Agentic commerce should change roughly a tenth of your BFCM 2026 plan, not most of it. Fix the product data that every AI rail reads, and leave your promotional calendar, email flows, and fulfillment plan alone.
Adobe measured a 693% year over year jump in AI traffic to retail sites last holiday season, then added the sentence almost nobody quoted: the base of users remains modest. Both halves are true and only one of them gets repeated.
Black Friday lands on November 27 this year, which puts ten weeks between today and the weekend that decides your quarter. In those ten weeks you will be told, repeatedly, that agentic commerce has changed the fundamentals. Five things shipped in the last three weeks alone to support that claim: Anthropic open sourced a working shopping agent, Mastercard launched Agent Connect, OpenAI put a ChatGPT Ads app in the Shopify App Store, M11 Labs launched a marketplace where brand agents compete for the recommendation, and Shopify kept extending Catalog.
Here is the number that should govern how much of your plan moves in response. eMarketer forecasts that US ecommerce sales through AI platforms will exceed $20 billion in 2026 and top $144 billion by 2029. Twenty billion is a real number. Against a US ecommerce market running well over a trillion dollars a year, it is also somewhere in the neighbourhood of one and a half percent.
Both of those facts are true simultaneously, and the entire question of how to plan Q4 sits in holding them at once. A channel growing that fast off a base that small deserves attention proportionate to where it will be, and investment proportionate to where it is. Most of the advice circulating this month gets that ratio badly wrong in the expensive direction.
Agentic commerce is currently a small, fast growing, unusually high intent slice of ecommerce, and the growth rate is the part that gets quoted while the slice is the part that should size your budget. Adobe’s final read on the 2025 holiday season found $257.8 billion in US online spend, up 6.8% year over year, with traffic to retail sites from generative AI tools up 693.4%. In the same analysis Adobe noted that the base of users remains modest, and that paid search and email remained the primary drivers of sales volume.
The intent quality is where the story gets genuinely interesting. Adobe’s 2026 Prime Day analysis found AI sourced traffic converting 40% better than non AI channels including paid search, email, and social, a complete reversal from the prior year when AI traffic underperformed those channels by 23%. Shopify’s Q2 numbers point the same way, with AI driven traffic and orders each tripling year over year.
One Shopify finding deserves more attention than it got: 75% of AI attributed purchases came from outside its top 100 product categories. That is a specific, useful signal. If you sell something narrow, specific, and hard to describe in three words, the agentic surfaces are disproportionately good for you right now. If you sell a commodity in a crowded category, they are not, and the growth rate in the headlines is not describing your business.
None of this adds up to a channel that carries your Q4. It adds up to a channel worth being correctly configured for, which costs far less than being early to.
You do not need to pick a protocol before Black Friday, and any vendor telling you otherwise is selling implementation hours. Two standards genuinely matter and they are not symmetrical. The Universal Commerce Protocol launched at NRF in January 2026, led by Google with a coalition including Shopify, Visa, Mastercard, and Stripe. The Agentic Commerce Protocol is OpenAI and Stripe’s, centred on ChatGPT. I broke down how these relate to MCP in the retailer’s guide to AI shopping protocols, and the relationship has not changed since, only the number of companies announcing adjacent products.
Mastercard is the clearest example of why watching beats acting. It announced Agent Connect and an updated Agent Suite for Merchants on September 9, promising one integration covering merchants, agents, platforms, and payment providers. It published no merchant deployment timeline. A product with no availability date is a roadmap, and roadmaps do not belong in a Q4 plan.
Run the eighteen month test on each of these rather than the this week test. In eighteen months, will it have mattered that you implemented a hosted UCP endpoint in October 2026 rather than March 2027? For a merchant under $10M, almost certainly not. Will it have mattered that your product attributes were complete? That one compounds from the day you do it.
Every agent rail in that table reads your product data, which makes catalog quality the only agentic work with a guaranteed return regardless of which standard wins. UCP, ACP, MCP, Shopify Catalog, Anthropic’s agents, and the ChatGPT Ads app are all, underneath the branding, systems that ingest structured product information and make a recommendation from it. They disagree about transport, fees, and who holds the checkout. They do not disagree about needing to know that your boot comes in a wide fitting.
