Two-Year AI Retrospective
AI changed search and buying. Here's what I got right - and wrong.
In January 2024 I published a two-part prediction that AI would change how people search and buy, forever - before "zero-click" was even a phrase people used.
That was five weeks before Gartner's own headline prediction on the same shift - the moment most of the industry marks as when this conversation started.
The original prediction
Two posts, January 2024.
Pinned as Featured on LinkedIn ever since. Here's the original wording, unedited - not a paraphrase of what I remember saying.
"AI functionality could even make calls on the users behalf and book an appointment. No website visited by the user."
"...being quicker to respond to change than the competition could not only safeguard your existing growth, but also accelerate your business further."Read the original on LinkedIn →
"...future websites may be less focused on experience and more about knowledge-bases and databases with semantic content."
"headless-CMS could provide the hybrid solution... API gateways will be set-up from company websites as dedicated entry points for AI to retrieve specific content elements."
"AI will favour solutions from brands with strong review credentials."Read the original on LinkedIn →
This retrospective is also on LinkedIn as a 12-slide carousel, with the full source list in the comments - see it here →
AI-assisted, human-led: I briefed this, reviewed every word, and stand behind it as if I'd typed it myself. Full detail ↓
Why revisit it now
Two years is long enough to find out.
Most predictions about AI age badly in one direction or the other - either overtaken by how fast things moved, or quietly forgotten when they didn't land. This one deserved an honest check against what actually happened, not a victory lap. So I went back through the research properly: what the original posts actually said, what the market did instead, and where the two differ.
The verdict
Six claims, checked against what happened.
Net: four of six claims weren't just right, they were ahead of the market by 12-18 months - genuinely rare for a two-year-old AI prediction. The two corrections are specific and instructive, not embarrassing, and as you'll see below, they both point the same direction.
The boldest claim, tested
"No website visited by the user" - what actually happened.
The clearest test of the original claim came from OpenAI's Instant Checkout, launched in September 2025 - full in-chat purchasing, the purest version of "AI books the appointment." Walmart put around 200,000 products through it starting that November.
The result: in-chat purchases converted at roughly a third of the rate of clicking through to walmart.com. Walmart's own EVP of product called the experience "unsatisfying." Two structural gaps explained most of the drop - no basket (single-item purchases only), and narrow payment options at launch.
By March 2026, OpenAI had pulled back from in-chat checkout entirely, shifting toward a model where the merchant owns the transaction. Walmart's response was its own agent, embedded inside ChatGPT, that handles discovery in the chat but completes the purchase inside Walmart's own environment - Walmart login, Walmart cart, Walmart payment. It's now converting at around 70% of walmart.com's own rate - more than double what Instant Checkout managed.
So the honest update isn't "still early, give it time." It's that OpenAI itself concluded the same thing this page is arguing: AI can own the conversation, but the merchant should own the transaction. That's a sharper, more current version of "no website visited" than the original claim - and it lands closer to the real lesson than a straight prediction of full bypass ever did.
The architecture, corrected
Why headless CMS was the right idea in the wrong place.
A headless CMS was, and still is, a genuinely workable API layer for a meaningful slice of the market - content-led, service, or B2B businesses without deep real-time inventory or multi-market pricing to manage. Where it breaks down is genuine SKU-scale retail, where a PIM or ERP exists precisely because it specialises in exactly that operational complexity.
The more interesting question is why the wider architecture didn't spread faster, even though the technology was always capable of it. Two things compounded:
Wrong buying centre. Headless CMS gets bought to solve a marketing problem - faster publishing, one system feeding web, app and kiosk. IT shows up for implementation and security sign-off, not to frame the opportunity. The people who could see where this was heading - IT, at the technical conferences - weren't the ones holding the budget or the business case.
No proof point yet. Even if the right people had been in the room, there was no ChatGPT shopping, no AI Overviews, nothing to point to in 2022-23 that would have won the internal argument. It wasn't one gap, it was two, compounding each other.
Consistent with this: the infrastructure that actually won - Google Merchant Center, OpenAI's product-feed APIs, the emerging agentic commerce protocols - wasn't built or sold by CMS vendors extending a marketing tool. It came from AI and search platforms engaging retailers' commerce and data teams directly. A different buying centre entirely.
