How AI Buyer Agents Make Purchase Decisions

AI buyer agents don't shop the way people do. They pull structured data off a product page, run it against a checklist, and score whatever survives. By the time a human sees a recommendation, the brand's odds were already locked in, and most brands don't even know the vote happened.

Human shopping runs on feel: a hero image, a layout that reads as trustworthy, a stack of five-star reviews nobody actually reads before trusting anyway. An agent doesn't register any of that; it pulls attributes off the page, checks them against a stated or inferred requirement, and scores what's left, leaving the lifestyle photo shoot to go unnoticed entirely.

Here's the order an agent actually reads in: schema markup and structured fields first, then offer data like price, stock, and shipping terms, then review aggregates and Q&A content. Only after that does it bother with the paragraph of marketing copy, and usually just to settle a conflict somewhere else on the page. Deloitte has described agentic commerce as compressing what used to be days of comparison shopping into something closer to instant evaluation, so a brand's window to make an impression is now milliseconds of parsing, not minutes of scrolling. A page built to convert a human, gorgeous photos, warm copy, can still score badly if the data underneath is thin or broken, because good marketing and good machine-readability are two different jobs now. Treat them as one job and invisibility to the systems actually doing the shopping follows.

The structured signals AI buyer agents actually rank on

Agents don't weigh everything at once. They work down a list, and the first pass is elimination rather than ranking.

Tier one is pass or fail. Out-of-stock items get dropped before scoring even starts, and a product missing a GTIN or other standard identifier can fail to resolve to the right item across databases, which is enough on its own to knock it out of consideration. Price and shipping terms need to sit in a field the machine can read; if that number only lives inside a banner image or a PDF spec sheet, the agent can't see it, and the product never clears the gate.

Tier two is where the real ranking happens, among whatever survived tier one. Review scores and review volume carry weight, and agents seem to read sentiment rather than just average the stars. Attribute completeness matters just as much: material, dimensions, compatibility, intended use. The more fields filled in, the more the agent has to match against. Third-party citations count too, expert reviews, editorial write-ups, community threads, and Reddit shows up constantly in research on where large language models pull product answers from. Platform endorsement badges, an "Overall Pick" label and things like it, carry real weight, and Columbia Business School's ACES framework found agents systematically favor these organic platform endorsements over anything tagged as sponsored.

Tier three only kicks in when products are still tied after all that: how often the brand gets mentioned across the open web, and how specific the product description is when the structured fields are otherwise a wash. Research on the mention-frequency point has found branded mentions correlate with AI visibility far more than backlinks do, a different game entirely from the one SEO people spent twenty years learning.

The tiers don't carry equal weight everywhere, either. Perplexity leans hard on third-party editorial citations, Amazon's Rufus draws mostly from Amazon's own reviews and Q&A, and ChatGPT Shopping pulls from Google's product feed, with Microsoft Copilot tracking pricing through Bing. The tier-one gates, though, stay close to universal across all of them. A brand that only tunes its Google Merchant Center feed will look great to some of these systems and simply won't exist to the rest.

The position bias and instability problems that make agent rankings less predictable than search rankings

Columbia Business School's ACES framework, published at the ACM Web Conference in 2026, is the most rigorous peer-reviewed look at agent purchasing behavior out there right now. Its findings cut against the comfortable idea that agents behave like clean, rational optimizers.

The headline result is choice homogeneity: demand piles onto a small set of "modal" products while comparable alternatives get ignored almost completely, even when there's no real quality gap between them. Any brand assuming a good product eventually earns its fair share of agent recommendations should look hard at that finding, because the data doesn't back it up.

Then there's instability, and it's worse. Agent preferences shift with model versions, not with anything the brand actually did. A product ranking well under one model release can fall out of favor after the next update with the underlying catalog data completely unchanged, and market share among recommended products can reshuffle without one single brand touching its own listing. Walmart's CTO has publicly flagged concern about third-party agents that route around traditional merchandising and advertising signals altogether, a notable thing to admit from a company that spent decades mastering exactly those signals.

There's a sponsored-content problem too. Agents avoid anything flagged as sponsored while favoring organic platform endorsements, so traditional paid advertising may hurt AI visibility instead of helping it. Position bias is real, it varies by provider and model version, and it shows up even in headless, text-only interfaces where there's no visual real estate to fight over in the first place. No brand can buy its way into an agent recommendation the way it buys a paid search slot, which leaves the data layer and the reputation layer as the only ground left worth building on.

Why catalog data quality is the competitive surface that determines AI eligibility

Ranking well on Google tells you almost nothing about whether AI systems will cite a brand. A meaningful share of brands with strong organic search presence get zero AI citations, which tells you the two forms of visibility have come apart from each other entirely.

Catalog enrichment, done right, is its own discipline, separate from feed optimization. It means filling in missing attributes, fixing data that's inconsistent across SKUs, and writing descriptions structured so a machine can parse them, not just so a human finds them pleasant to read. The goal is a product record an agent can identify with confidence, compare against alternatives, and recommend without guessing at whatever's missing.

