Recommendation prompts, the "what should I buy" class, are where AI answers concentrate hardest, and concentration is the economics of this whole chapter. In OppAlerts' study, the top five brands in an industry hold 49% of all AI answer share, against 24% for the top five in Google organic across the same 100 industries.From the OppAlerts ranking-factors study: 403,000 prompts, 100 industries, 13 models. Method and tables in The Research Behind This Guide, part of the AI Search Visibility research. An answer that names three to five products replaces a results page that listed ten, so the channel pays roughly twice as much of its attention to the winners, and near zero to everyone else.
That would be grim news for challengers if the shortlist were one list. It is many lists, and that is the opening.
How the shortlist gets picked, and how personas reshape it
Mechanically, a recommendation answer assembles the way Model Memory and Live Retrieval describes: the model starts from its memory of the category, and if search runs, retrieved pages rewrite part of the list. Both stages compress. Memory favors the brands most represented in training data; retrieval reads a handful of top-ranked pages, mostly roundups and comparisons; and the final answer keeps only the few names that both stages agree deserve slots, plus whatever the retrieved pages argued strongly for. In the memory-versus-search comparison, 44.1% of brands in an average searched answer were absent from the same model's memory-only answer, so the retrieved pages are doing real selection, not decorating a fixed list.
The persona wedge is the second force. Add who is asking to the prompt ("for a college student", "for a contractor who tows") and the shortlist can rewrite wholesale. In OppAlerts' persona analysis, taste-driven categories like apparel rewrote most of the brand list when the persona changed, while utility categories moved far less, and category giants routinely collapse to near-zero for specific personas that a niche brand owns outright. The aggregate leaderboard and the persona leaderboards are different competitions. A challenger that cannot crack the generic top five can still be the answer for the personas that map to its actual customers, and Prompt and Topic Research shows how to find out whether your category is one where personas move the list.
Pick the personas where you have a genuine product edge and make yourself easy to justify there: pages that say, in plain sentences, which buyer you are for and why. That is the challenger's wedge, and in taste-driven categories it is often the only realistic entry.
Machine-readable product truth
When a model composes a product recommendation, it needs specs, price, and availability to reason with, and it gets them from what it can retrieve and parse. Product data completeness becomes visibility work. The checklist: current prices on crawlable pages rather than behind scripts, spec tables in real HTML, availability stated, Product structured data with price and availability fields filled, and a maintained product feed where platforms accept one. Structured Data and Machine-Readable Facts covers the markup layer; Technical Accessibility for AI Crawlers covers making sure the AI crawlers can fetch any of it.
The failure mode is a product page that renders beautifully for humans and parses as marketing prose for machines. A model that cannot find your price will either omit you from a "best under $100" answer or guess, and a guessed price attached to your brand in a purchase conversation is worse than absence.
Comparison and alternatives pages
When recommendation prompts trigger search, the generated queries look like "best project management software for small teams" and "X vs Y", so the pages retrieved are overwhelmingly comparisons, roundups, and alternatives pages. Two moves follow. Earn slots on the third-party roundups that rank for your category's versions of those queries, which is link-and-PR work covered in Links: The Signal That Refuses to Die. And build your own: honest "X vs us" and "alternatives to X" pages, with real spec-level comparisons, are retrievable documents in exactly the format the query fan-out hunts for. Write them factual enough that a model can quote them without repeating ad copy; Writing for Retrieval and Citation covers that discipline.
This inherits from classic SEO more than any other part of the book: the comparison pages that win these retrievals are usually the ones that already rank. In the study data, 81% of AI top-10 recommendations also rank in Google's organic top 10 for the category, so ranking remains the entry ticket; the filtering that happens after retrieval is what the rest of this chapter influences.
Agent-ready commerce
The next step past recommending is transacting. OpenAI's Instant Checkout lets U.S. ChatGPT users buy from Etsy sellers directly in chat, with over a million Shopify merchants announced as coming next, single-item purchases first and multi-item carts on the roadmap.OpenAI, Buy it in ChatGPT (September 2025), announcing Instant Checkout and the Agentic Commerce Protocol. The rails underneath are the Agentic Commerce Protocol, an open standard OpenAI codeveloped with Stripe for letting an AI agent transact against a merchant's existing commerce backend.Stripe, announcement of Instant Checkout and the Agentic Commerce Protocol, which Stripe operates the payments layer for.
What a buying agent needs from you is the machine-readable product truth above, plus transactional plumbing: accurate structured product data, live pricing and stock, and support for the agent protocols your platforms adopt as they mature. The assessment worth holding: agentic buying is early, the announcements above are the first infrastructure, and the brands with clean product data will be default-compatible with each new agent surface while everyone else retrofits. Watching this space is part of The Road Ahead: Agents and What Comes Next.
Priorities for this arena, in order: be retrievable and parseable on the pages recommendation queries fetch, win the personas where you have a real edge, then wire up for agents as the protocols settle. The concentration numbers say the aggregate shortlist is nearly closed; the persona lists and the agent era are still open.