Share of Model

AI search is already here—it’s just not evenly distributed. ChatGPT’s prompt volume is already roughly 18% of Google’s daily searches, and AI Mode queries are doubling every quarter. McKinsey found 44% of AI-search users now treat it as their primary source of insight—ahead of traditional search at 31%. For B2B it’s starker: 51% of software buyers start purchase research in a chatbot, not a search engine.

The fragmentation is the strategic point, not just the growth. Ultimately, this is a half-dozen new channels that don’t all agree with each other. Only 6–8% of URLs ChatGPT cites also rank in Google’s top ten. After decades of search being winner-take-all, this feels chaotic.

SEO was a game of position: There was a list, you had a slot on it, and the slot was objective—everyone searching the same words saw the same ten links. AI answers don’t work like that. What replaces the search rankings page is a set of preferences, and increasingly memories, the model has about the user that helps inform the unique ways it chooses to represent your brand and products to users. This is even true for what we would think of in search as branded terms.

We watched a brand that ranks #1 on Google for its main commercial terms show up in AI answers with a warning attached—effectively, “don’t buy from these people.” The complaints came from Reddit threads about pricing, baked into training data. Peer-reviewed audits confirm this is systematic: models carry trained-in brand preferences before any web search happens.

And look at who actually shapes the preference. It’s estimated that in 2026, Wikipedia is 13.2% of ChatGPT citations and Reddit is 12%; the Wall Street Journal and the New York Times don’t crack the top twenty. 63% of all AI citations point to listicles—comparison content is the new shelf space. And while all that’s true today, there’s no guarantee it will still be true tomorrow.

To influence the preference you have to know where it lives. Models have two memories: the weights—what they absorbed in training, updated every few months at best—and retrieval, what they look up in the moment.

The useful frame here is pace layers, a Stewart Brand idea that Alephic has written about applying to AI work: complex systems have fast layers that absorb shocks and slow layers that hold the structure. AI search is exactly this. Mentions and citations move daily. Your owned website moves weekly. Core search rankings—which still power most AI retrieval—move over months. And the training run that decides what the model believes about you moves over quarters and years.

Most GEO tactics on sale today only touch the fastest layer. Self-serving listicles work right now because retrieval weights them heavily—which is why it’s a short-term game with a short-term prize. The arbitrage gets crushed the moment engines pre-compute answers and filter self-interested sources, the way Google has spent twenty years learning to do. Worse, flooding the fast layer with AI slop poisons the slow one: it damages how you’re represented in the next training run, and recovery operates on training-run time, not sprint time.

The fast layers are for experiments; the slow layers are for strategy. A plan rooted in principles and fundamentals—real reputation, real evidence, real infrastructure—keeps paying as it sinks down the layers. A plan built on chasing listicles because they work today expires with the next model update.

For thirty years the web ran on a simple trade: you let crawlers read your content, and search engines sent you visitors. Google still crawls about five pages per visitor it refers back. OpenAI crawls 857. Anthropic, more than eleven thousand. The traffic isn’t coming back—AI clickthrough fell from 0.8% to 0.27% in 2025, and publishers expect another 43% drop by 2029.

The web’s response has been to wall itself off.

And the walls now have infrastructure. Cloudflare, in front of roughly a fifth of the web, now asks every new domain at signup whether to allow or deny AI crawlers, and HTTP 402 “Payment Required” became live pay-per-crawl plumbing. Where traffic stopped flowing, payments started: 91 content licensing deals and counting—News Corp–OpenAI near $250M, Reddit–Google at $60M a year—plus Anthropic’s $1.5B authors settlement putting a price on unlicensed training data. A real market for machine-readable information is forming, with prices.

Underneath the plumbing, this is a disintermediation story—the one that gutted publishers, now running for everyone else. The assistant sits between you and your customer, keeps the relationship, and hands back a transaction at best: the sale without the loyalty or the insight. You don’t have to call yourself a publisher for this to reach you. If your economics depend on owning an audience, it already has.

Which makes every brand a publisher with a data posture, whether it wants one or not. Block, license, or feed—but choose deliberately. Your robots.txt is a commercial term now, not an IT setting. And blocking doesn’t even keep you out of AI answers; it just removes your voice from how you’re described.

