What Comes After the Composable CDP?

For a bunch of smart people, the technology community can be awfully slow learners. One mistake they keep repeating is believing that the latest technology is a “silver bullet” that solves all problems. This despite years of experience finding that (a) no technology solves all problems and (b) every new technology is eventually replaced by something else.

The Customer Data Platform industry is as prone to these errors as everyone else. The current silver bullet is composability, which is touted as the ultimate solution to problems of CDP cost and complexity. Just hearing the claim should lead anyone to realize it’s not true: composability solves some problems in some situations (i.e., unnecessary data movement at firms with sophisticated technical resources) but it won’t help companies that lack the necessary data infrastructure or staff resources. I suspect that even the most ardent composable enthusiasts would agree with this more nuanced view if you pinned them to a wall.

It's a step in the right direction to move past “composability is great” to discussing “when composability is the best solution.” But it’s also important to address the second flaw in the “silver bullet” fantasy, which is to assume that the best current technology is the best future technology. In other words, it’s worth asking, what comes next?

Remember, composability isn’t a solution: it’s an architecture. The fundamental argument for composable CDPs is that it’s better to have an architecture that uses the enterprise data warehouse as the primary store for customer data, rather than an architecture where customer data is assembled, stored and accessed in a separate CDP database. So the question about what’s next is really a question of what other architectures are possible.

I can think of two candidates.

• Read customer data directly from the source systems, without assembling it in any persistent database (i.e., neither a data warehouse nor a CDP). This vision of real-time, on-demand customer profile assembly is what many people thought a CDP should be back in the early days of the industry. At that time, ten to fifteen years ago, the existing technology simply couldn’t make it work. Many source systems didn’t allow real-time access, internal networks couldn’t handle the volume, and the processing costs to assemble profiles in real-time were too high. The problem pre-dates CDPs, of course: the need to make data accessible by pulling it from source systems into a separate database is why data warehouses were first created in the 1980s. There has certainly been progress in recent years, although I can’t say that the necessary technology is fully available. Still, the scope of what can be done in real time continues to expand, which diminishes the share of data that must be assembled in a warehouse. This, in turn, diminishes the relevance of the warehouse-centric composable architecture.

• Create customer digital twins. I’m highly skeptical that AI can accurately mimic customer behavior: this seems most likely to reinforce mediocrity by failing to predict unexpected behaviors (the most interesting kind) or reactions to novel situations (which includes the most important  innovations). But I can still imagine creating a digital twin for each customer, updated in real time with data captured across all company systems. The twins would be self-contained objects (or maybe agents) that can be queried any time a company system wants to find the best action for a particular customer in a given situation. While their role might resemble the customer profiles assembled in today’s data warehouses, I’m sure the technology will be quite different. I can't predict the specifics of that technology but am sure that clever people somewhere are already working on it. My only point here is that, again, the architecture won’t look like the warehouse-based composable CDP.

There are surely other possibilities that haven’t occurred to me. In fact, I’d offer better than 50/50 odds that the next CDP silver bullet will be something I haven’t listed. Placing the right bet matters a lot to investors and developers but not so much to users, who can wait to see what becomes available. 

What does matter to users is identifying their requirements.  This is something that debates about architecture only distract them from doing. So my fundamental recommendation is that CDP users think less about chasing the silver bullet and more about building the golden record – that is, how to create the accurate, unified, accessible customer data that a CDP using any architecture is intended to provide.

Fitting Agents into the Sales and Marketing Mix

Much has been written recently about how marketing and sales processes change when human buyers and sellers are replaced by buyer and seller agents: abbreviated, inevitably, as “A2A” marketing. It’s a fascinating topic but just one model that will coexist in the near future with human (or, more precisely, non-agentic) buyers interacting with agentic sellers, agentic buyers interacting with human sellers, and, lest we forget, humans interacting with humans. Any consultant will immediately recognize that this cries out for a 2x2 matrix, or perhaps a pair of 2x2 matrices if you want to distinguish business marketing from consumer marketing. For the moment, let’s stick with the single matrix model:

It’s worth making these admittedly-obvious distinctions because each situation raises separate issues, which are otherwise easily jumbled into a confusing heap. Let’s look at each situation in turn.

