Whitepaper · September 2026
From AI answers to sales
How AI assistants became a front door for buying, what brands can measure today, and how to connect AEO work to leads, deals and revenue.
In eighteen months, AI assistants went from answering questions about products to taking orders for them. ChatGPT, Google’s AI Mode and Gemini, Microsoft Copilot, Perplexity, Amazon’s shopping assistant and, this month, Meta’s Muse all now help buyers compare products, and several can complete a purchase without the buyer ever opening a store’s website.
For brands, the easy reading of this is “get mentioned by AI.” The data points somewhere more useful. AI-referred shoppers are still a small share of traffic, but they arrive further along and buy more often. Buyers use AI to narrow the field, then check what it told them — on the brand’s own site, with a dealer, or with a sales rep. The sale still closes with the merchant. That makes AI visibility an input to sales, not a result in itself.
Six findings frame the rest of this paper:
The practical conclusion: treat the buying opportunity — a buyer question your product information doesn’t yet answer — as the unit of work, act on it in every channel where the product is sold, and follow it past the answer to inquiries, deals and revenue.
The change happened in public, one announcement at a time. The table below lists the milestones that matter most for brands. The pattern: first assistants learned to show products, then to check out, and in 2026 the industry began to argue about where checkout should happen.
| Date | What changed | Where the purchase completes |
|---|---|---|
| 29 Sep 2025 | OpenAI launches Instant Checkout in ChatGPT and the Agentic Commerce Protocol, built with Stripe11 | Inside the assistant |
| 13 Nov 2025 | Google’s agentic checkout goes live in the US; shopping comes to the Gemini app12 | Google completes the purchase on the merchant’s site |
| 25 Nov 2025 | Perplexity makes shopping free for US users, with Instant Buy through PayPal13 | Inside the assistant |
| 8 Jan 2026 | Microsoft launches Copilot Checkout in the US14 | Inside the assistant |
| 11 Jan 2026 | Google introduces the Universal Commerce Protocol (UCP), co-developed with Shopify, Etsy, Wayfair, Target and Walmart8 | Inside AI Mode and Gemini |
| 24 Mar 2026 | OpenAI lets merchants use their own checkout and focuses on product discovery7 | Hand-off to the merchant |
| 24 Jun 2026 | Salesforce makes a B2B Buyer Agent generally available, with cart-to-quote negotiation15 | The brand’s own B2B commerce, in messaging channels |
| 8–24 Sep 2026 | Meta launches Muse, a personal agent that can check out; Shopify announces Shop Pay checkout for Muse16 | Inside the agent |
| 16 Sep 2026 | Google: UCP “already enables direct checkout for hundreds of thousands of brands and retailers”6 | Inside the assistant, or cart transfer to the merchant |
The audiences are large. Alphabet reports more than 1 billion monthly users for AI Mode and 950 million for the Gemini app.17 Amazon says its shopping assistant Rufus was used by more than 300 million customers in 2025 and helped deliver “nearly $12 billion in incremental annualized sales.”18 These are company-reported figures, and each describes an assistant inside one company’s own ecosystem.
The traffic that reaches brands is smaller, and growing. Adobe, which analyses more than a trillion visits to US retail sites, measured AI-referred traffic up 693% in the 2025 holiday season and 62% year over year in July 2026, while noting that the base “remains modest.”1,2 Across 1,215 enterprise domains, Conductor found AI referrals made up about 1% of all visits.19 What has changed faster than volume is quality: AI-referred visitors to US retail sites went from converting 23% worse than other traffic during Prime Day 2025 to 40% better during Prime Day 2026, and 60% better in July 2026.2,20
Buyers are not handing over the decision. Gartner found only 11% of US consumers willing to let AI make a purchase decision, even in low-stakes categories.4 They use the assistant to narrow the field and learn the tradeoffs — and in BCG’s survey of more than 13,000 consumers, AI introduced a brand the shopper would not otherwise have considered in more than half of AI-assisted purchase journeys.21 That is the opening, and the risk: the assistant decides which brands get considered, using whatever facts it can find.
