Your Best Customer This Black Friday Might Not Be a Person

AI shopping agents are a measurable share of Black Friday traffic. What they read on your store, what makes them bounce, and how to sell to a buyer that never sees your hero image.

The reader that never sees your hero image

At 2:14am on Black Friday, something arrives on one of your product pages. It does not scroll. It does not look at the hero image you spent three weeks art-directing. It ignores the countdown timer entirely, not out of principle but because it never loaded the script that draws it.

It reads your price. It reads your stock status. It looks for a delivery date and a return window. Then it leaves, and somewhere a human being is told either "this one arrives before Christmas and you can send it back until January 15" or "I could not confirm whether this will arrive in time."

That sentence is your conversion rate now. You did not write it. You do not get to review it. And the thing that wrote it was not reading your storefront the way a customer does.

We have written a lot on this blog about what agentic commerce is and which protocols are forming underneath it. This piece is about something narrower and more urgent: what happens when the machine shows up during your busiest 72 hours of the year, and whether your store is legible to it.

Is this actually happening yet, or is it a 2030 problem?

It is small, and it is growing fast, and the growth is concentrated exactly where you do not want a surprise.

The bot-management firm HUMAN Security tracked agent traffic across its customer base through Black Friday and Cyber Monday 2025. Against a pre-holiday baseline of November 21 to 26, agent traffic to its ecommerce clients rose 144.70% to a peak on November 29. The mix shifted too. In October, roughly half of agent traffic went to product and search pages. During the Black Friday window, 83.3% did. Attempts on payment and checkout pages went from under 1% of agent traffic to more than 2%.

Akamai, measuring a different network, reported that AI bots made up 47.9% of commerce traffic as of December 2025, though that figure is dominated by training crawlers rather than shoppers acting for someone. DataDome found LLM crawler traffic rising from 2.6% to 10.1% of verified bot traffic across the first eight months of 2025.

Three caveats worth stating plainly, because this space is full of numbers that do not survive contact with a footnote. All three sources are vendors selling bot management, all three measure only their own customers, and none of them are comparable to each other. HUMAN itself notes that agent traffic is in the millions per month against trillions of total interactions.

So: not the majority of your traffic. Not close. But the shape of it is the interesting part. Agent traffic during peak is overwhelmingly research traffic, not checkout traffic. Something is reading your catalogue at scale and then going away to tell a person what it found.

That is the part you can influence.

What does an AI shopping agent actually see on your product page?

Much less than you think, and almost none of what you designed.

There are two doors into your store for a machine, and they behave differently.

Door one: your feed

OpenAI's Product Feed Specification, published with Stripe under the Agentic Commerce Protocol, is explicit that merchants push data to OpenAI rather than waiting to be crawled, and that the system "accepts updates every 15 minutes." It also notes that frequent updates "improve match quality and reduce out-of-stock or price-mismatch scenarios." Microsoft's Copilot shopping surfaces run on Merchant Center feeds. Google's own agentic surfaces pull from Merchant Center too, and in January 2026 Google added dozens of new Merchant Center attributes, including answers to common product questions and compatible accessories.

A feed is not a nice-to-have channel any more. For a growing share of machine buyers, the feed is your storefront.

Door two: your raw HTML

This is where most stores quietly fail.

In a study of their own network traffic, Vercel and MERJ found that no major AI crawler renders JavaScript. GPTBot fetched JavaScript files in 11.50% of requests and ClaudeBot in 23.84%, but neither executed them. Only Googlebot and AppleBot render. That study dates from late 2024 and no AI vendor has publicly announced adding rendering since, so treat it as the best available evidence rather than a permanent law.

The implication for a Shopify or headless store is uncomfortable. If your price, your variant availability, your delivery estimate or your "only 3 left" indicator is injected client-side by an app or a theme script, there is a real chance the agent reading your page simply does not see it. It does not see an empty space where the price should be. It sees a page with no price on it, and it moves on to a competitor whose price is sitting in the HTML.

Do AI agents fall for countdown timers and scarcity banners?

This is where I want to be careful, because the intuitive answer is wrong in an interesting way.

The tempting claim is that agents are immune to persuasion, so urgency mechanics are wasted on them. The research does not support that.

A 2026 CHI paper by Tang and colleagues, "Dark Patterns Meet GUI Agents," found that browsing agents frequently avoided manipulative interface patterns, but not because they recognised them. The authors describe a paradox of "avoidance without awareness." Agents sidestep some dark patterns through procedural blind spots and a bias toward completing the task, and they fall for others for exactly the same reasons. A separate benchmark, DECEPTICON, estimated dark-pattern effectiveness against a GPT-4o agent at 78.5% with no defences in place, dropping to 59.6% with in-context prompting.

So agents are steerable. They are just steerable by different things.