Shopify’s own guidance on agentic ready product data names four attributes: parent and child variant grouping so an agent understands options, taxonomy specificity (“men’s insulated winter boots” rather than “footwear”), literal descriptive language instead of marketing copy, and real time pricing and inventory by API rather than scraping. That list has not changed since June and will not change when the next protocol ships, because it describes what a machine needs to answer a question, not what any particular company built.
The quiet advantage here is that none of that work is agentic specific. Specific product types improve your Google Shopping feed. Complete variant data reduces your return rate. Literal descriptions help the shopper who lands from any channel and is trying to work out whether the thing fits. You are not placing a bet on agentic commerce by doing it. You are doing the work that the last decade of ecommerce already rewarded, for a second set of reasons.
That is the test I would apply to every agentic recommendation you receive this quarter. If it only pays off when agentic commerce grows, it is a bet. If it pays off either way, it is maintenance you were behind on.
Three things change, and all three are cheap. First, your product data is now read by machines that will describe your product to a buyer without you present, which raises the cost of vague attributes from a conversion rate problem to a representation problem. Second, a paid surface opened inside AI answers that did not exist during your last peak, which is worth a small test rather than a budget line. Third, AI sourced traffic now converts better than your other channels rather than worse, which means the tracking gap is no longer a rounding error.
That third one is where most merchants are genuinely exposed. If AI referred sessions are landing in your analytics as direct traffic, you will finish BFCM with a channel that outperformed everything else and no evidence it existed. You will then plan 2027 from the wrong numbers. Fixing attribution is a one afternoon job and it is the highest leverage agentic work available to you before November, ahead of anything protocol related.
What does not change is the shape of the quarter. The promotional calendar, the inventory buy, the email and SMS flows, the customer service staffing, the shipping cutoffs, the returns policy. My own BFCM 2026 data report found demand effectively guaranteed, with 99% planning to shop and 42% planning to spend more, and roughly 80% of shoppers starting before Thanksgiving week. Nothing about agents changes when people start shopping or what makes them abandon a cart.
The cost side has moved more than the channel side this year, and it is getting a fraction of the attention. Five inputs shifted against importing merchants between June and August, which I covered in the Q4 margin math breakdown. A merchant who spends October implementing a protocol and never reprices against current freight and tariff costs has optimized the wrong variable by an order of magnitude.
The parts of BFCM that decide the outcome are the same parts that decided it seven years ago, and they are all unglamorous. Inventory depth on your top sellers. A site that stays up under load. Shipping cutoffs you can actually hit. A returns policy that does not eat the margin you just earned. Customer service coverage sized to the volume rather than to last year’s volume. Email and SMS lists warmed before the week rather than during it.
Adobe’s data makes the point better than argument does. In the season where AI traffic grew 693%, paid search and email were still the primary drivers of sales volume, and social commerce moved from 3.3% to 4.6% of revenue. The channels that carried the quarter were the channels that carried the previous quarter. That is what a small base looks like from inside a growth story.
There is a version of this argument that becomes an excuse to ignore the shift entirely, and that is not the argument. The merchants who are correctly positioned in 2028 will be the ones who did the data work in 2026 while everyone else was picking protocols. The point is that the data work and the protocol work are not the same work, and only one of them has to happen before November.
If you want a sanity check on any given recommendation, ask whether it would have been good advice in 2019 with the AI language stripped out. Clean up your product titles would have been good advice in 2019. Make sure your analytics can see where orders come from would have been good advice in 2019. Implement a hosted endpoint for a six month old standard backed by a coalition that did not exist last January would not have been.
Run the cheap, universal work first and defer everything that requires a bet on a standard. Weeks one and two, which is now, export your top 20 revenue SKUs and check them against the four attributes above. Ninety minutes for most catalogs at that size, no budget, and it improves every channel you already run. In the same fortnight, verify that your analytics can distinguish AI referred sessions from direct traffic, because everything you learn this quarter depends on that being true.
Weeks three and four, reprice against your actual landed costs rather than last year’s, and finalize the promotional calendar. This is the work that determines whether the quarter is profitable, and it has no agentic component whatsoever. Do not let it slip because something shipped.
Weeks five and six, if you are above $500K and the catalog work is genuinely finished, run one small paid test on the AI surface with a budget you would write off and a stop rule set in advance. You are buying information, not revenue. Anything you learn goes into the 2027 plan, not this one.
