Of the many futures anyone could imagine, more and more dream of worlds in which something called “agentic commerce” is commonplace, or even the dominant mode of online shopping. In these tomorrows to come, agentic AI does the shopping for people based on rich context about their particular needs and deep wells of data about their preferences.

If you’re a hand steering investment in commerce companies or helming one of them directly yourself, you’re going to want to know where this ghost shopping ship is sailing and when it’s leaving port-if it is at all.

An attempt to bring ads into this potential future from square one also has many people asking many questions that definitely need answering.

We’ve got you covered, after some scintillating industry updates.

Many companies will seek to influence LLMs and the agents downstream of them in the anticipation of a new paradigm in which agents make lots of buying decisions.

A publisher and ad tech company joined forces to purposefully serve ads aimed at LLMs, an experiment already drawing opposition from a major lab and pointing to problems with trying to pay to directly influence AI recommendation and commerce engines.

The complex history of SEO and related e-commerce may offer the best lessons in terms of how investors, legal professionals, and business operaots should proceed.

Internet content will likely be the oil that runs the recommendation engine of agentic commerce, and this content presumably will come from publishers and UGC driven platforms, the line between which is increasingly blurred as conventional editorial content and “user generated content” increasingly meld.

Brands may want to use this influence over models to sell things-they’ll probably want to be at the top of any list of recommendations requested, or better yet, to be interjected and suggested in conversations where their product or service is at the furthest possible edge of relevance.

They’ll aim to do this with their digital commerce assets and storefronts, plus with content they produce themself, but this will be fairly minor compared to the battle that will play out in where the real attention goes, and any tactics the brands use will be minor mirrors of their campaigns to capture audiences in major digital media channels.

How said channels work with the LLMs and the recommendation + commerce engines downstream of them is important to understand.

For web publishers and other content creators, identifying some LLM crawlers has been straightforward for a while, as the major labs and companies often have fully declared and well documented crawlers that are easily identifiable, although this is not always the case.

Concerns have been surfaced that there may be an increasing number of third-party crawlers and data sources being user by LLM labs, adding some ambiguity here.

The initial primary concern that drove publishers and platforms to care about identifying LLM crawlers was for the purposes of using robots.txt and other, newer mechanisms to control what content they could touch. Server load was a serious problem, as the sheer volume of requests coming from some of these crawlers could increase the number of requests a website was getting sufficiently to strain its resources and technology.

More complex questions also emerged around intellectual property protection, and how to address the erosion of audience value via LLM results, even if presented with citations, supplanting traffic to actual websites. Publishers are all engaged in a lengthy, largely public community debate about restricting LLM access to their valuable content to protect their traffic-based revenue.

This is part of a larger fundamental debate that’s been raging since Google’s “Zero Click Era” started ramping up in 2024, and various debates about the value of surfacing content to search engines and web portals versus the drawbacks are even older still.

One positive action use case for identified crawler traffic has been to surface your organic content and related metadata in the most LLM crawler friendly way, which is generally though not always some sort of markdown, in the hopes of positioning your content more favorably in such a way that your content appears as a visible source or recommended thing in LLM queries.

This is, again, very similar to the old paradigm’s technical SEO, wherein various website parameters and code documents were programmatically applied to large bodies of web content in order to hopefully get them to rank higher in search results.

Unsurprisingly, many of the most visible thought leaders in this space slid over from the realm of SEO, and dubbed this new discipline G(enerative)EO or A(nswer)EO or something similar.

Since it’s already advantageous for publishers to segment content and code specifically meant for LLM crawlers, and to treat this part of their business differently, it’s natural to see this as a segment of their business, and for publishers to look at it and ask “how much money am I making off of this?”

This has mostly followed standard publisher logic for content monetization. Publishers look at the amount of traffic they’re getting from LLMs, assess the quantitative financial value of that traffic, and determine how much they want to continue investing in generating more of that traffic, relevant to other potentially traffic driving investments they could make.