Human to Human (H2H)

Beyond the literal situation of one seller talking to one buyer, I’d argue this also includes humans interacting with traditional broadcast media, web search, and even non-agent websites. The common thread is that the human buyer does most of the work of asking questions and processing answers. The seller is largely reactive, although there are some situations where she makes choices such as selecting a personalized “next best action”, embedding dynamic content in a website, and setting up conventional search engine optimization. Those choices may be informed by predictive models or some other type of AI, but every step in the workflow is ultimately managed by humans, not agents.

I can’t point to specific data but am pretty sure that H2H interactions still account for the vast majority of today’s sales and marketing activity. This means that marketing and sales teams should still give significant amounts of attention to improving them, even though agentic interactions are vastly more fun to think about. If you absolutely must bring AI and agents into the picture, you can use them behind the scenes to speed up workflows, optimize performance, and analyze results.

Agentic Buyers to Human Sellers (A2H)

This is probably the situation that gets the most attention today. It includes true “buyer agents” (controlled directly by buyers) and “buyer-supporting” agents such as AI search engines and browsers. I call these “buyer-supporting” because they’re not controlled by the buyer, but instead by a company like OpenAI or Google which provides them to buyers at little or no cost.

The distinction matters because companies that offer “buyer-supporting” agents have their own agendas, which don’t necessarily align with the interests of actual buyers. In particular, these companies are increasingly interested in monetizing their products by serving ads within AI search and browser results. Some of these ads will be clearly labeled while others may be subtly embedded in the results themselves. These ads are an opportunity for marketers but may be problematic for users, who could be led to question the objectivity of the AI results.

Concern about biased AI search results could in turn lead to significant interest in true “buyer agents” that consumers pay for themselves. History suggests this will be an uphill battle: as we’ve seen with streaming video, large majorities of consumers typically chose free, ad-supported services over paid, ad-free subscriptions. Still, as streaming video has also shown, a significant fraction of consumers will pay for subscriptions in return for a better experience. This could be a large enough market to support a profitable business. Business buyers are even more likely to purchase agent subscriptions, since they don’t pay with their own money and can easily justify the expense based on better quality results. The precedent here is ad-supported versions of office productivity apps, which have never been broadly successful. There’s a chance that agents could be funded by charging advertisers for access to their owners, although such models have also failed in the past.

Advertising aside, most A2H discussions in martech and adtech circles focus on how sellers can adapt their systems to get the best results from buyer-side agents. This often involves advice on optimizing website design to accommodate search and browser agents, so a given brand receives the best possible treatment. Traditional SEO vendors are frantically expanding their products to meet this need and new AEO (AI Engine Optimization) specialists are also appearing. So far, the solutions are pretty basic: systems run sample queries to measure how often a given brand is mentioned in AI search results and vendors offer design tips to expose the kinds of data that AI agents are looking for. The next level is to look beyond measuring and influencing whether the brand is presented, to how it’s presented in terms of positioning and value. We’ll surely see more of that.

The thing to remember about “buyer-supporting” AI search and browser agents is they are generally driven by a big LLM model that draws from the same information for all users. True “buyer agents” would supplement the more-or-less static LLM models with custom research that visits seller websites to find answers to buyers’ specific questions. For example, one buyer might be interested in pricing details while another cares more about product quality. Beyond exposing all possible information, a seller might aim to present its product differently depending on what appear to be the buyer’s priorities. This is largely similar to today’s (non-agentic) website personalization. What’s more intriguing is the possibility that sellers can find a way to identify individual buyers’ agents over time, perhaps by requiring registration in exchange for detailed information. This would let the seller build a buyer profile and tailor responses to this profile. Piercing the buyer agents’ veil of anonymity would be hugely valuable.

There is a third situation: where the “H” in “A2H” is an actual human, not a non-agentic system. One current example is humans responding to agent-generated Requests for Proposals, which will likely be joined by other formats such as email inquiries or even telephone surveys. The growing volume of agent-generated requests is already a nightmare for business sellers faced with the cost of responding to them. The obvious solution is to let seller agents respond to the buyer agents, but it may be a while before most firms can deploy this capability. In the interim, sellers will be increasingly pressed to qualify buyers before deciding how to respond. Insofar as responding to qualification questions requires effort by the buyer, this imposes a cost on the buyer that should help to eliminate frivolous requests. At some point it might make sense for sellers to impose a literal cost – that is, to charge a fee – for responding to agent-generated sales queries. A less obvious concern is that buyers who rely on agent-generated research questions may fail to understand their true needs, removing a substantial portion of the value gained from a good purchasing project.

