A test on e-commerce, verifiable constraints, and seven ways the process breaks down.
Christmas is a week away. I wanted to do something simple: order an artisanal panettone online.
So I asked GPT in agent mode to execute a concrete task with explicit constraints:
- artisanal panettone
- bakery with a Google rating ≥ 4.2
- weight ≥ 750g
- price < €40
- purchase, with my final confirmation
This is exactly the kind of task agents are built for. Instead, the agent spent 31 minutes arriving at a frustrating result: it couldn’t complete a purchase that met all the requirements simultaneously.
What it produced along the way, though, is a precise diagnostic of how fragile many e-commerce systems are when a purchase is mediated by a non-human actor. This isn’t a SEO problem, or a GEO problem. It’s a problem of agent-actionability: of AXO.
The method (for those who like things verifiable)
Prompt: “Search online and, if possible, purchase an artisanal panettone… Google rating ≥ 4.2… weight ≥ 750g… price < €40.”
Session duration: ~31 minutes
Observed outputs: e-commerce navigation, attempts to verify requirements, fallbacks to third-party sites for ratings, technical dead-ends (403/404), blank pages, anomalous redirects, infinite scroll, cookie overlays.
This is not a universal benchmark. It’s a robustness test of e-commerce shops selling panettone during peak demand season. The point isn’t whether panettone is available online — it’s how possible it is to verify and complete a purchase through an AI agent operating under explicit constraints.
The key point: an agent doesn’t buy “by feel”
A human, faced with an unfamiliar site, intuits, tolerates, approximates: “it’s probably 750g, I’ll take it anyway.” An agent has to do something different: read the data, verify it, make a decision consistent with its constraints, complete the action. When any one of those steps is missing, the result isn’t friction — it’s a block. The process stops.
These are the seven failure points I observed.
1. Availability discovered late, or unreliably
The agent finds a promising product, moves forward in the process, then discovers it’s out of stock. It has to start over.
If availability isn’t exposed in a stable, readable way from the beginning of the journey, the decision-making process breaks at the worst possible moment — after time and resources have already been spent on exploration.
2. Pages too dynamic to parse: the agent falls back to text extraction
At several points, the agent attempts workarounds — “reader mode,” opening adjacent elements, searching for strings in the page text (“750”) — because the page isn’t interpretable through standard means.
The translation of this behavior is straightforward: “I can’t read this site the way I should, so I’m trying to treat it like a text document.” If an agent has to treat your e-commerce as a badly formatted PDF, the purchase process is already compromised — not because the agent is limited, but because the system was never designed to be read by an actor that doesn’t improvise.
3. Verifying a Google rating becomes a hunt across aggregators
The “≥ 4.2 on Google” requirement forces the agent off the merchant’s site. If it can’t find an immediate match on Google, it has to chase evidence across Wheree, Wanderlog, RestaurantGuru, TripAdvisor, and similar platforms.
The problem is twofold. The reputational evidence is outside the brand’s control, often outdated, and not clearly attributable to Google. And the brand ends up being interpreted elsewhere — through sources the merchant doesn’t manage, with data it can’t verify. If the site doesn’t provide a reliable primary reference to its own Google Business profile, reputation becomes a fragile variable in the agent’s decision-making process.
4. 403/404 errors and blank pages: dead-ends with no exit
A human facing a blank page will reload, try a different browser, wait. An agent reads the same event as a process failure and changes course — often toward a competitor or an alternative source.
What is a temporary interruption in a human flow is an exit point in an agentic flow. The distinction isn’t technical: it’s a design distinction.
5. Infinite scroll: a catalog built to keep people browsing, not to help them decide
At a certain point in the logs: “There might be infinite scroll here… I’ll wait.”
Infinite scroll works for humans because it’s designed to retain them — more products, more stimuli, more time on-site. For an agent that needs to compare options and make a decision against explicit constraints, the same mechanism becomes a limbo: expensive exploration, difficult comparison, no closing point. Catalogs designed for human scrolling are not designed for deterministic selection.
6. Cookie banners as a source of non-deterministic errors
The logs contain a fairly long sequence describing the agent’s attempts to handle a cookie overlay: it closes the X, nothing happens, it tries again, it gets a 404 that was probably caused by the interaction with the overlay.
For a human, a cookie banner is an annoyance. For an AI agent, it’s a source of unpredictable states: it intercepts clicks, produces behavior the system can’t interpret, and after a few failed attempts gets classified as a block. The issue isn’t the banner itself — it’s that it was never designed to be handled by a non-human actor.
7. The panettone becomes a pair of Chopard sunglasses
When the agent seemed close to completing the purchase, the navigation derailed toward a completely different product. The causes can be various: unstable slugs, confused routing, dynamic components, interactions between overlays and open sessions.
In a human purchase flow, this kind of anomaly is strange but manageable — the user notices and goes back. In an agentic flow, it means loss of context and interruption of the sequence. It’s not a curious bug: it’s a structural failure point.
What this test tells us
The agent didn’t fail because panettone doesn’t exist online. It failed because many e-commerce systems don’t expose stable signals — product data, availability, reputation, purchase paths — in a way that a non-human actor can verify and act on reliably.
SEO asks: “can they find me?” GEO asks: “do they cite me?” AXO asks something different: “can they verify me, compare me, choose me, and buy me?” It’s a change of interlocutor, not another step on the same scale.
When selection is mediated by an agent, the experience to design isn’t the navigation — it’s the process of decision, verification, and action. An e-commerce that can’t hold up under that process isn’t a poorly optimized store: it’s a store that, for a growing part of the ecosystem, simply doesn’t exist.
If you sell online and want to understand whether your shop is verifiable and purchasable by agents and assistants, write to me.