Once an assistant has a shortlist, the purchase can end in one of three places. Brands rarely choose just one; the same product may sell through all three.
Google, Microsoft, Perplexity and now Meta let buyers complete a purchase without leaving the assistant. Google, Microsoft and Perplexity each say the merchant stays the merchant of record — in Google’s words, “the retailer always remains the merchant of record.”8,13,14 What the merchant loses is the page: the buyer never sees your product page, your cross-sells or your loyalty prompts, and on some surfaces your analytics don’t fire. Shopify notes that Google Analytics and custom pixels “won’t fire in Meta’s direct checkout,” although the orders appear in Shopify attributed to Meta.22
In March 2026 OpenAI said its first version of Instant Checkout “did not offer the level of flexibility that we aspire to provide,” and moved to letting merchants use their own checkout while it focuses on product discovery.7 For Shopify stores, ChatGPT users now complete the purchase on the store’s own checkout in an in-app browser.23 Here the product page, identity and loyalty still apply — and the facts on that page have to match what the assistant just said.
For considered purchases — a car, a home, equipment, a B2B contract, a mattress for a 200-room hotel — the assistant’s job ends with an introduction. AutoTrader and Zillow run apps inside ChatGPT that hand buyers to dealers and agents; Meta’s Business Agent can “qualify incoming leads” for businesses in messaging; Salesforce’s B2B Buyer Agent handles procurement, including cart-to-quote negotiation, over channels such as WhatsApp and SMS.15,24,25 Shopify states plainly that “B2B-only products aren’t supported for agentic storefronts.”23 In this path, AI shapes the shortlist and a person closes the deal. Gartner’s finding that 69% of B2B buyers prefer to validate AI-generated insights with sales reps describes exactly this hand-off.5
Not every door is open. Amazon’s Conditions of Use now include Agent Terms under which no agent may use Amazon’s services unless it identifies itself.26 Channel rules are part of the landscape, and they change.
What the three paths share: the facts the assistant repeats — specifications, conditions, prices, policies, the evidence behind claims — come from the brand, its channels and its partners. A wrong fact costs more than a missing one; a fast checkout only makes a wrong answer arrive sooner.
“Are we showing up in AI?” is a reasonable first question and a poor last one. Between being named in an answer and closing a sale there are at least five distinct events, and each needs its own evidence.
| Layer | What it tells you | How to observe it today | Watch out for |
|---|---|---|---|
| 1 · Mentioned | Whether answers name your brand and model | Answer checks run through model APIs; platform reports such as Google’s share-of-voice view across AI Mode and AI Overviews6 | API answers can differ from consumer apps. A mention is not a recommendation. |
| 2 · Cited with a link | Which pages the answer relies on — yours or someone else’s | Answer checks that keep the source links | McKinsey found a brand’s own sites make up only 5–10% of the sources AI search references.27 |
| 3 · Visit | Whether answers send buyers to you | AI referrals in your analytics (e.g. GA4) | Misses zero-click journeys and off-site checkout. When Google showed an AI summary, users clicked a result in 8% of visits versus 15% without one.28 |
| 4 · Inquiry or deal | Whether buyers act on what they learned | CRM records, quote forms, dealer and sales notes | Buyers rarely volunteer where they started. Ask at intake which assistant they used and what they asked. |
| 5 · Sale or contract | Revenue | Order and contract records | Before-and-after is not incrementality. Keep attributed, influenced and incremental revenue separate; never add them together. |
Three rules keep this honest. Measure each layer on its own terms — a rising mention count does not prove more sales, and a sales spike does not prove the answers changed. Hold the question set steady when you compare over time: same buyer questions, same models, same products. And write down what you cannot see: zero-click journeys, purchases completed inside an assistant, and deals where nobody recorded how the buyer found you.