Human Buyers to Agentic Sellers (H2A)

Traditional websites may use AI-driven personalization but they are still non-agentic systems. In the future, we can expect true agentic interactions to become increasingly common. The best current example would be chat interfaces connected to an agentic back-end, enabling them to engage in true conversations with potential buyers. These have already evolved in some situations to full-scale agentic business development reps (who send those those super-annoying emails complementing your latest blog post and asking for an appointment) and sales reps (engaging in lengthy dialogs).  Agentic customer support reps are even more common and, often, better than humans at many tasks. While the distinction between AI-based and agent-based interactions can be vague, it’s fair to say that agentic interactions will be significantly more responsive to individual situations. This, in turn, makes them more reliant on capturing real-time data, both for customer behaviors and surrounding context.

Letting autonomous agents interact directly with customers raises major concerns about governance, output quality, and risk. These are widely recognized, as are the challenges of integrating agent-based systems with existing infrastructure. That being the case, I won’t rehash them here, apart from noting that they currently present substantial barriers to adoption of H2A models.

Agentic Buyers to Agentic Sellers (A2A)

Agents selling to other agents is the obvious endpoint of agentic adoption. It’s appealing if only for the amusing prospect of agents merrily jabbering with each other without any human involvement. But apart from a few highly structured interactions, such as programmatic advertising, it’s still largely in the future. A2A can’t become more common until the industry first solves the separate challenges of agentic buyers and agentic sellers. It must then overcome the additional challenges of connecting the two. Once the plumbing issues are addressed, there will be another level of adoption as buyers and sellers work to turn the interactions to their advantage. How will price negotiations work when buyers want the lowest price possible and sellers want the highest price? How will sellers discover the actual needs of buyers so they can make the best recommendations – and is what’s best for the seller necessarily what’s best for the buyer? How will seller agents decide which information to offer and which to exclude? How will agents build trust with each other? And how will companies manage the computing costs of agent-to-agent interactions, which could be substantial if the interactions are extensive?

Plenty of smart people are surely working through these issues. We already see some technical foundations being laid in protocols such as MCP and Google’s A2A. But it’s probably too soon for most marketers to put much energy into worrying about A2A deployment. Mastering the intermediate steps of A2H and H2A should come first and will put them in a better position to deal with A2A when the time is right.

Summary

The impact of AI in general, and agentic AI in particular, is overwhelming. While this piece offers some ideas and makes some prediction, my real goal is much simpler: to suggest that distinguishing the different types of human and agent interactions is a way to split the topic into smaller, more tractable pieces. I hope that helps.

Let's Debate CDP Functions, Not Definitions

What’s the definition of a CDP? It's a bad question because it diverts attention from what really matters: What capabilities do CDP users need? Still, buyers keep asking and sellers keep answering, typically in ways that promote their own interests. Looking for an unbiased perspective on the topic, I recently asked ChatGPT what answers it was finding. It came back with a reasonable cluster of responses and particularly interesting details on who was using each one (see below for the full response)*:

• Unified customer database: 70–90% of analyst pages, trade articles and vendor docs 
• Marketing activation / audience building platform: 60–85% of vendor docs, blogs and many press releases  
• Real-time/streaming profile & interaction engine: 30–60% depending on whether the source is vendor marketing (more likely) or neutral analyst articles (less likely to require “real-time”). 
• Privacy, governance & identity management layer: 15–35%, increasingly present in analyst pieces and vendor positioning  
• Part of a larger ‘data cloud’ or enterprise data stack: 10–30%, especially in vendor/marketing copy from big cloud vendors

These are categories that ChatGPT identified without me defining them in advance.  So it's particularly interesting that there’s no mention of composability, warehouse-centricity, no-copy, hybrid, embedded, integrated, stand-alone, or other architectural details that have dominated recent industry discussions. In a way, this represents a failure by the CDP Institute to propagate our view that a CDP must build a separate database of its own. But a less parochial response is to be pleased that the main distinctions reflect system functions, which is where the focus belongs.

