Black Friday is three weeks away. ChatGPT Search has been GA for a year, the Agentic Commerce Protocol was announced this fall, and the percentage of e-commerce traffic that originates from an AI agent has been climbing every quarter. This is the first BFCM where agent-mediated discovery is large enough that ignoring it has a price tag.
Most merchant prep this year has gone into the usual suspects: ad budget reallocation, email warmup, inventory positioning, page speed. All necessary. None of it covers the new surface.
This is the catch-up list for that surface. Eight items, scannable, in priority order. The merchant teams who run this checklist between now and Thanksgiving will go into BFCM with their store legible to agents. The teams who don’t will find out what invisible looks like.
1. Get your Product and Offer JSON-LD right, and put GTINs on everything
Agents read structured data before they read prose. That’s not a stylistic preference; it’s load-bearing. Gemini extracts JSON-LD before parsing the page HTML and weights it roughly 2.5x heavier than visible text. Copilot’s Knowledge Graph entity resolution depends on GTIN/UPC. Without a clean Product schema with price, availability, and identifiers, you are functionally invisible to the agents most likely to recommend you.
Check that every product page renders Product JSON-LD with name, image, description, brand, gtin13, mpn, sku, and a nested Offer block with price, priceCurrency, availability (InStock or OutOfStock, pick honestly), and priceValidUntil. If your platform doesn’t emit GTINs, fix that this week. By BFCM weekend, the agents you missed will be quoting your competitors.
2. Ship an llms.txt, and make it honest, not marketing copy
llms.txt is the discoverable summary an agent reads to understand what your business is, what you sell, and what policies you have. Grok actively fetches it. ChatGPT and Claude crawlers reference it when present. The agents that find a clear, accurate llms.txt index your catalog with higher confidence; the ones that don’t fall back to inferring from your homepage prose, which is slower and lower-fidelity.
The mistake to avoid: writing it as marketing copy. Agents downgrade content that reads as promotional. The version that works is matter-of-fact. What you sell, your shipping cutoffs, your return policy, your size-availability honesty, any product limitations worth knowing. If you’d be embarrassed for a customer to read it, rewrite it. Agents will be the customer reading it.
3. Pick a side on agentic checkout
ACP (Agentic Commerce Protocol) is the new path for an agent to complete a purchase without the buyer leaving the conversation. Shopify shipped support in October. BigCommerce and several headless commerce platforms are following. The merchants who have ACP wired by BFCM weekend will capture buyers who never load a product page.
If you’re on Shopify and you haven’t enabled the merchant-side ACP toggle, do it this week. If you’re on a custom stack, at minimum publish an ACP manifest at /.well-known/acp/manifest.json that declares your catalog endpoint and supported actions. The investment is small. The downside of skipping it is that any agent considering you and a competitor with ACP enabled defaults to the competitor.
4. Audit your FAQ for the questions agents actually ask
The questions a human reads on your FAQ page (“What’s your return policy?”) are not the questions an agent asks (“What’s the return window for unopened items shipped to California after Nov 25, and does the holiday-extended-returns policy apply?”). Agents construct specific multi-constraint queries because they’re trying to give the buyer a confident answer.
Walk through your FAQ as though you were briefing an agent. Are the specific dates explicit? Are the conditional cases (sale items, gift returns, international, custom orders) actually addressed, or hand-waved? Where the FAQ is vague, the agent will either refuse to answer or guess. A guessing agent reads “I’m not sure about this brand, let me check a competitor.” Specificity is conversion.
5. Tell the truth about availability, including the bad news
The single most expensive lie a store can tell an agent right now is fake stock. An agent that recommends an out-of-stock product to a buyer, drives them through the cart, and lands them on a sold-out page learns two things: don’t trust this catalog, and downrank in future comparisons.
Real-time inventory sync to your product feed, your ACP manifest, your llms.txt, and your structured data. If you can’t push it to all four, push it to the ones you can and explicitly mark availability with a timestamp in the others. Honesty about being sold out costs nothing this BFCM. Lying about it costs the next twelve months of agent recommendations.
6. Do not blanket-block AI crawlers
A surprising number of merchants entered 2025 with User-agent: GPTBot and User-agent: ClaudeBot set to Disallow: / in robots.txt. The reasoning, when there’s any, is “I don’t want my content training their models.” The cost is being invisible to ChatGPT Search, Claude product browsing, and Copilot Knowledge Graph indexing. Those are the surfaces the buyers using these tools rely on for recommendations.
Audit your robots.txt this week. Block training crawlers (GPTBot, Google-Extended, anthropic-ai) if you have a strong opinion about that. Do not block fetch crawlers (ChatGPT-User, Perplexity-User, Claude-Web). These are the agents serving live queries on behalf of buyers. Blocking them is blocking your own discovery surface.
7. Honest reviews beat manicured ones
Agents don’t trust review widgets the way humans do. Gemini and Claude both flag specificity asymmetry: five-star reviews that gush in vague terms while one-stars cite concrete problems. The agents have learned what manicured reviews look like and they discount them.
The fix isn’t to remove negative reviews. The fix is to surface the full distribution honestly. Aggregate rating + review count via JSON-LD with the actual distribution, not a curated subset. Where a product has a real limitation (runs small, requires assembly, ships in two boxes), say so in the description. The agents reward that honesty by recommending you more confidently.
8. Instrument the new surface
If you can’t see your agent traffic, you can’t optimize for it. By BFCM weekend, the bare minimum is: User-Agent logging that surfaces known agent IDs, referrer tracking that distinguishes ChatGPT/Perplexity referrals from generic search, and a way to look at what queries the agent issued before landing on your site.
You don’t need a vendor for the basics. Server logs and a few hours of work in your analytics tool get you the first version. You need it because the BFCM Monday post-mortem is going to ask “where did the lift come from,” and “AI / unknown” without further detail is the answer that ends most merchants’ agentic commerce program before it starts.
The point of the list
None of these are exotic. The hard part isn’t the technical work. Most items take a few hours each. The hard part is doing them now, before the surface you need to optimize for has produced enough revenue to be politically obvious inside your company.
The merchants who win BFCM 2025’s agentic share are the ones running this list this week. Everyone else is going to be running it in February, retroactively, after the postmortem makes them.
Three weeks. Eight items. Worth it.