Agentic commerce: ChatGPT converts 42% better than Google in 2026. Here is how your site gets shortlisted by AI agents.

The context: why this article is written in 2026, not 2028

We were patient on agentic commerce. In 2024 it was demo-ware. Early 2025 it was a promise. Then Adobe Analytics published its April 2026 report with a number that changed our read: AI-referred traffic (ChatGPT, Perplexity, Claude, Gemini) converted 38% WORSE than Google traffic in March 2025. Twelve months later, that same traffic converts 42% BETTER. That is an 80-point swing in 12 months on the same dataset from the same merchants.

On market size, Salesforce documented $262 billion in sales attributable to AI agents during the 2025 holiday season (20% of global orders). Adobe measured +805% AI-referred retail traffic on Black Friday 2025 and +670% on Cyber Monday. On Q1 2026, +393% on average. These are not consultant projections, they are measurements from trackers already installed at the merchants themselves.

The real question for an SMB merchant in Quebec is no longer will this happen. It is: what do I change on my site this week so that when ChatGPT picks three backpacks for a user, one of the three is mine.

What an agent looks for when it shops for your customer

An AI agent does not match keywords. It matches constraints.

When a user types into ChatGPT durable backpack, for a 3-week backpacking trip in Asia, under $200, delivered before March 1, the agent breaks the query down into 4 distinct sub-tasks:

1. Filter category + price + delay: backpacks, under $200, availability + shipping OK before March 1.
2. Match use case: Asia 3-week backpacking trip implies 40-65L capacity, laptop compartment, water resistance, TSA lock, compression straps.
3. Qualitative filter: durable implies materials (Cordura, 500D+ ripstop), reinforced stitching, multi-year warranty.
4. Final ranking: the PDPs that explicit the highest number of matches, with proof (descriptive reviews, precise technical specs, mentions by third parties), rise to the top of the shortlist.

The trap: a PDP written for traditional SEO (blue backpack, premium leather, urban look) does not match any of the 4 dimensions. It is invisible to the agent, even if it ranks position 1 on Google for blue backpack.

What we find in PDPs that get cited:

• Named use cases: ideal for 3-4 week trek, cabin carry-on, Southeast Asia backpacking. Not 1 generic case, but 3 to 5 precise scenarios. That lets the agent say here is a match for your situation.
• Explicit exclusions: not ideal for winter alpine hiking, not designed to carry more than 18 kg. Counter-intuitive, but exclusions help the agent eliminate false matches and increase confidence on true ones.
• Concrete compatibility: dimensions in cm, max carried weight in kg, cabin bag compliance on the 3 main airlines, dishwasher-safe yes/no, iOS 26+ compatible yes/no. Every constraint made explicit is a potential match.
• FAQ anchored on real customer questions: pull from support chat, reviews, product Reddit threads. Every Q/A becomes training data for the PDP.
• Specs in structured table format, not prose. Models extract lists and tables far better than paragraphs.

The robots.txt battle: block or allow

Amazon chose to block. In its robots.txt, ChatGPT-User and OAI-SearchBot are disallowed. Similarweb-measured result: Amazon's share of ChatGPT shopping queries dropped below 3%, falling around 18% month over month. Etsy, which did the opposite (ACP launch partner in September 2025), moved above 20%.

Amazon's logic makes sense: their on-site ad business is worth $56 billion per year. If agents bypass the storefront and buy without seeing the sponsored ads, that value evaporates. So they blocked the crawl, sued Perplexity in November 2025 (Comet was browsing Amazon disguised as Chrome), and won a preliminary injunction on March 10, 2026 blocking Comet from Amazon accounts. The 9th Circuit ruling expected after the June 11, 2026 hearing will set the first federal precedent on buyer-side AI agents.

For an SMB in Quebec this is not theoretical. Blocking AI crawlers in robots.txt = disappearing from the shopping panels of ChatGPT / Perplexity / Google AI Mode. Allowing them = accepting that the agent scrapes your PDP, presents it next to 3 competitors, and potentially negotiates the sale without the user ever setting foot on your site. Our default position with merchant clients: allow, optimize to get picked, and double down on post-click KPIs (repeat rate, LTV, brand searches), because the initial click becomes rare but qualified.

The 3 protocols shaping the market

Since September 2025, three protocols have locked in. They do different things, and you need to understand which one applies to what:

The 8 changes we ship on merchant sites before Q4

We grouped the interventions by layer. A merchant who ships the first 3 already captures 80% of the gain, because those are the layers agents read first.

The real 2026 risk: margin, not visibility

We spend a lot of time on visibility because it is the first bottleneck. But the real strategic conversation starts after you are visible.

When a user buys through ChatGPT Instant Checkout, OpenAI takes a 4% fee, Stripe takes 2.9%, and you lose the customer data (email, address, purchase history consolidated in ChatGPT, not in your CRM). It is the marketplace pattern we know from Amazon and Etsy for 15 years, except the marketplace is no longer a website, it is a conversation. And the user will probably tell ChatGPT order me the same product next time without ever thinking about your brand.

Two strategies we see with our merchants:

1. Treat the channel as paid acquisition: the ChatGPT channel is treated like paid, with a CAC equivalent (the 6-7% margin). The KPI is no longer transaction ROAS, it is the direct-repeat rate you can pull back to your own site within 90 days (via post-delivery email + product quality).
2. Block on low-margin SKUs, activate on new drops and hero products: we turn on ACP only on the 20 to 40 strategic SKUs where we want fast volume and awareness, and keep the repeaters (consumables, accessories) on our own site where the margin holds.

The activate everything, we will see position on a 1,000-SKU catalog with no margin strategy - we have not seen a merchant come back happy from that one.

Verdict: where to start with 2 weeks and a $5,000 budget

If you are an SMB merchant with a catalog of 50 to 500 SKUs and you want to be present when your customers ask ChatGPT or Perplexity to find them a product, here is the order:

Week 1 - Audit + foundations:

• Check robots.txt + firewall (Cloudflare, Sucuri) and unblock ChatGPT-User, OAI-SearchBot, PerplexityBot, ClaudeBot, Google-Extended.
• Audit the Merchant Center feed. Confirm that product_highlight is filled and use-case custom attributes are in place.
• Pick the 20 to 40 strategic SKUs where you will invest layer 2 (page structure).

Week 2 - Rewrite the strategic PDPs:

• On those 20 to 40 SKUs, rewrite the page in 5 blocks: Ideal for / Not ideal for / Technical specs (table) / FAQ (6-10 questions) / Concrete compatibility.
• Add Product, Offer, FAQPage, AggregateRating JSON-LD schemas. Confirm the schema matches what is visible on the page (no drift).
• Turn on Shopify's Agentic Storefronts channel (Winter 2026) to syndicate to ChatGPT / Perplexity / Copilot from a single catalog.

After 30 days - Measure:

• Look at AI-referred traffic in GA4 (segment session_source containing chatgpt, perplexity, claude, gemini).
• Install a share of model tracker (Profound, TruIntel or equivalent) on the 20-40 SKUs.
• Compare AI vs non-AI conversion. If AI conversion is lower on your site, it is probably a PDP structure problem, not a channel problem.

We wrote this article because the question should I be worrying about AI agents has come up in every e-commerce strategy call since March 2026. The answer is yes, now, not Q4.