Of course, the variety of definitions is still problematic. While it’s usually safe to assume that a system labeled as a CDP will provide a unified customer database, it’s less certain that it will also offer marketing activation and downright dicey as to whether it will offer real-time streaming profiles and interactions. This means the label provides little useful guidance: imagine a can of soup labelled “contains tomatoes and maybe chicken and could also have mushrooms, rice or shrimp”. The only way to know what's inside would be to open it -- which defeats the purpose of a label.  

(And, yes, the problem is worse for CDP than other categories. When I ran the same prompt for the term "customer relationship management software," a single answer dominated: 71% defined CRM as “a software/system/platform to manage customer interactions and data.”  The next highest share was just 29% for “integrated suite (sales, marketing, service automation)”. It’s true that the dominant answer is exceptionally broad, but at least most people understand this and won’t expect anything more specific.)

So, although industry understanding has not been entirely destroyed by architectural debates, there is still enough disagreement on the scope of a CDP to limit the term’s utility. (If the CRM example is typical, it may be a natural progression for popular categories to expand their meaning over time. That would be an interesting hypothesis to explore if anyone out there is looking for a thesis topic.)

The industry could fight to restore a more specific CDP definition, but that’s probably a losing battle. It’s more likely productive to shift the discussion away from defining the term "CDP" to defining the functions required to manage customer data. 

Yes, I’m proposing that the solution to our problem is a checklist. Don’t roll your eyes: whole books  have been written on the topic. (Ok, maybe just one book.)  But in an industry that has long been driven more by theory than practical requirements, anything that gets buyers to focus on what they actually need is a win.  

Getting the industry to agree on a shared requirements checklist wouldn’t be easy, since every participant would want to add or remove items depending on whether their products supports them. Indeed, the very notion of a comprehensive requirements list favors broad, integrated products over narrow point solutions. But I’d still invest a few embers of hope in a project to forge a complete customer data framework. The potential benefits, for users and vendors alike, are well worth the effort.

Why Agentic Campaigns Are Not The Future of Marketing

My last post, presenting ChatGPT’s description of AI-based marketing, generated thought-provoking comments on LinkedIn. Key insights included:   

• Organization issues as roadblocks to successful deployment (Chris Adelman and Lee Hammond)
• Buyer agents as increasing the importance of trusted information sources (Jon Miller, Dan Cote)
• Some vendors are already delivering end-to-end agentic campaigns (Matthew Niederberger, Duarte Garrido)
  

Thanks to everyone who contributed.

What I didn’t see was push-back against the vision that ChatGPT presented. Maybe people were being polite or maybe they fear that ChatGPT could hold a grudge. But my own initial take was triggered by words including “narrative” and “journey”. These led me to think that ChatGPT was still describing today's standard of multi-step marketing campaigns applied to customer segments. On closer examination, the story is a bit more complicated: the various components included in ChatGPT’s list are not necessarily a consistent whole, but rather a collection of predictions that ChatGPT has gathered by scraping of industry discussions. Oddly enough, ChatGPT's summary provides a clearer statement of an alternative approach:

AI-native marketing is:

• Continuous (not campaign-based)
• Conversational (not broadcast)
• Collaborative (AI agents on both sides)
• Generative (creating narratives, products, and experiences dynamically)
• Ethically-audited and explainable (trust is as important as persuasion)

Of course, that ChatGPT says something is no guarantee that it’s correct. So, although I intuitively agree with the prediction, I still need to convince myself (and, hopefully, others) that it’s right. One test is to compare the prediction with cutting-edge industry developments. Here’s a list of trends I’ve been tracking:

• Agentic search: AI-generated summaries replace traditional, link-based search results
• Ads in chat replies: Chat responses include paid advertising, identified with varying degrees of transparency
• Social commerce (a.k.a."live shopping"): Make purchases of items featured in social media discussions without leaving the social media platform
• Commerce via agentic browsers: Advertising and shopping links presented by browsers based on what they see the user doing (a privacy nightmare but we'll ignore that for now)
• Interactive ads: Advertisements that link to a conversational interface rather than conventional landing pages or websites
• Buyer agents: AI agents that do product research and, perhaps, make purchases on behalf of users without users visiting seller websites
• Conversational websites: websites that present chat-style interactions or hyper-personalized contents in dialog with the user

All but the last of these occur at the discovery stage of the buyer cycle. That's encouraging because marketers have been very concerned about they reach new customers when traditional search no longer drives traffic to their websites and consumers increasingly ignore traditional email and display messages. But what's really important is that these items all change the role of marketing from creating campaigns that broadcast messages to consumers, to executing continuous conversations with consumers. This is indeed what ChatGPT described. It’s not something that is accomplished by using AI generate conventional broadcast campaigns.