The minimum schema for reliable citation across the major AI shopping surfaces: Product, Offer, AggregateRating, Review, BreadcrumbList, FAQPage, and Organization markup. Well-formed schema earns citation not just through search engines but through the language models themselves, which reference schema documentation directly when they build an answer about a structured product. Across the different AI shopping channels, Google Merchant Center feed quality paired with Schema.org Product attributes forms the shared baseline. For most DTC brands, that's the highest-leverage place to start, full stop. Kinect, a DTC-focused AI sales and catalog intelligence platform, structures this data layer so both on-site agents and outside AI surfaces can read it.

Branded mentions matter here too, and not as some brand-awareness side project. Research has found they correlate with AI visibility far more than backlinks do, which means PR placements, community presence, and editorial coverage now function as catalog strategy as much as brand strategy. Research on generative engine optimization adds another wrinkle: content with statistics and cited sources gets a real lift in generative engine visibility, because specificity and sourcing make a claim easier to cite. None of this stays fixed, since models update, new surfaces launch, and a catalog that passed muster this quarter can drift out of compliance by the next release. Enrichment has to run as a standing job, not a project with an end date.

How multi-agent systems and buyer-agent delegation change the decision sequence

The first version of agentic commerce was simple: one agent, one query, one recommendation. That's already getting replaced by something more layered.

Multi-agent negotiation is showing up now, where the outcome of one completed transaction feeds the parameters of the next. Agents increasingly optimize toward a buyer's broader goals across multiple purchases and a longer stretch of time, rather than treating each query like an isolated event. Trust in these systems keeps deepening too, and shoppers are handing agents actual purchasing authority, not just the discovery work, so the human might never look at the shortlist at all.

The infrastructure for this already exists. OpenAI launched a protocol with Stripe that lets agents transact on a buyer's behalf, and Google released a protocol standard for agents to discover and talk to each other. Shopify's agentic storefronts let checkout happen inside an agent's own interface, no need for the buyer to ever land on the brand's site, and Perplexity's Shopify integration makes eligible merchants automatically discoverable, no separate onboarding needed.

Here's the hard part: a brand that isn't structured to be read by an agent can get excluded before a human ever makes a delegation decision, and the recommendation shows up already filtered by the time anyone sees it. The paid-advertising problem gets worse in this setup, too, because in a multi-agent system the agent optimizes for the buyer's stated criteria, not for anyone's ad revenue.

What "Share of Model" means and why it replaces traffic as the primary visibility metric

Standard web analytics fall apart in a world where AI sits between the buyer and the brand. A shopper asks an assistant which brand to buy, gets an answer, then types the brand name straight into a browser bar, and that visit logs as direct traffic. The agent's role in the decision vanishes from the data, and whatever brand equity got transferred in that exchange stays invisible to whoever's reading the dashboard.

"Share of Model," or SoM, is the metric built to fill that gap: how often a brand gets named, cited, or recommended by major AI systems in response to the kinds of questions a real buyer would type. You measure it by running purchase-intent prompts through these systems and auditing which brands come back, how prominently, and with what framing. It's replacing Share of Voice as the main visibility benchmark because it captures the moment the decision actually gets made, not the moment traffic happens to land somewhere.

Branded mention frequency drives SoM more than backlink count does, consistent with the finding that mentions correlate far more strongly with AI visibility. SoM gets built through reputation, not link equity, which marks a real departure from how SEO people have thought about authority for a couple decades now.

Vendor-reported "AI-influenced revenue" figures deserve some skepticism. Those numbers typically count any order where an AI touchpoint happened anywhere in the journey, a much looser bar than direct attribution, and the gap between the two methods is wide enough to change the entire business case. Meanwhile, research on DTC marketing priorities shows brands shifting hard toward profitability metrics, conversion rate, customer acquisition cost, lifetime value, average order value, over raw traffic or visibility counts. SoM fits that shift well, since it ties back to conversion quality instead of reach for its own sake.

What DTC brands need in place to be chosen by agent systems, and where to start

The order follows straight from everything above. Fix the tier-one gates first: availability data, complete GTINs, offer fields a machine can actually parse. Brands that fail here get excluded before ranking even starts, so nothing downstream matters until this part is solid.

From there, build out schema completeness, Product, Offer, AggregateRating, Review, BreadcrumbList, FAQPage, Organization, since that's the foundation covering the widest range of AI surfaces. Next is Google Merchant Center feed quality, which feeds several AI shopping channels at once. Only after that does it make sense to onboard platform-specific channels: Stripe's and Shopify's commerce protocols cover ChatGPT and Copilot, and Perplexity eligibility comes free to Shopify merchants already.

This never really finishes. Model updates reshuffle agent preferences on their own schedule, so a catalog that's fully compliant this quarter might need a fresh audit after the next major release. The brands that survive that instability treat catalog quality as a standing job, not a one-time migration.

On-site AI is the one layer a brand still fully controls. Agent-referred traffic converts at meaningfully higher rates than organic traffic across multiple datasets, and capturing that intent well means building an on-site AI experience that matches structured product data to whatever the buyer actually asked the agent. That's the piece that closes the loop between getting discovered and getting bought.

For most DTC operators, the real problem is execution. Catalog data, review content, policy details, and brand voice all need to live in one well-maintained place that both on-site AI and outside agents can reach, not scattered across five tools that each need their own upkeep. Speed matters more than it looks like it should, too: brands building citation presence and clean structured data now are the ones stacking up the branded-mention history that drives AI visibility later. Delay, and the gap widens with every new model release, while the brands that got their storefronts AI-ready early keep pulling further ahead on conversion.