Two of the main ways they actually get paid for this LLM traffic is the way they get paid for any other traffic: advertising and subscription revenue. While some publishers have struck lucrative licensing deals, the only opportunity for many of them fundamentally takes the same form as any other channel they engage with: traffic to their site.

Things haven’t been looking good for publishers in terms of LLM traffic driven revenue being able to replace the decimation that has happened for many of them on the organic search side since 2025. Keep in mind that publishers have been under an immense amount of fiscal pressure in the past few years to replace dwindling revenue from organic search engine traffic when we look what kind of paths forward they’re seeking.

Well, what about that trusty old source of revenue, digital ads?

Since the beginning, ads have probably been “shown to LLMs” in some way, shape, or form, because LLMs are trained on large quantities of web page content, and some ad content was probably difficult enough to parse out that it passed into model data.

In conventional models of thinking, these would be “bot impressions” and categorized as somewhere between “waste” and “fraud” in many marketing measurement and validation systems, even if there were no ill intent on the side of the LLM doing the crawling and scraping.

Bot traffic can formally be considered General Invalid Traffic (GIVT) or Sophisticated Invalid Traffic (SIVT), and ad impressions shown in these sessions may be categorized the same ways. One might has worse intent behind it, but both are wasted ad impressions.

Recently there’s been a relatively short, if not terribly sweet, and somewhat predictable, story about what happens when someone makes the simple “logical” jump to decide digital ads should be used to theoretically influence LLMs.

Using technology developed by a company called Mobian, Time magazine deployed ads specifically meant to be processed by LLM crawlers on their site, on behalf of several brand advertisers.

Within two weeks of this news hitting the ad industry trade publications, Perplexity announced that they would block these ads, categorizing them as “deceptive.”

The way that Mobian and Time, have thought about commercial agents as something you can advertise to, and how clearly against this notion Perplexity is, serves as a good overview of where everyone is with agentic commerce, in our opinion.

Anyone who arrived at a pro-advertising-to-agents stance followed a simple reasoning chain even Digiday calls “logical” at the beginning of their article, which goes like this:

Companies served ads meant to influence the overarching preferences of people over time so that their shopping habits would change; when agents do the shopping instead, we should move these advertising efforts to sway them.

The thing is that a “role” being transitioned from humans to machines has conventionally always meant a change in behavior towards the machines doing that “role” from how we treated the humans prior.

If a shoe factory made sandwiches at lunch for all its human workers, and then one day replaced them with machines, it would be considered odd if you pitched sandwiches for machines.

When you dial a customer service line and get an automated recording or a bot, you interact with it very differently than you would a real human.

The belief on the pro side would seem to be that digital ads are different than most objects / processes that transition to being from humans to being for machines, and so the transition from showing ads to human online shoppers to agentic online shoppers only necessitates small tweaks and converting the ads to markdown formats.

We haven’t been able to find any convincing evidence or precedent that this is true.

So the burden of proof that it makes sense to show digital ads to LLM crawlers arguably falls on the pro-digital ads for LLM audiences side; digital ads meant to influence LLM systems are clearly distinct from anything that has been done before in the digital marketing space.

It’s worth noting that early work from Kantar and WARC suggests that people and machines, when they see ads, see them very differently, which presents another hurdle.

If the only thing we know so far is that humans and machines see ads very differently, than might the relative effectiveness and potential of showing ads to LLMs instead of to people be different? Can advertisers hope to capture the same value advertising to machines on publisher sites as they did to humans? Half the value? Double the value?

If that value figure still indicates that the juice is even worth the squeeze, does anyone yet understand how different an effective advertising program for machines needs to be from an effective advertising program for humans, and what the operational cost of adding that to the human-facing advertising already being done would be?

I think the answer to all those questions from the pro side would be “no, but we need to start showing ads to machines in order to learn all these things,” which is at the very least an important caveat emptor to any would-be early adopters of these ads.

Outside of that, we have identified five larger concerns that come from older, more well understood problems and models that are applicable or have clear parallels to persuading shopping machines.

The historical default when it comes to digital advertising is that a great deal of effort is put into not showing ads to bots, period.

