How AI Agents Decide Which Brand to Recommend (And Why Yours Isn't on the List)

Agent commerce · Ranking factors

Open ChatGPT. Ask it for the best of whatever you sell, under your price ceiling, for your buyer's most common use case. Read the brands it surfaces. Now answer one question: do you know why those brands ranked above yours. If the answer is no, you are operating in the agent era the same way brands operated in 1998. You are assuming the surface is a black box, hoping for visibility, watching competitors with no obvious advantage rank above you.

The agent's ranking logic is not a black box. It is a set of signals. Some are technical, some are editorial, some are reputational. They are knowable, observable, and influenceable. This post lists the nine that move the needle in 2026, sorted by influence, with the audit you can run on your own store for each one. Twenty minutes per signal. Most operators will find at least three they're failing on today.

How the ranking actually works

Every modern shopping agent (ChatGPT, Claude, Gemini, Perplexity, Amazon Rufus) runs the same three-stage pipeline behind the conversational surface. Stage one, retrieval: the agent assembles a candidate set of brands and products that could plausibly answer the buyer's query, pulling from its training corpus, live web search, and commerce APIs like ACP. Stage two, scoring: the agent ranks candidates against the buyer's stated and inferred constraints (price, use case, preferences, history). Stage three, trust: the agent decides which of the top-ranked candidates it is confident enough to actually recommend, with reasoning the buyer can act on.

Each stage has its own signals. Retrieval signals decide whether you are even in the candidate set; if you fail here, none of the downstream work matters. Scoring signals decide your position within the set once you are in. Trust signals decide whether the agent surfaces you confidently or buries you behind hedges. The nine signals below map to one of the three buckets, and the bucket determines what kind of work moves the signal: retrieval is mostly technical work, scoring is mostly reputational, trust is mostly operational.

The 9 signals, ranked by influence

Sorted by how much each one moves agent rankings against a typical 2026 baseline, weighted from published research, vendor documentation, and observation across hundreds of buyer-side prompts. Each signal includes the audit you can run, the work that moves it, and the benchmark that counts as good. Where the public evidence is thin, the section flags it.

01. Review density and recency

The single highest-weighted scoring signal. Research referenced in Bain's analysis of agentic commerce shows agents weight review aggregate and review count as primary confidence inputs when scoring within a candidate set. The same research also finds recency dampens older reviews materially: a 2019 review carries less weight than a 2025 review, even when both are positive. The brand with more reviews from the last ninety days wins the disambiguation tiebreaker against the brand with more total reviews but a stale recency profile.

Audit: Count reviews on your top five product pages versus the same SKUs on Amazon and on your top two DTC competitors. Note the date of the most recent review on each surface. Improve: Ship a post-purchase review flow that fires at the moment the product is most fresh in the buyer's mind (typically 7 to 14 days after delivery for consumables, 30 to 60 for durables). Layer in review-syndication that publishes to your JSON-LD so the count surfaces structurally. What good looks like: if you sell on Amazon, aim for 30 to 40% of the Amazon listing review count on your DTC product page. If DTC-only, 100 or more reviews per top SKU with 20% or more from the last ninety days. Recency matters more than total count once you're past a baseline of roughly 50 reviews.

02. Structured product data completeness

The retrieval signal that gates everything else. Agents read structured product data through JSON-LD blocks embedded on the page (and, for ACP-supporting merchants, through the ACP-format feed Stripe handles). The fields they care about: product name, brand, GTIN, price, availability, image variants, dimensions, materials, shipping window, return policy, review aggregate. Missing fields do not block indexing outright, but they collapse the agent's confidence in your candidacy. The technical baseline for this is covered in detail in the OAI-SearchBot and robots.txt post; this section is the ranking-layer summary of why it matters.

Audit: Run your top ten product URLs through schema.org's structured-data validator. Note completeness per page. Improve: Implement complete JSON-LD Product schema with every field above. For Stripe merchants, enable ACP-format feed support (configuration, not engineering). What good looks like: schema validator returns zero warnings on every product URL in your top decile of revenue, GTINs present where applicable, real-time inventory hook rather than a daily snapshot.

03. Third-party citations from authoritative sources

Agents weight editorial citations on tier-one publishers as canonical evidence that a brand belongs in the candidate set. A Wirecutter recommendation, a New York Times round-up mention, a high-engagement Reddit thread, a category-publication review: these are the citations that show up in the open-web corpus the agent was trained on and the live-web search the agent runs at retrieval time. Self-published content does not substitute for them; the agent's retrieval logic discounts content on your own domain relative to content about you on someone else's.

Audit: Search "best [your category]" on Google. Read the top fifteen organic results. Count how many cite your brand by name. Repeat for "[your brand name] review" and note which publishers come back. Improve: PR cycles into category publications, expert positioning for founder bylines, Reddit/community presence on the actual subreddits buyers in your category use. What good looks like: three or more named citations from tier-one publishers per top-three SKU, with at least one citation per quarter to keep the recency signal alive.

