AI Shopping Agents at Peak: Getting Found When One in Five Holiday Visits Is a Bot

Salesforce expects 20% of all 2026 holiday ecommerce traffic to come from AI chat agents: shopping assistants answering live questions, autonomous agents doing tasks, and competitor scrapers checking prices (Salesforce, 20 July 2026). That makes AI shopping agents an ecommerce problem for this peak season, not a future one. Agents do not browse, they parse. They read your product data, your prices, your shipping and your returns, and they compare them against everyone else's. Stores whose information is structured, specific and consistent get shown to the shopper. Stores with vague promises get skipped. This guide covers what agents read, what they ignore, and a checklist to finish before Black Friday.

What an AI shopping agent actually does

"AI agent" covers three different kinds of traffic, and Salesforce's forecast counts all of them: consumer-facing bots handling live queries, autonomous agents executing backend tasks, and competitor scrapers feeding algorithmic price-matching. So "one visit in five is a bot" is fair. "One shopper in five uses an agent" is not.

The human side is growing fast from a small base. Adobe measured a 693% rise in shoppers clicking through from generative AI tools to US retail sites over the 2025 holiday season, and those visitors converted 31% more than other traffic (Adobe, 7 January 2026; Adobe, 12 January 2026). For 2026, Adobe forecasts that AI traffic to US retail sites will rise another 130%, with the biggest jump on Thanksgiving (Adobe, 28 September 2026).

What these assistants look at is documented. OpenAI says ChatGPT's shopping research reads "price, availability, reviews, specs, and images" (OpenAI, 24 November 2025), and that when several merchants sell the same product it weighs availability, price, quality and whether the merchant is the primary seller (OpenAI, 29 September 2025). The pattern is the same everywhere: the agent assembles a comparison, and the shopper chooses from what the agent could read.

What agents read

Agents take product information from two places: feeds and pages. Feeds matter more than most store owners realise.

Product feeds. If you sell on Shopify, eligible products are in Shopify Catalog by default, and that catalog is how ChatGPT, Microsoft Copilot and Perplexity find Shopify products (Shopify, 17 June 2026; Shopify, 24 March 2026). On Google, Merchant Center feeds feed both AI answers and regular results. Google's own guidance says feeds "can help your products and services to be visible in both AI responses and other Google Search results" (Google Search Central).

Structured data on the page. Product markup with price and availability, shipping details with delivery times, and a return policy. Be precise about what this does. Google says structured data "isn't required for generative AI search", but shipping and return markup makes your listing eligible to display delivery times, shipping costs and return windows (Google merchant listing docs). It earns you a clearer listing, not a ranking boost.

Clean specifications. Materials, dimensions, weight, compatibility, what is in the box. Adobe tested how readable retail sites are to language models and flagged product pages as a problem area, at 66%, against 82% for returns and exchanges pages (Adobe, 16 April 2026). The product page, the one that sells, is the one agents struggle with most.

Site-level files such as llms.txt. This is where honesty matters. llms.txt is a 2024 proposal by Jeremy Howard for a markdown file that gives language models a clean map of a site (llmstxt.org). Google states plainly that Google Search does not use it and ignores it, and Google's John Mueller has said none of the AI services have said they use it (Search Engine Journal, 17 April 2025). It is cheap to publish and harmless, but it is not a lever. Feeds and accurate pages are.

The delivery-time problem

Delivery is where most stores sourcing from China lose the comparison without knowing it.

When an agent can display delivery information, the shopper sees your promise next to everyone else's. Google's shipping markup exists to show "shipping costs and estimated delivery timeframes based on their location", and OpenAI lists delivery ETAs on its roadmap for agent shopping. A listing that says "7 to 15 days" sits next to one that says "4 to 8 days to the United States" and loses, even if both parcels arrive on the same day.

There is no published evidence that faster delivery makes an agent rank you higher, and we will not claim it. The loss happens one step later, when a human reads the options the agent laid out and picks the one that sounds reliable. A specific window per destination reads as reliable. A wide range reads as a guess.

So publish delivery windows per country, not one global range. Ours are on each country page, for example shipping to the United States, and the trade-off between services is in express vs standard shipping from China.

How we did it

We publish two files for machines, and both are live.

/llms.txt is a plain map of the site: what we do, the service pages, blog posts and help articles with one-line summaries, our main shipping destinations, and a short section telling AI assistants how to describe us. It exists because an assistant asked about "China fulfillment" should not have to guess from our homepage design.

/pricing.md is our pricing in plain markdown: the plans, what each includes, and how the per-order rate is charged, with a pointer to live shipping prices by destination. It is written for agents and procurement workflows that need facts without parsing a web page.

The lesson from maintaining them is not about format. It is about rot. Nobody reads these files except machines, so nobody notices when they go stale. Writing this post, we re-read ours against our live pages and found several lines that had drifted, including a duty rule that changed last year. We fixed them the same day. Put your machine-readable files on the same review cycle as your pricing page, or they will quietly tell agents things that stopped being true.

One more first-hand signal. In our own Search Console data over the last three months, dozens of queries read like prompts rather than searches, for example: "context: location: united states ... question: how do fulfillment providers in china calculate storage, pick/pack, and outbound shipping fees?" Those are assistants searching on a shopper's behalf. The questions they ask are specific, and they reward pages that answer specifically.

A pre-Black Friday checklist for your store

Six items, in the order we would do them:

1. Product structured data with price and availability on every product page, matching what the page shows. Out-of-stock products marked out of stock.
2. Shipping details. Either shipping markup on each product or a single store-wide shipping policy, which Google recommends providing under your Organization markup, with delivery times per destination.
3. Return policy markup. The window, who pays return shipping, and any fees. If product-level and store-level policies both exist, Google uses the product-level one, so keep them consistent.
4. Specific delivery windows. Per country, in days, on product pages and in your shipping policy. Drop "7 to 15 days worldwide".
5. Prices consistent across every page and feed. The product page, the collection page, the feed, the cart and any markdown files must agree. Agents compare them and treat disagreement as a warning sign.
6. A clean product feed. On Shopify, check that your products are eligible for Shopify Catalog; on Google, fix Merchant Center feed errors before peak. Google says merchants who adopt core feed best practices see, on average, a 5% rise in conversions the following month (Google, 16 September 2026).

What not to do

Agents cross-check, and they remember.

• Inflated claims. "Waterproof", "medical grade", "lifetime warranty" without the paperwork. An agent that reads a certification-free claim next to a certified competitor has an easy choice.
• Fake or undisclosed incentivised reviews. Google added a guideline against them to its review snippet documentation in July 2026.
• Inflated list prices. Price history is now a shopping feature: Amazon's Alexa for Shopping shows up to 365 days of it. A "70% off" badge on a price that was never charged is easy to catch.
• Prices that differ between pages. The single most avoidable reason an agent treats a store as unreliable.

For the numbers behind the season, see Black Friday statistics 2026 and how brands are preparing for Q4 2026. Social platforms follow the same logic of fast, specific promises; see TikTok Shop fulfillment from China.

Ready to make your delivery promise specific? Connect your store and take real windows per country from day one, or see pricing for the plans.

Last reviewed 29 September 2026.