Agentic commerce is already here. AI influenced $262B of global online holiday sales—a fifth of the total. Shopify orders from AI searches grew 13× in Q1. And AI-referred traffic now converts 42% better than everything else—a full reversal from 2025, because the model already did the comparison shopping.

The rails underneath are still trying to catch up—two rival protocols fought over in standards bodies and in court, the map redrawn three times in eighteen months. Watch Walmart’s path through it:

The protocol wars · 18 months

Sept 2025

OpenAI launches Instant Checkout

The Agentic Commerce Protocol (ACP), built with Stripe. The buy button moves inside ChatGPT.

Oct 2025

Walmart signs on

The world’s biggest retailer agrees to sell inside the chat.

Jan 2026

Google and Shopify answer with UCP

The Universal Commerce Protocol, announced at NRF with 20+ retailers. The retailer stays merchant of record and keeps the data. Amazon, Meta, Microsoft—even Stripe—join the council.

Mar 2026

OpenAI walks checkout back

Discovery stays in ChatGPT; checkout goes home to the merchants. Target, Sephora, Best Buy, and Home Depot stay aboard.

Mar 2026

Walmart pulls out—and returns on its own terms

Native yield ran 3× worse than its own properties, so Walmart left Instant Checkout—and embedded its own agent, Sparky, inside ChatGPT instead. Rent the audience. Keep the relationship.

Two protocols, one full reversal, eighteen months. Anyone selling you a definitive agentic-commerce playbook is guessing.

Amazon, meanwhile, blocked ~47 AI agents and sued Perplexity rather than integrate—while its own assistant drove roughly $12B in incremental sales inside its walls. Nobody knows where this lands yet—not OpenAI, not Google, not Walmart. It is early. The wrong move is waiting for a winner; the right one is staying light—options open on every protocol, the customer relationship held where you can defend it.

AI drives an explosion in customer preferences—through personalization and longer prompts. The funnel changes shape: the middle—comparing, shortlisting, checking reviews—disappears into the model, and what comes out is a decision. And in agentic commerce, agents don’t run keyword searches; they match catalogs against preferences no one would type into a search box—budgets, brand vetoes, ethical lines, fit. Discovery becomes personalization. Being readable gets you considered; being described richly enough to match a real preference gets you bought.

SEO was about clicks and performance. AI search is about opinion. The model has a view of your brand, and every interaction puts it into words—describing you, in its own words, to every buyer who asks. The question stopped being do you rank and became what are you like.

It’s also a conversation you never see. The buyer asks, the model answers, a view of your brand forms—and nothing reaches you. No click, no visit, no line in the logs. Search was the one place a brand impression came with a number attached. Now the impression happens and the number never does.

It changes the day job too. SEO came with a free map of demand: every keyword your buyers typed, with volumes attached. GEO deletes the map. Prompts are 23-word briefs, never typed twice, invisible to you. There is no Ahrefs for prompts. When you can’t study the queries, you have to study the customer.

This is the oldest job in marketing: awareness, preference, reputation. How Brands Grow calls it mental availability—not whether buyers know your name, but whether you come to mind in a buying situation. A prompt is a buying situation, typed out in full.

It’s already costing deals: 69% of B2B buyers chose a different vendor than they planned based on AI guidance, and a third bought from one they’d never heard of. The shortlist is settled in a conversation you can’t see, before your pipeline knows the deal exists. And most enterprise brands are in the worst spot: perfectly crawlable, never shortlisted.

So you navigate by the two things brand marketers have always owned: what your customers actually need, and what you actually stand for. Not keyword lists—need states, the situations that send someone asking. Not rankings—your attributes, and whether the model repeats them back.

Citation monitoring is table stakes, but it doesn’t tell you much on its own—ask a model the same prompt twice and the brand list barely repeats. We believe serious brands should be investing in strategy from first principles, backed by first-party research and data. And innovating and experimenting in playgrounds and prototypes.

That’s the work Alephic is built for: strategy, search, and AI in one team—the combination AI search actually demands.

Know your customers, know the models, and connect the two.

In practice, that’s custom explorations and working systems:

Trusted by teams at

Strategy in times of uncertainty requires adaptation. We embed with your team, ship the first working system in weeks, and leave you with infrastructure that compounds—the forward-deployed way we work on everything.