The most useful evidence here comes from a Columbia Business School study, "What Is Your AI Agent Buying?", which ran randomised product listings past frontier models in a controlled marketplace. The findings are worth sitting with:

• Agents penalised sponsored tags and rewarded endorsements.
• All models showed strong position effects, favouring the top row, but each preferred different columns. Position bias is real and model-dependent.
• Price, ratings and reviews moved decisions in the expected direction, but sensitivity varied sharply between models.
• A seller-side agent that made minor tweaks to product descriptions captured meaningful market share.

That last point is the merchandising lesson. The lever that moves a machine buyer is specificity in your product data, not urgency in your layout. "Runs generous through the shoulder, model is 5'9" wearing a size 6" is a sentence an agent can act on. "Everyone's favourite" is not.

None of this means you should strip urgency from your site. Humans still buy from you, and most of them still respond to a genuine deadline. It means that for the agent reading in parallel, that layer is noise, and the layer underneath it had better be good.

How do you make holiday shipping and returns readable to a machine?

This is the most concrete opportunity in the whole piece, and most stores are leaving it on the floor.

Google's structured data vocabulary already has a property built for exactly this problem: returnPolicySeasonalOverride, part of MerchantReturnPolicy, designed for "special events, such as holidays." If you extend your return window for the gifting season, as most brands do, you can express that in machine-readable form rather than burying it in a paragraph on a policy page that no agent will parse reliably.

The base markup needs applicableCountry and returnPolicyCategory, with merchantReturnDays becoming required once your window is finite. Recommended properties include returnMethod, returnFees and who pays return shipping when the customer simply changed their mind.

Why bother? Because return terms appear to be a trust unlock for the human on the other end of the agent. In a survey of 1,840 consumers across four markets fielded in mid-2026 by RTB House, 42% of American millennials said they would let an AI agent buy on their behalf within a $250 budget if the purchase could be returned within seven days, against roughly a third without that safeguard.

Shipping works the same way. Google's ShippingService type supports transit times through ServicePeriod, with businessDays and duration ranges, and order-value thresholds for free shipping. The OpenAI feed spec carries a delivery_estimate field and requires it to be a future date, which means your estimates need to roll forward daily as cutoffs approach rather than sitting static through December.

USPS dates are for contiguous US delivery by December 25; FedEx dates target December 24. UPS had not published its 2026 schedule at time of writing. Verify against the carriers before you encode anything.

One more thing about peak specifically. Google treats price mismatches between your feed and your landing page as critical errors that can lead to account suspension, and it will stop performing automatic item updates entirely if it finds incorrect microdata or a large-scale mismatch. During a flash sale, your page, your JSON-LD and your feed have to move together. If you change prices more than once a day, Google's own guidance is to move to the Inventory API or Merchant API rather than relying on scheduled feed uploads.

Are you accidentally blocking the agents you want?

Quite possibly, and the reason is usually a robots.txt file written in 2023 by someone reasonably worried about model training.

The distinction that matters is between three different kinds of machine, and they do not share a user agent:

• Training crawlers collect data to train models. GPTBot, ClaudeBot, Google-Extended.
• Search indexers build the index that answers questions. OAI-SearchBot, Claude-SearchBot, PerplexityBot.
• User-triggered fetchers go and look at your page right now because a human just asked a question. ChatGPT-User, Claude-User, Perplexity-User.

Block the first group if you want to. That is a legitimate business decision. But OpenAI's documentation is explicit that opting out of OAI-SearchBot removes you from ChatGPT search results, which is a very different outcome from declining to be training data. OpenAI also notes that because user-triggered actions are initiated by a person, robots.txt rules may not apply to them in the same way.

DataDome found that 88.9% of domains disallow GPTBot. How many of those made a deliberate, separate decision about the search indexer? In my experience reading client robots.txt files, very few. If you are working through the prep side of this, our AI visibility strategy for BFCM and holiday peaks covers the discovery half of the same problem.

There is a second layer of risk below robots.txt. Cloudflare began blocking AI crawlers by default in July 2025 and introduced a "signed agents" class in August 2025 using HTTP Message Signatures, with ChatGPT agent among the first cohort. Visa's Trusted Agent Protocol builds on the same Web Bot Auth work. If your WAF or bot-management rules are set to challenge anything non-human, a legitimate shopping agent acting for a real customer with a real credit card may be getting a CAPTCHA it cannot solve.

I could not find a single named, documented case of a merchant catching this in the wild, which tells you something in itself. The risk here is structural rather than anecdotal. Nobody is monitoring for it.

If you are running through our 11 signs your store is ready for agentic commerce, add this one to the list.

Can a shopping agent talk to your store's own assistant?

Here is the gap nobody has closed yet, and it is the most interesting thing I found.

Shopify's Storefront MCP server exposes catalogue, cart and policy data to external agents, including a tool for searching shop policies and FAQs. That tool matters more than it sounds: it is the only mechanism in the stack that lets an outside agent ask a store something open-ended rather than looking up a structured field.

As the catalogue capabilities moved to newer UCP endpoints through 2026, that policy and FAQ capability did not move with them. The independent tracker UCP Checker noted in September 2026 that it "has no counterpart on the UCP endpoint," with a proposal still open to add one.