In fact, I’d argue that the current broadcast campaign model developed precisely because the then-existing technology couldn’t handle a conversational approach. Branching campaigns with personalized content nodes were the best that could be done when marketers were limited to deterministic, manually-crafted workflows. If you recall the once-popular cliché of marketing automation as a way to replicate the personal relationship of customers to their “corner grocer,” you quickly realize the real-life grocer had conversations, not campaigns.

Think a bit deeper and you see that the reason the grocer was able to have conversations was data. He knew the history of each customer and was able to discern their current situation by seeing and talking to them. Multi-step marketing funnels exist because marketers don’t have that kind of information available in real time. Instead, they have to make a guess at where each customer might be in the buying cycle. The funnel is a framework for making those guesses. Provide real information and it’s no longer needed.

I think this is important. Marketers, vendors and analysts have focused largely on how AI can reduce the cost and time of generating marketing campaigns. This allows companies to create near-infinite amounts of personalized content. But unless that content is driven by near-complete, real-time data, there’s a severe limit to how relevant it can be. Data, not content creation, is what ultimately limits the effectiveness of AI marketing.

This may strike you as bad news. Data has always been a problem for marketers, and real-time data is a particular challenge. Agentic campaign creation won’t solve the data problem, although perhaps it will highlight data's role as a bottleneck. The good news is that other AI and agent applications can in fact improve data access, something that data management vendors are already pursuing. 

It’s true that better data management isn’t as exciting, or directly tied to revenue, as better customer engagement tools. This is probably why most vendors focus on adding engagement features. But as autonomous engagement become increasingly common, the need for better data will become more obvious. Vendors who use AI to deliver better data (and convince the market that they are leaders in doing it) will have an increasing advantage in growing their business.

I’ll give the final word to Abhi Yadav of iCustomer, one of those very smart people who see further, sooner than most of us. He recently shared a substack post that captures my opinion quite eloquently:

We used to drag customers to our properties (site, app, landing pages). Now we take capabilities to where customers already live: agent/app ecosystems, workflow-native GPTs, and search/ad surfaces that embed decisions at the edge. ... If your data stays scattered and your channels siloed while this shift accelerates, AI won’t transform your business it’ll just perform visual eye candy. AI should drive a paradigm shift in your business, not a party trick.

What would an AI-native marketing process look like? ChatGPT's answer is better than I expected

If you look at recent vendor announcements, it seems that the cutting edge in applying AI to marketing is 'end to end' campaign generation, where a collection of agents work together to build and execute a complete marketing campaign with no human effort (although, for practical purposes, still under human supervision).  As I've written before, this strikes me as an advanced type of 'substitution', where the new tech is used to execute the old processes better and faster.  That's fine and valuable but it's only the first step in taking advantage of the new technology.  The final step is to develop new, "transformed" processes that are optimzed to make full use of the new tech's capability.  

I've recently been asking myself and others what they think that "transformed" process would look like for marketing.  I'm still collecting my thoughts but, as part of the process, asked ChatGPT for its opinion.  The answer was better than I expected.  I don't necessarily agree with all of it but thought it was interesting enough to present to the larger world for comment.  So, please, let me know what you think.

Prompt: if marketing techniques were designed from scratch to take full advantage of AI, what would they look like?

That’s an excellent and forward-thinking question — and the answer’s pretty transformative.

If we were to design marketing from the ground up for AI (instead of retrofitting AI into traditional marketing), we’d end up with a system that looks less like “campaigns” and more like continuously learning ecosystems that respond in real time to individuals and social dynamics.

Here’s a breakdown of what that might look like:

---

1. Autonomous, Adaptive Market Intelligence

Instead of periodic market research and segmentation,
AI-first marketing would run on constant sensing loops:

• AI agents continuously scan social media, purchase data, and search patterns.
• They detect cultural shifts, emerging memes, and micro-trends as they form.
• Personas aren’t static — AI builds and dissolves them dynamically based on live data.

Analogy: You don’t define “target audiences.” You discover living patterns of attention in real time.