Some of these bots are benign or beneficial, and help these efforts by self-declaring as bots; others are either negligent or malicious, declining to declare themselves or even masking their nature in many sophisticated ways to purposefully appear to be real users.

Every wrinkle added to the equation of blocking bot fraud makes things more difficult for the detectives, and adding a classification of “good bot” that should be allowed to see ads opens up a whole new kind of obfuscation, with new problems, that fraudsters can exploit.

The way this fraud would be perpetrated is that entities that were not actually the LLM crawlers would pretend to be them by altering and spoofing their characteristics. Essentially, a hijacked device or virtual machine names itself “OAICrawler” and sends a browser agent, hardware version, and all those other parameters that mimic the real OAICrawler.

The best way to solve this would be to collaborate with OpenAI, Anthropic, Perplexity, and everyone else to develop the best possible ongoing verification protocol for the real crawlers, to help other parties more easily identify the bots.

In the case of protecting ads aimed at LLMs against this fraud, the first and only indication from Perplexity is that there is no appetite for collaboration on this, which would mean publishers and antifraud verification tech will be on their own to tackle this whole second category of fraud they now need to deal with.

Digital ad fraud can be measured and fought deterministically, but another way to detect it is to also acknowledge the ultimate problem you’re trying to solve by fighting it: it doesn’t get you any results.

When you’re advertising a product you want humans to buy, failure can manifest very clearly. One indication that you bought fraudulent ads could be that you spent lots of money, on what should have been ads seen many times by a great many people, and you got absolutely no results.

There are obviously many other factors that can make a campaign unsuccessful besides fraud, and you would want to control for those, but at the end of the day a potential sign of fraud, and arguably the most important one, is that nobody’s mind was changed in any meaningful way about your product.

Well, what about when you’re trying to change the outputs of a model having tens of millions of “conversations” with people every day, or in this agentic future, skipping the conversations and shopping for them on autopilot?

This is where you get into the GEO / AEO related topic of LLM visibility monitoring, and while it’s early days for this field, so far it’s been very difficult and full of many known and unknown unknowns, specifically when compared to its forebearer, search engine rank monitoring.

There will be no easy Share of Search, or aided awareness lift, or similar measure that can specifically tell you how you’re impacting a model.

Sales lift and MMM and things may eventually be able to tell you something about those efforts, but that will be about outcomes that ultimately came from purchases by humans-until you have an entirely separate sales channel for agents, which will probably only be the domain of the Amazons and Walmarts of the world for a long time.

That leaves a big too-messy-to-investigate-well middle in between “we did a thing” and “sales went up” if something less than a resounding success leaves you wanting to understand what happened with your LLM influencing ad campaign.

The growing GEO, AEO, and paid LLM ad industries are full of people with a monetary incentive to convince you that this is all so violently novel that there is no predictor for how any of this will go, but so far the short history of paying LLMs to persuade people to buy things rhymes aggressively with the history of search engines coming to be paid ad platforms.

This is partially due to the fact that a major player in the early LLM adjacent ads game are Google, a corporate empire built on a search engine.

The bargain between search engines and their users has always been pretty clear when it come to organic search results versus paid ads on search engines:

The organic results are the real best answers to your query, unfettered by commercial influence. It’s arguable how true that is, but it’s the tacit agreement that drives all of this.

Ads are the only placement that can be outright paid for, and will be clearly demarcated.

This establishes a clear promise to consumers that they will know when they might be influencer by advertising spend, because the only way this will happen is when they see an ad. They will not be influenced unknowingly by advertising, via seeing an allegedly organic result that is influenced by paid advertising.

Is it possible to just spend money and influence these results, though? Yes, through action not sanctioned by search engines and LLMs, which is another potential strike against the idea of using ads to influence LLMs if policies from the major products continue to go against the practice.

Tactics bucketed under “black hat SEO,” whether they be technical or content related, have been around a long time.

Between contracting those services, and buying backlinks, a common and but contested practice, there’s a clear truth to the claim that one could spend money to influence organic results in the short term.