A buying opportunity is a kind of buyer need that your product information does not yet answer well: a question about fit, conditions, tradeoffs or proof that an assistant, a shopper or a sales rep cannot resolve from what you publish. It is a market-level idea. A deal is different: one customer’s purchase in progress. Keep the two separate, and link them.
Buying opportunities are found by working from the buyer’s side:
Illustrative example · sample product, not a customer result
| Buyer question | What the brand publishes | State | Action |
|---|---|---|---|
| Will it sleep cool for a hot sleeper who sleeps on their side? | “Cooling gel foam” | Unconfirmed | Publish the cooling test method and result for this model, or narrow the claim |
| Is it firm enough for a back sleeper over 230 lb? | Firmness 6/10; no weight guidance | Missing | Add weight guidance and support-layer specifications |
| How long is the trial, and who pays return shipping? | Brand site: 100 nights. Retail partner listing: 30 nights | Conflicting | Reconcile the policy, or state clearly how it differs by retailer |
| Does it limit motion transfer for couples? | “Great for couples” | Unconfirmed | Cite a motion-isolation test for this model, or remove the claim |
None of these fixes is exotic. What makes them valuable is that each maps to a real buyer question, and that one fix — a published cooling test, say — can serve several models in a portfolio while each model keeps its own specifications and evidence.
AI assistants read across everything published about a product: your pages, marketplace listings, retail partners, feeds and catalogs, reviews. An established brand’s product exists in many of these at once, and they drift apart.
Answer engine optimization (AEO) — also called generative engine optimization (GEO) — is the work of getting products understood and chosen in AI answers. On its own, it stops at the answer. For an established brand, most revenue still closes somewhere the answer can’t see: on the brand’s checkout, at a retailer, with a dealer, or with a sales rep. Connecting the two turns AEO from a visibility project into a sales discipline.
A brand can start this today with the systems it already has:
The feedback runs both ways. Buying opportunities create deals; deals show which opportunities are worth the work. The questions sales teams hear are the most reliable list of buying opportunities a brand has — and today they rarely reach the people who write the product information.
| Weeks | Goal | What you do | What you have at the end |
|---|---|---|---|
| 1–3 | Baseline | Pick the 10–20 products that matter most to revenue. List the buyer questions for each category. Record what assistants answer, which sources they cite, and your AI referral visits. Ask sales which questions they hear most. | A dated baseline and a list of buying opportunities |
| 4–8 | Fix | Resolve the highest-value opportunities: publish missing facts, reconcile conflicts across channels, back or narrow unconfirmed claims. Approve and publish channel by channel. | An approved change log per product and channel |
| 9–12 | Measure and connect | Re-run the same questions on the same models. Compare referrals. Start tagging inquiries and deals with the buyer question. Decide the next round. | A before-and-after on a fixed question set, and a first link between opportunities and deals |
Hasmord is the AEO-native sales platform. Today it finds the buying questions your products can’t yet answer, by category; checks the evidence behind your claims; drafts channel-specific improvements your team approves — product page copy, an Amazon listing export, and on paid plans a Hasmord-hosted page written for AI answer engines (it does not write into your store); re-runs AI answer checks through the Anthropic, Google, OpenAI and Perplexity APIs; and reads AI referral visits from your Google Analytics 4.
Lightweight lead-to-contract management — tying inquiries and deals to the buying questions and products behind them, and feeding sales results into the next round — is (coming in v3.0).
Talk to us: info@hasmord.com · Try it on your own products: www.hasmord.com
Figures are quoted as their publishers state them and were checked against the original page on 25 September 2026. Adobe publishes growth rates and ratios, not absolute AI visit volumes, and has revised some earlier series; we use its latest releases and do not compare figures across releases. Survey results depend on how each question defines “AI,” so adoption figures from different studies should not be compared directly. Company-reported figures (Alphabet, Amazon) describe each company’s own products. The product example in section 4 is illustrative. This paper does not report Hasmord customer results.