---

2. Hyper-Personalized Narrative Engines

Instead of a single ad creative for millions of viewers,
AI-first marketing generates unique narrative journeys per person:

• Language models craft individualized story arcs and product framing that match someone’s identity, tone preferences, and current emotional state.
• Multimodal models (text, image, video, audio) assemble personalized content on the fly — like an AI “showrunner” creating micro-stories for each user.
• Brand consistency is enforced by a central “brand DNA model” ensuring all outputs feel coherent and on-brand, even when autonomously generated.

---

3. Predictive and Generative Commerce

Instead of advertising then selling,
AI-first marketing anticipates needs and builds products or bundles before people articulate them:

• Predictive demand modeling uses behavioral and environmental data to suggest what products should exist.
• AI may prototype and A/B-test new offerings via virtual storefronts, gauging interest before production.
• Marketing and product design merge — you market possibilities before they exist, and only produce what the market proves it wants.

---

4. Multi-Agent Relationship Systems

Instead of one-way brand broadcasts,
AI-first marketing uses relationship agents that represent both consumers and brands:

• Consumers might have personal “shopping AIs” negotiating with brand AIs.
• Trust, transparency, and reputation become measurable assets traded between AIs.
• The marketing goal shifts from persuasion to mutual optimization — aligning a person’s needs and a company’s offerings through agent collaboration.

---

5. Real-Time Ethics and Trust Architecture

AI-driven marketing would require built-in ethical governors:

• Explainability modules ensure users understand why they’re being shown something.
• Consent becomes dynamic and contextual (you can allow certain data uses for limited time).
• Bias-detection systems continuously audit generative campaigns for fairness and cultural sensitivity.

---

6. Emotionally Intelligent Interaction

Instead of static messaging,
AI-first marketing converses, empathizes, and co-creates:

• Emotional state detection (via tone, text, or visual cues) guides responses.
• The AI adjusts its persuasive approach — sometimes offering information, sometimes humor, sometimes retreat — like a human skilled in timing and empathy.
• Brands become more like adaptive characters than faceless institutions.

---

7. Closed Feedback Loop: Learn → Generate → Test → Learn

Every interaction feeds a continuous reinforcement loop:

• Each ad, email, or product suggestion is an experiment.
• The system measures micro-responses (hover time, tone change, conversion intent).
• Those signals retrain the creative and strategy models daily.

---

In summary

AI-native marketing is:

• Continuous (not campaign-based)
• Conversational (not broadcast)
• Collaborative (AI agents on both sides)
• Generative (creating narratives, products, and experiences dynamically)
• Ethically-audited and explainable (trust is as important as persuasion)

 

How Will GenAI Change the Consulting Industry?

Reports on the threat that generative AI poses to knowledge workers usually focus what might be called “structured creativity”: tasks like copywriting or computer programming where outputs are clearly specified through things like creative briefs or system requirements. One group that generally doesn’t consider itself at high risk are the analysts and consultants who write the reports themselves. After all, their work requires insights and judgements that are beyond the current capabilities of genAI, which ultimately does nothing more than play back the training data it's fed.

Case in point: This widely reported BCG study found that genAI tools substantially improved performance of consultants at product planning, with the greatest improvements among below-average performers. A second finding of the same study was less widely noted: AI depressed performance at generating insights from interviews and raw data. This dovetails nicely with the idea that AI can help with structured tasks such as campaign management but isn’t (yet) capable of developing new insights. One reason that AI struggles with insights is its well-known tendency to hallucinate, which would translate into false conclusions about the meaning of input data. (See, for example this BBC study that found 47% of AI chatbot answers contained significant errors about data sources or accuracy.)*

So it seems that the survey analyzers are indeed safe, at least for the time being. But while insights are an important part of the value provided by analysts and consultants, they’re just a small fraction of their work. In my experience, the main reasons that companies hire consultants are to access their skills at organizing a project and their knowledge of vendors to consider hiring.** Those are things that genAI does extremely well: ChatGPT will gladly provide a plausible project plan for nearly any task and will just as merrily answer your questions about which vendors to consider, providing detailed comparisons on relevant features. It's true that those comparisons will often be inaccurate (see the BBC study or do your own test in a field you know well). But, let’s face it, consultants also make mistakes about product capabilities, which change much faster than even the most diligent industry expert can keep up. So, whether a vendor list comes from genAI or a human consultant, it still has to be validated with current information before you base any decisions on it.

