searches to focus on new leads.

The biggest inherent risk for a company like Ancestry is that it will fall victim to an agentic army of freeloaders. Such freeloaders could continually create new accounts every two weeks and continue searching.

Ancestry could foil that by tying trials to credit cards rather than email accounts. There are ways like that to make freeloading harder. And humans who are determined to freeload have spent decades swearing, "Information wants to be free!" while coming up with ways to avoid paying the creators and brokers of such information. Freeloaders are going to freeload, agents or no agents.

There are risks of excluding agents, too. Instinct can search for the original sources that feed into Ancestry. Additionally, if Ancestry proves so useful to my bot, I may let the free trial convert into a subscription.

Even agent-friendly sites will need to consider some forms of rate limiting. And genealogy offers a great example as to why.

If I'm doing a lot of genealogy research, I might do 50 to 100 searches a day for a few days (I get very obsessive); I want that kind of access with a paid account. If I'm a genealogy researcher doing this for others, I might want to do hundreds of searches a day — but presumably that might come with a higher fee. 

If I'm doing 10,000 or 100,000 searches a day, it means I'm trying to scrape the entire database, and that shouldn't be allowed. The agent, hitting a rate limit, could then try to simulate 1,000 users each doing 10 searches a day. That requires an inordinate amount of work, which a site like Ancestry could block via methods described above to make the effort cost-prohibitive.

Determining how to accommodate agents will require us to have a better understanding of human versus agentic behavior. There's a lot we don't know, and it will evolve. Will agents be more or less efficient than human behavior, for instance? Consider some differences:

Humans are constrained by time, and bots are constrained by tokens, or some equivalent.

Humans often have a mission that they get sidetracked from. Agents may not get sidetracked, but they often have to iterate as they interpret their mission, and it may take many more permutations to achieve a result.

Humans are often irrational. Agents often misunderstand orders and hallucinate.

Humans are vulnerable to upselling and cross-selling. Will that be the case for agents?

So what can you do about this?

Relatively few publishers and brands can afford to close their doors to agents. On the product side, it's like intentionally avoiding distribution channels such as Amazon or Walmart. On the content side, there are only so many outlets (like Marketecture, of course) that give you information you can't get anywhere else. 

The trick won't be how to let in the agents. It'll be how not to be abused by them in a way that hurts your bottom line. Agents doing the bidding of humans means there are humans who might find what you have worth paying for. The biggest risk is when agents seem to be doing their own bidding.

— David Berkowitz, Chief Community Officer, Marketecture Media

Who: Retail Marketers, E-commerce Leads, Brand Marketers, Agency Strategists

What: Well, this one is timely. Walmart is building infrastructure to welcome outside AI shopping agents while Amazon walls its platform off, Axios reports. Its new Agent Gateway, built on MCP, lets approved third-party agents shop at Walmart and Sam's Club, while Agent Profiles let customers set spending limits, require purchase approval, and cut off access, with Walmart keeping hold of sensitive customer data. Amazon, meanwhile, blocked Meta's Muse agent less than two weeks after launch, citing security; analyst Neil Saunders argues Amazon is protecting a closed ecosystem whose ad business agentic shopping could undermine. Only 16% of shoppers are currently comfortable letting an AI buy on their behalf (Coveo), but Walmart's own Sparky assistant saw 60% quarter-to-quarter growth.

Why it matters: The two biggest retailers just picked opposite sides of the agentic commerce question, and every brand that sells through either one inherits the consequences. Walmart's bet — that "everyday low prices" win when agents hunt for deals — is a media-strategy question as much as a retail one: What happens to retail media when the shopper is a machine?

(Axios)

Who: Publishers, Brand Marketers, Content Strategists, SEO/GEO Leads

What: The Daily Mail is running 12 to 15 GEO tests across the U.S., UK, and Australia. It’s tracking consumer prompts across AI platforms, auditing which sources get cited, and creating bespoke branded content to lift a brand's visibility — ahead of launching an official AI visibility product early next year, Digiday reports. In one early test for an unnamed personal finance brand, visibility across target prompts rose from about 50% to nearly 75%. Director of SEO Carly Steven frames it as a revenue opportunity, not a defensive play, but admits attribution is the hard part: One brand found Claude drove just 0.1% of referrals yet roughly 11% of conversions once it asked customers directly how they heard about it. Future already has more than 30 GEO clients; Forbes and CNN are building consultative versions.

Why it matters: GEO is crossing from agency service to publisher product, which changes who brands pay and what they're buying: citations, not clicks. The 0.1%-of-referrals versus 11%-of-conversions gap is the stat to steal; referral analytics are nearly blind to AI's actual influence.

(Digiday)

Who: Agency Executives, Creative Technologists, Production Leads, MarTech Decision-Makers

What: WPP's research and creative engineering teams benchmarked NVIDIA's new Vera Rubin hardware on Google Cloud for agency production workloads, publishing the results on WPP Research. A 120-billion-parameter Nemotron vision-language model ingested an uncompressed 84-second 4K commercial master — 4.18 GB of frame data — and produced a frame-accurate audit with scene logging, OCR, and continuity checks in 19 seconds. It flagged 28 scene cuts and cross-campaign brand-compliance issues in just over a second. In a second test, a single 4-GPU node served 1,024 concurrent users and autonomous creative agents at over 43,500 total tokens per second with zero failures. WPP says the system learns from creative decisions: Every user choice between generated assets becomes training telemetry.

Why it matters: This is the production math behind the holdco AI pitch, published with real numbers instead of sizzle: Quality assurance that took days now runs faster than playback. The "every run is recorded" learning loop is the strategic tell; WPP is trying to turn client work into proprietary training data, which is exactly the asset agencies can own that platforms can't.

(WPP Research)

Liking AI Brief? The Refresh is a weekly recap of all the most important advertising news, written by AdTechGod.

Thanks for reading this week’s AI BRIEF by Marketecture. 

Marketecture Media is committed to bringing you the latest news, opinions, and community updates. 

If you are interested in advertising on this newsletter, our podcasts, The Advertising Forum, or in our Slack community, you can contact us here.