Weeks seven through ten, freeze. No new integrations, no new apps, no protocol implementations. The failure mode I watched most often across merchant accounts in the $500K to $2M band was not missing a channel. It was adding one too many things to the stack in the six weeks before peak and discovering the conflict on Black Friday morning.
The specific failure waiting for merchants in this band is rebuilding the Q4 plan around a channel that will deliver a low single digit percentage of their revenue this year. It is a seductive mistake because it feels like being early rather than being distracted, and because the growth rates are genuinely extraordinary. A 693% increase is not a small thing. It is just not, yet, a large thing.
The way I would frame it for anyone at this stage: agentic commerce deserves the share of your attention that matches its share of your revenue, plus a small premium for where it is heading. If AI surfaces send you 2% of your orders, give the category something like 10% of your Q4 planning attention. That premium buys you the catalog work and the attribution fix, which is exactly the right amount, because those two things are also what you would do if the category stalled tomorrow.
Where this reasoning breaks down is worth naming honestly. If you sell in a narrow category, and you are seeing AI referred sessions convert well above your site average, the 10% figure is too low for you and you should weight it higher. Shopify’s finding that three quarters of AI attributed purchases came from outside the top 100 categories suggests a meaningful number of merchants are in exactly that position. Check your own numbers before accepting my ratio, which is the same advice I would give about any benchmark, including the ones that contradict each other.
The signal that should make you revisit all of this is not another protocol launch. It is your own analytics showing AI referred revenue crossing something like 5% of the total, sustained across a full quarter. At that point the category has stopped being a small fast growing slice of your business and started being a channel, and a channel gets a channel’s worth of planning. Until then, the catalog work is the whole job.
Roughly a tenth of your planning attention, concentrated on two cheap tasks: cleaning product data on your top revenue SKUs and making sure your analytics can distinguish AI referred sessions from direct traffic. The promotional calendar, inventory buy, email and SMS flows, shipping cutoffs, and customer service staffing should be unchanged, because nothing about AI agents changes when shoppers start buying or why they abandon carts. A useful ratio is to give the category attention matching its share of your revenue plus a modest premium for its growth trajectory. For most merchants sending fewer than 5% of orders through AI surfaces, that lands near 10%.
No. Neither protocol requires implementation before peak season for a merchant under roughly $10M, and the engineering hours are better spent elsewhere in a ten week window. The Universal Commerce Protocol launched in January 2026 with Google leading a coalition including Shopify, Visa, Mastercard, and Stripe, and typically involves hosting an endpoint and eight to sixteen hours of work. The Agentic Commerce Protocol from OpenAI and Stripe is faster to adopt but carries a platform fee. Both will still be available in the first quarter of 2027, when you can evaluate them against real data from your own store rather than against vendor forecasts.
US ecommerce sales through AI platforms are forecast to exceed $20 billion in 2026, which against a US ecommerce market running well over a trillion dollars annually works out to somewhere near one and a half percent. Growth rates are genuinely dramatic, with Adobe measuring a 693.4% year over year increase in AI sourced traffic to retail sites during the 2025 holiday season, but Adobe noted in the same analysis that the base of users remains modest and that paid search and email remained the primary drivers of sales volume. Both the growth rate and the small base are true, and planning requires holding both.
Yes, and that reversed recently, which is why it is worth measuring rather than assuming. Adobe’s 2026 Prime Day analysis found AI sourced traffic converting 40% better than non AI channels including paid search, email, and social, a full reversal from the prior year when AI traffic underperformed those channels by 23%. Shopify has reported similar direction from its own merchant base. The likely mechanism is intent rather than technology: shoppers arriving from an AI conversation have already done comparison work before they land. The practical implication is that if those sessions are being recorded as direct traffic, you are underinvesting in a channel that is outperforming.
Fix attribution before anything else, because every other decision you make this quarter depends on being able to read the results. If AI referred sessions are landing in your analytics as direct traffic, you will finish BFCM unable to tell which channel actually performed, and you will plan 2027 from the wrong numbers. Once attribution is trustworthy, the next highest value afternoon goes to product data on your top 20 revenue SKUs: correct variant grouping, specific product types rather than generic categories, literal descriptive language, and accurate real time inventory. That work improves every channel you run, not only the agentic ones.
