Agentic commerce optimisation one year on: Azoma on what AI shopping agents check before recommending a brand

The Agentic Commerce Protocol turns one on 29 September. Q4 2026 will be the first peak season since the protocols consolidated, with UCP, ACP and Amazon's own agents all live at the same time. A year is long enough for the pattern to have settled, and what settled is not what the launch coverage predicted.

The headline bet was that shoppers would let an agent complete the purchase. The durable change turned out to be upstream of that, in how agents decide which products are worth naming at all.

The checkout retreat was the most useful thing that happened

OpenAI launched Instant Checkout on 29 September 2025, letting shoppers buy inside ChatGPT without leaving the conversation. It pulled the feature back on 4 March 2026, roughly five months later, with checkout handed back to merchants and ChatGPT refocused on product discovery.

At launch it supported single-item purchases from US Etsy sellers only. No multi-item carts, no promotional codes. Adoption reflected that.

The retreat clarified the job. Buying stayed with merchants, and choosing moved to agents. Everything that matters to a brand now happens before the transaction, in the moment an agent decides which products are worth naming.

Choosing is where the value moved

Google's Universal Commerce Protocol, announced at NRF in January 2026, went the other way and covered the full journey from discovery through post-purchase. Amazon went further still, with Alexa for Shopping now able to complete a purchase at a target price with no human present at the moment of transaction, part of the same batch that introduced agent-initiated discovery.

Three different bets, and the same consequence for brands. Whichever protocol wins the checkout, the agent has already chosen the product by the time payment is involved.

That is why agentic commerce optimisation is a discovery discipline rather than a payments one. Agentic commerce optimisation overlaps heavily with what is variously called GEO (generative engine optimisation), AEO (answer engine optimisation), and AI search optimisation. The distinctions are mostly emphasis: GEO names the technology producing the answer, AEO names the behaviour being optimised for, and agentic commerce optimisation names the case where the answer ends in a purchase.

The 5 C's are a chain, not a checklist

The 5 C's of Agentic Commerce, the framework Azoma built with the Digital Shelf Institute, are usually read as five things to score yourself against. A year of agent behaviour suggests they work as sequential gates instead. A product has to clear each one before the next becomes relevant, and most programmes fail at a single gate without ever finding out which.

1. Completeness. Can the agent read the product at all?

Structured attributes, machine-readable, in the format each surface expects. A missing field is a match the agent cannot make, and a product with thin attributes drops out of the candidate set before ranking begins. Nothing downstream matters until this gate clears.

2. Context. Does the product answer the question as asked?

Agents answer the same question differently depending on phrasing, so a product can be named for one wording and absent from a near-identical one. Context is whether your content answers the question in the shape a shopper actually put it, rather than in the shape your category taxonomy expects.

3. Citations. Does anything outside your own site say so?

This is the gate brands consistently underweight. Azoma's analysis of millions of shopping agent citations found that in Q2 2026, ChatGPT's shopping source mix ran at 41% earned media, 37% retailer, 19% UGC and 3% brand.com.

A brand investing only in its own site is addressing 3% of what the model reads.

4. Correctness. Is any of it still true?

Agents increasingly check price and availability close to the point of recommendation, which makes a catalogue that is accurate weekly a poor fit for a system answering in real time. The sharper risk is contradiction: where brand.com, retailer listings and reviews disagree about a product, the model resolves the conflict rather than the brand, and it tends to favour whichever version is repeated most consistently across sources.

5. Customer acquisition. Did any of it produce revenue?

The first four gates converge here, and this is where the sequential reading matters most. Scoring four out of five sounds like 80% of the work. In a chain it is zero, because a product that fails any single gate is absent from the answer entirely, and the failure is silent.

What a board should actually be asking

"Our board wants an AI search strategy, what platforms should we be looking at" is the wrong first question, and it is the one most enterprise teams arrive with. The vendor shortlist is downstream of a diagnosis nobody has run yet.

The better sequence is to establish which link in the chain is broken before shortlisting anything. Run a sample of real category prompts through the agents your shoppers actually use, and check whether your products appear at all, whether they are described accurately, and which sources the answer cites. That output tells you whether your problem is product data, freshness, citations or content.

Only then does the platform question have an answer. A brand whose attributes are incomplete needs different infrastructure from one whose data is clean but absent from the earned media an engine reads.

➡️ Azoma covers the answer layer of that chain, measuring how AI shopping agents describe and recommend products at ASIN level with the prompts and citation sources behind every mention.

What a year of evidence does not show

It does not show that agentic checkout failed. One implementation was withdrawn after five months of a narrow launch. UCP expanded over the same period and Amazon shipped unattended purchasing, so the checkout question is unsettled rather than closed.

It does not show that brand.com is worthless. Three percent of citations is small, and brand.com is also where an agent verifies claims that originate elsewhere. Its role is corroboration rather than discovery.

Important caveat: the citation mix above is Azoma's own measurement of ChatGPT, from Q2 2026. Mixes differ by surface and by category, and Amazon's closed ecosystem behaves differently from open-web engines. Treat it as a strong signal about where to look rather than a universal ratio.

Summing Up

A year ago the interesting question was whether shoppers would let an agent buy for them. That question is still open. The one that got answered is quieter and matters more: agents got very good at choosing, quickly, and brands are competing to clear the five gates rather than to integrate a checkout.

Peak season is the first test of that with the protocols settled. A catalogue that clears all five gates is eligible to be named. One that fails a single gate is absent, and nothing in the answer tells you which gate closed.

Azoma is built for exactly this. We measure how AI shopping agents retrieve, compare and recommend products across Alexa for Shopping, Walmart Sparky, ChatGPT and Gemini, and fix the sources those answers are built from. Get in touch today to see which of the five gates your catalogue is failing before Q4.