Ahead of market, not just early
Predating the benchmark everyone else cites.
Gartner's widely-cited industry prediction on this exact shift - that traditional search volume would drop 25% by 2026, with search marketing losing ground to AI chatbots - was published on 19 February 2024, in its Predicts 2024: How GenAI Will Reshape Tech Marketing report.
8-12 January 2024 - original two-part series published.
19 February 2024 - Gartner publishes the prediction most of the industry now cites as the moment this conversation started in earnest.
That's roughly five weeks. Not just early - ahead of the specific, most-referenced mainstream source on the same claim.
Worth being straight about the limit here too: this comparison holds for the broad "AI will reduce search traffic" claim, which is well documented. No equivalent 2024 analyst source was found making the narrower, more technical claim - headless CMS with dedicated API gateways as the delivery mechanism. That one appears to have been a more original, rarer thesis at the time, not independently corroborated elsewhere - which is a more honest way to put it than claiming credit for something no one else was measuring.
Where this leaves it
It was never really a CMS story.
The durable advantage was never "headless CMS" as a category, and it was never really about the website layer at all. It's ownership of a governed, canonical, API-exposed data layer - most often a PIM, ERP and DAM working together - with any CMS as just one consumer of that data among several, alongside search AI, commerce platforms, and the emerging multi-protocol standards now being built for agentic commerce.
AI discovery has moved upstream to structured feeds. AI-completed transactions are still an open, maturing question rather than a settled one - separable from discovery, and worth watching as the product layer keeps developing. But the businesses ahead right now aren't the ones who bought the trendiest frontend. They're the ones who treated their product data as a governed, single source of truth years before anyone made them.
Evidence trail
Sources for the claims on this page.
Every dated or quantified claim above is checkable. Closed by default - open it if you want to verify any of it yourself, which is rather the point.
Show sources (6)
- Gartner's 25%-by-2026 search-volume prediction, published 19 February 2024. Gartner Newsroom, 19 Feb 2024 →
- OpenAI's Instant Checkout launch and Walmart's ~200,000-product rollout, from November 2025. Search Engine Land, 19 Mar 2026 →
- In-chat purchases converting at roughly a third of Walmart.com's own rate; Daniel Danker's "unsatisfying" quote. MarTech, 20 Mar 2026 →
- OpenAI's pivot away from in-chat checkout toward merchant-controlled models, and Walmart's Sparky rollout inside ChatGPT and Gemini. The Paypers, 24 Mar 2026 →
- Sparky-in-ChatGPT converting at roughly 70% of Walmart.com's rate. Caveat: this figure is widely repeated across industry coverage, but every instance traces back to unnamed "early" or "internal" data rather than an on-record company statement - treat it as directionally reported, not confirmed to the same standard as the figures above. Industry coverage, Apr 2026 →
- Original two-part LinkedIn series this retrospective is checked against. Post 1, 8 Jan 2024 → · Post 2, 12 Jan 2024 →
How AI was used
This page was AI-assisted - I briefed it, it wrote.
I lead AI with my wisdom, not the other way around. Nothing reaches you unread: I review, verify and take responsibility for every word, the same as if I'd typed it myself. Every judgement call here is mine - AI can draft and test an idea, but it never decides what goes in front of you. Read the AI Policy →
That last point is where this connects to the work I do now. A Web Experience Transformation diagnostic looks at exactly this gap - where a business's data, its website, and its commercial operation have quietly drifted apart from each other, and what that's actually costing. The pattern above isn't unique to AI search; it's the same gap that shows up in conversion, in reporting, and in how confidently a leadership team can trust its own numbers.
If any of this sounds like where your own business is, I'd be glad to talk it through - no pitch, no obligation.
Get in touch
Let's talk about what's stalling.
Whether it's a full diagnostic or a fifteen-minute conversation to see if there's a fit - reach out directly.
Hypernova Transform
AI Search & Zero-Click Commerce
A January 2024 prediction on AI search and agentic commerce, checked against what actually happened by 2026.