Even more to the point: just as it’s always been hard for clients to tell truly knowledgeable consultants from pretenders, it’s now equally hard to tell accurate genAI outputs from hallucinations. This applies to frameworks, vendor lists, and even those insights where humans are still clearly superior. GenAI is famously capable of providing reasonable-sounding arguments based on false premises, citing false evidence, and leading to wrong conclusions. A true optimist would argue those can lead to fruitful discussions even when they start with an error. A more realistic expectation is that some large fraction of people will accept them as delivered. 

This is especially true when it comes to comments on industry trends and futures, where there is no clearly right answer. A large part of the industry conversation today, especially in media like LinkedIn, is exactly that sort of discussion. AI-generated comments are increasingly common and, truth be told, can be hard to tell from the human-generated ones. Sure, that's a tough situation for industry members trying to get good information. But save some sympathy for the poor human consultants who are trying to be heard – and recognized as legitimate experts – amid all the noise. 

Indeed, the flood of content is arguably the biggest change in marketing in recent years. Not so long ago, publishing good content on a regular basis was an effective, low-budget path to success. The mechanism was simple: good content was picked up via social media sharing and, if it became popular enough, registered in search engine results. Today, not only is there too much content for anyone who’s not already famous to gain a significant share of attention, but AI search summaries will scoop up good information without leading readers back to the source. The alternative path of publishing on high-traffic news sites is equally blocked since AI search engines stifle traffic to those sites as well. Of course, a handful of influencers who are skilled at attracting attention can still become Internet-famous.  But what gains attention isn’t necessarily what’s of true value.***

I’ll skip the usual moaning about bad information driving out the good, since the consequences are all around us. Instead, I’ll posit that the situation is untenable because people simply will not be willing to consume such huge volumes of content indefinitely. There is already a backlash among people who want to reduce the time they spend trying to keep up. (In fact, time spent on social media peaked in 2022.)  The obvious solution is to let AI view media for you and summarize what’s worth seeing. Some people may find that acceptable but I’d like to think that most will recognize that AI can’t be trusted to make the right choices – especially if the AI is run by corporations with interests of their own. The alternative would seem to be relying on human curators – remember “editors?” -- to assess the available materials and select those worth consuming. Of course, deciding which curators to trust will itself require consumers to do some work. This path also requires an alternative to the failing economics of traditional, ad-supported media, which will probably be consumer-paid subscriptions that compensate both the content creators and the curators. A middle ground might be AI curators that are funded by consumers rather than self-interested corporate providers.

I won’t pretend to know approach will dominate. Perhaps different groups will adopt different approaches, with access to accurate information becoming a luxury good like ad-free TV. For now, people can already be selective in their content consumption by avoiding the undifferentiated (or, more precisely, platform-selected) streams in news and social media feeds. The dilemma is sharper for experts relying on content creation to build their business: do they try to keep up with content bots that can easily outproduce them, recognizing that even consistently brilliant, unique, and insightful content won’t necessarily be noticed? Or do they find other ways to reach prospective clients, perhaps through live events or less mediated platforms like reddit and medium?

Mind you, this is occurring in a world where AI can legitimately replace some of the expertise consultants are used to selling, such as methodologies, technical skills, and vendor knowledge. This leads to smaller engagements where clients are buying only the AI oversight and creative insights that humans alone can provide. So does the total market for consulting services shrink at exactly the same time that selling those services becomes harder? Or, more optimistically, does lower cost enable consultants to expand the market by selling to firms that couldn’t previously afford their services? (See, for example, this new WPP offering of AI-enabled “self-service” ad campaign development.) While AI-powered services don’t themselves create more human jobs, they do presumably offer an opportunity to sell insight-based projects that do require paying humans.

I'll say it again: the future is unknown. But I do believe that analysts who are calmly assessing the threat that AI poses to other people’s jobs might give more urgent thought to the threat AI poses to their own. 

Taking Advantage of AI Search

The past few weeks has seen a flurry – more of a blizzard, really – of announcements from Search Engine Optimization companies that have added tools for AI search ranking. This is a natural outcome of the meteoric growth in AI search: referrals from ChatGPT grew 20% from June to July this year while ChatGPT agent visits doubled in the same period. 