This is where a store's own AI agent stops being a support cost and starts being infrastructure. If Alhena is answering on your storefront, it is already grounded in your catalogue, your policies and your order data, and already answering the fit, ingredient, dimension and delivery questions that structured fields cannot express. The shift worth making now is to stop thinking of that agent as something that talks to shoppers, and start thinking of it as something that has to answer whoever asks, at whatever volume peak brings, including a machine working on a shopper's behalf.

The peak season support post covers the human side of that volume problem. The machine side is new, and it is additive - and we go deeper on what happens when those agents reach your inbox in how to prepare your support stack for agentic traffic.

Where does the agent hand the customer back to you?

At the moment it matters most, and usually with less ceremony than you would like.

Walmart told WIRED in March 2026 that purchases completed inside ChatGPT converted three times worse than click-outs to Walmart.com, attributing it largely to single-item checkout. Shopify's Agentic Storefronts documentation notes that on desktop, ChatGPT links out directly to the merchant's store rather than completing in-app.

So for a large share of agent-influenced sessions, the handoff is literal. The agent makes a promise, and then a human lands on your product detail page to verify it.

If the page they land on says something different from what the agent told them, you have manufactured a trust failure at the exact moment of purchase. Same price. Same stock status. Same delivery date. Same return window. The consistency between feed and page is not a technical hygiene issue, it is the conversion event.

Consumer trust data suggests this is where the category actually sits right now. Accenture's survey of more than 25,000 people across 16 countries found 32% would let an agent decide what to buy if they confirmed payment themselves, and only 9% were open to fully autonomous purchases. Checkout.com found 24% of consumers say they will never delegate purchases to AI, while merchants reported that just 3% of transactions involve agents.

The agent is a research and recommendation layer for most people, not a buyer. Which means the handoff is not a footnote in agentic commerce. For the next couple of peak seasons, it is agentic commerce. We went into what went wrong the first time this was attempted at scale in our ChatGPT Instant Checkout post-mortem.

How do you measure a customer who was never a person?

Badly, at the moment. That is the honest answer, and anyone selling you a clean dashboard for this is ahead of the evidence.

Some progress: in May 2026, GA4 began assigning a dedicated "AI Assistant" channel, with the medium ai-assistant, to sessions arriving from recognised chatbot referrers including ChatGPT, Gemini and Claude. That is a real improvement on watching AI traffic scatter across Referral and Direct. ChatGPT also appends utm_source=chatgpt.com to some outbound links.

But that only captures LLM-referred humans, the people who clicked through. It tells you nothing about the agent fetch that happened twenty minutes earlier, because agent fetchers frequently never execute client-side analytics at all. Your GA4 property cannot see a customer who was never rendered a page.

Which leaves server logs as the only honest source of truth. Segment by user agent. Track OAI-SearchBot, ChatGPT-User, PerplexityBot, Perplexity-User, Claude-SearchBot and Claude-User separately from GPTBot and ClaudeBot. Watch which URLs they hit and how often they come back. If agent fetches against a product URL rise and referred sessions do not follow, something in that page is failing the machine read.

No published source currently tracks agent-completed purchases end to end at the merchant level. If you want the broader picture of how LLM-referred traffic is behaving as a channel, our Agentic Commerce Report for DTC founders goes deeper on what is measurable today, and the BFCM data brief covers where shoppers stall on the human side.

Five checks you can run this afternoon

Not a timeline. Our BFCM AI Playbook already has the phased version. These are five things you can verify today, in roughly an hour, with no project plan.

• Curl your bestseller. If price, availability and delivery estimate are not in the raw HTML, fix that before you touch anything else. It is the single highest-leverage change on this list.
• Read your own robots.txt as three decisions, not one. Training, search indexing, user-triggered fetch. Make sure you have deliberately chosen each one, and confirm your WAF is not undoing the choice.
• Add returnPolicySeasonalOverride for your extended holiday window. It already exists, it is already supported, and almost nobody uses it.
• Put the 2026 carrier cutoffs into your shipping markup and your feed, and make sure delivery estimates roll forward daily rather than going stale in mid-December.
• Ask your own support agent the five hardest questions you get at peak. Fit, compatibility, ingredient safety, delivery-by-date, return eligibility. If the answers are good, they are good for a human and a machine alike. If they are vague, they will fail both.

The actual shift

For twenty years, ecommerce merchandising has been an exercise in persuading a person. Photography, copy, social proof, urgency, the small psychological architecture of a well-built product page.

None of that is going away. Most of your revenue this November will still come from human beings looking at pictures.

But a second reader has arrived, and it does not respond to any of it. It responds to whether your data is present, current, specific and consistent between every surface it can reach. That reader is small today and it is compounding quickly, and unlike a human shopper it will never give you the benefit of the doubt on a missing delivery date.

The brands that do well out of the next few peak seasons will be the ones that stopped thinking of structured data as an SEO chore and started treating it as product copy for a customer who cannot see.

Frequently asked questions