This has also been a major concern for all the major search engines, who care about the integrity of their results, lest users come to find them untrustworthy or useless due to heavy gaming of the ranking algorithm by parties that might not actually be offering them useful content, goods, or services.

A great starting point for understanding this is the storied career and life of Matt Cutts, a Google software engineer tasked with anti-spam and search safety measures for many years at Google who became a well known figure in both the technology and SEO communities.

One reason he became so well known in the SEO community is that he also served as a spokesperson and liaison to publisher organizations and web masters, communicating some Google policy changes and fielding questions about them.

The reason Google and Cutts did this is that there was an understanding that it might be possible for some websites to accidentally engage in what looked like suspicious rank gaming behavior in the course of making good faith changes to their technical setup or content structure. Preemptively communicating some criterion could also help reduce the instances of this.

On the other hand, they had to be careful how much they communicated, lest they tip their hand about any new anti-spam measures to the bad actors.

In the background, there were always people arguing that it wasn’t necessary for Google to do this, nor was it necessary to pay attention to all of this information, because truly good faith entities focused solely on making good content for users wouldn’t actually often be running afoul of these policies.

Still other voices in the conversation said that webmasters shouldn’t pay attention to these Google updates, because they themselves were a bad faith ruse on Google’s part, or at the very least weren’t actually helpful. Many SEO professionals and webmasters have just never trusted Google.

Before we get to the validity and practical application of this kind of system by LLMs in the new era of discovery, an important question looms:

Will they even bother to do this at all? Will there ever be a Matt Cutts of Anthropic or OpenAI?

Google’s positive participation in the publisher and webmaster community has allegedly waned in recent years, and it’s hard not to notice that this development correlated closely with the waning importance of open web publishers in general to Google’s business. As Zero Click results became more common, Zero Care treatment of website owners allegedly did, too.

LLMs have been born into an era where even once staunch defenders of the open web are declaring it dead.

Many of the major sources of online content are even direct, or indirect, competitors to major LLM product makers.

YouTube is owned by the maker of Gemini, every one of the many Meta products is somewhat attached to their LLM efforts, and while it’s not 100% clear what is happening with the new US owned version of TikTok, ByteDance’s history of heavy ML investment must give the frontier labs pause in the context of a supposed US vs. China AI War.

The point is this: if you’re thinking of engaging in behavior that MIGHT not be looked upon favorably by a major LLM product, not only won’t you have a clear place to make your case short of a real court, but you might not even have anywhere clear to direct your questions.

Platform Capitalism has come to rule us all, and even slightly gray areas might now need to be treated like DMZs.

We’ve talked about shopping, a deeply human activity, and haven’t really said a whole lot about people and what makes them special-something we’re coming to understand more deeply the more we see LLMs try to imitate us.

Many people like to shop. Many love it, in fact. People even like the very messy, human things that happen in shopping like impulse buying.

It’s a social activity, as well.

There’s a real question as to how much shopping activity humans would ever even want to cede to any tool, LLM or otherwise.

Even if we wanted to cede a lot of shopping, the sustainability of that arrangement in the long run would depend on ceding it us to something that shopped for us well.

Some people are saying that taste is all that’s left, as this is what LLMs most clearly cannot replicate. Discernment seems to be a problem as well.

What is the point of showing a beautiful Bottega ad to something tasteless?

Why would we expect the delicate thread of exactly what makes us ourselves and how that manifests in what we choose to fill our lives to be grasped by this clumsy data thresher?

Unanswered questions, burden of proof, we said it above and you’ve read it before.

While it’s still a frontier, there are some useful resources emerging around LLM product visibility and best practices. Ensure any companies you counsel or invest in are sticking to those.

As far as a marketing or advertising budget, it’s probably more worth testing the nascent ad products available, or at least monitoring that space, than it is pinning too much hope on precise manipulation of LLM “preference,” via ads or the recreation of SEO’s wild west days on a new frontier.

Caution is a better approach until you have more data, and policies have been further clarified by the chatbot power players, or the lords of agentic commerce that come next.

The MRC’s Invalid Traffic Detection and Filtration Standards 2020 Addendum