Before you get too excited, bear in mind that LLM vendors’ share of global search is still a minuscule: 0.13% so far in 2025 (for ChatGPT and Perplexity combined). Given that Google still has nearly 90% of all search traffic, it’s much more important that more than half of Google searches now return an AI overview.

Probably the most important statistic is that organic click-through rates drop by 20% to 40% when AI Overviews are displayed.  One important caveat is that AI overviews show up primarily on “informational” queries, as opposed to commercial, transactional, or navigational queries. The caveat to the caveat is that informational queries are by far the most common: 88.1 % of the total, compared with 8.7% commercial, 1.4% transactional and 1.8% navigational. So, if your web strategy has been based on providing information to attract visitors, you have some adjusting to do.

Or do you? Most AI optimization tools replicate SEO by focusing on how often a brand appears in AI search results. This Evertune press release provides a detailed explanation of how their product works, which is probably typical: the system runs a set of common prompts on a topic and tracks how often the target brand is mentioned. That’s very similar to traditional SEO, which reports how the target brand ranks in search results. In both cases, the goal is to be seen more often.

It's true that tactics to reach those goals will differ. Traditional SEO is based on links and keywords, while AI search engines care more about detailed information and reviews. More broadly, content that attracts humans will often be insightful or entertaining, while bots respond best to clearly presented data. So some changes are needed if you want to switch your goal from SEO to AI search ranking.

The deeper question is whether appearing in AI search results is the right goal. If you’re a brand whose aim is building awareness and credibility, the answer is probably yes – so long as the AI results present you in a positive light. It will be even more important as search bots morph into shopper bots, which recommend (and will eventually execute) purchases on behalf of human buyers. Shopper bots will no doubt engage with brand-owned seller bots, a category (or e-commerce system feature) that I can't recall seeing but is surely gestating as you read this.

The answer is different if you’re a publisher. There’s little value in having your materials cited in an AI response that doesn’t generate a click-through to your site. In fact, showing your materials gives users less reason to visit. Adding injury to insult, responding to the AI search engine’s crawler actually adds to your website operation costs. Put simply, publishers need traffic.  Visitors provide inventory for ad impressions and names to capture for future promotions. The drop in organic traffic forces publishers to seek other traffic sources, including both paid search and social forums such as reddit, Instagram, and YouTube.

• Paid search itself will evolve along with AI search. Ads already appear alongside Google and Microsoft AI search results. Looking ahead, Google Research recently published a paper describing how advertisers could bid on the text of the responses themselves. There are also other ways to reward AI search engines for driving traffic to advertisers, such as affiliate relationships. Consumers won’t be pleased when paid and unpaid recommendations mingle indistinguishably in search results, but it’s not clear they’ll have much choice in the matter. The problem facing publishers is that many won’t be able to afford the cost of paid advertising or will need to reengineer their business to generate revenues to cover the expense.  

• Social traffic is a more appealing option for publishers since it’s “free” to the degree they can generate it. That should be possible: after all, publishers' core competency is to create appealing content. One implication is a growing overlap between traditional publishers and influencers, a distinction that has already become quite blurry. It’s worth noting that we’ve recently seen substantial growth in software and services to support influencer marketing, including tools to find influencers, build audiences, and streamline content creation, as well as ad networks and media buying tools to simplify ad placement on influencer sites. Most have been aimed at individual influencers, but enterprise-grade products will probably appear with publishers as the target clients. Social marketing will increasingly look like SEO, except aimed at optimizing content for social algorithms rather than search algorithms. Since SEO companies and social media management vendors both have a legitimate claim to this turf, it will be interesting to see who ends up owning it.

In short, we’re just beginning to see the changes resulting from AI search. It has already disrupted the business model of many web publishers, who no longer receive the organic search traffic that powered their audience creation. Traditional SEO will be supplemented by AI SEO, especially for brands who can benefit from appearing in AI search results. But while traditional search generated large volumes of free traffic, AI search results will likely be dominated by paid placements. Publishers, who gain less than brands from AI search ranking, will need to find alternative sources of traffic and will increasingly resemble influencer business models. Above all, we’ll see continued development of supporting technologies (“infrastructure” in terms of the STIB model) as entrepreneurs identify new opportunities to help companies operate in the new environment. Ultimately, again according to STIB, we’ll start to see new business models emerge that are best suited to the dynamics of an AI search-based economy.