04. Real-time inventory and pricing accuracy

The trust signal that catches operators by surprise. Agents penalize stale data heavily because recommending an out-of-stock or mispriced product damages buyer trust on the agent's surface, which is a cost the agent's reward model attributes back to the merchant. Deloitte's agentic commerce guide flags inventory accuracy as a top-five operational driver of agent recommendation share. A static daily snapshot is no longer good enough; agents reading your feed three hours after a stockout will route around you for the rest of the day.

Audit: Check the cache-control headers on your product feed and sitemap. Check the time delta between your inventory management system updating and the public feed reflecting the change. Check CDN cache TTLs. Improve: Real-time inventory hook (sub-15-minute refresh on stockouts), ETag and Last-Modified headers on the feed, CDN cache TTL capped at one hour for inventory-sensitive surfaces. What good looks like: stockout reflected in the public feed within fifteen minutes, price changes reflected within five minutes, no out-of-stock items surfaced anywhere a crawler can see them.

05. Server-side rendering and crawler accessibility

Agent crawlers read first-paint HTML and do not execute the client-side JavaScript that hydrates a single-page application. If your product name, price, reviews, or return policy only render after JavaScript runs, the crawler sees an empty shell and walks away with nothing usable. This is a retrieval signal because it gates whether the agent can extract structured data from your page at all. Common on older Shopify themes that overuse client-side personalization, and on custom React or Vue stacks that did not adopt SSR.

Audit: Open your top product page in Chrome, disable JavaScript in DevTools, reload. Note what is visible. Repeat with the OpenAI user-agent string set to confirm crawler-eye behavior. Improve: Server-side rendering for product pages, hydration after first paint, no critical content gated behind client-side JS. Next.js App Router gives you this by default; older stacks need explicit SSR work. What good looks like: product name, price, brand, reviews, return policy, and JSON-LD block all visible in first-paint HTML with JavaScript disabled.

06. Brand mention frequency across the open web

The training-data signal. Agents are trained on the open web at scale, and brands mentioned more frequently across more sources get stronger priors when the retrieval logic runs. This is the slowest signal to move and the most compounding once it does, because it operates on the corpus the next generation of models is trained on rather than on the corpus this generation already has. The Stripe and OpenAI ACP rollout and the agent commerce protocols ecosystem provide the structured layer for the agent to transact, but the brand-recognition layer that makes you a candidate in the first place lives in the open-web corpus the agent has internalized.

Audit: Run three searches. (1) Search "[your brand]" plus 'reddit', count threads where you're mentioned in the last twelve months. (2) Open ChatGPT in a fresh session and ask "what are some [your category] brands worth knowing about?", note whether you appear unprompted. (3) Search "[your brand] review" on Google, count distinct authoritative domains in the top twenty results. This is your baseline. The benchmark below is where you want to be in twelve months, not today. Improve: Brand-led editorial calendars, PR campaigns into category publications, podcast appearances, expert positioning on platforms the agent's training corpus weights heavily. What good looks like: hundreds of mentions across at least fifteen distinct authoritative sources, with growth tracked quarterly rather than monthly.

07. Return policy precision and clarity

Vague return policies create ambiguity that the agent passes on to the buyer as hedging. "Returns may apply" or "30-day return policy on most items" reads to the agent as low-confidence information, which lowers the confidence the agent has in recommending you over a competitor whose policy is precise. Precise policies ("free returns within 30 days, prepaid label included, no questions asked") read as high-confidence information and earn the agent's stronger recommendation. This is one of the cheapest signals to fix and one of the highest-leverage on per-recommendation conversion.

Audit: Read your current return policy. Count how many fields are vague ("may," "usually," "on most items") versus specific (numbers, named conditions, named exceptions). Improve: Rewrite to a six-to-eight-sentence policy that answers the buyer's most likely questions without ambiguity: timeframe, who pays return shipping, what condition the item must be in, what categories are exempt. Mirror the precise policy fields in your JSON-LD schema and (if applicable) your ACP feed. What good looks like: zero ambiguous qualifiers in the policy text, full field coverage in structured data, alignment between the storefront page and the feed.

08. Cross-platform consistency

Agents cross-check signals across surfaces. If your Amazon listing says one thing and your DTC product page says another (different title, different price beyond a reasonable delta, different return policy, different shipping window), the agent's confidence in either source drops. This shows up especially in the Amazon dynamic covered in the Amazon problem post: when the agent has to disambiguate between routing the buyer to your DTC site or your Amazon listing, inconsistency between the two pushes the agent toward Amazon's listing as the more authoritative source.

Audit: Pick your top three SKUs. Pull product title, price, shipping window, and return policy from your DTC site, your Amazon listing, your Google Shopping listing, and your Walmart Marketplace listing (where applicable). Note discrepancies. Improve: A single source of truth for product data (a PIM, or even a spreadsheet rigorously enforced) with automated propagation to every surface. Most discrepancies are not deliberate; they are stale updates on one surface that never made it to another. What good looks like: identical product titles across all surfaces, prices within 5% across all surfaces, identical return policy language, same shipping window claim.

09. Response time and uptime

The signal we have the least direct public evidence on, but the reasoning holds. Agents are responsible to the buyer for the quality of the merchant they recommend, and unreliable merchants damage agent trust in the same way unreliable hotels damaged Expedia's trust in early hotel-aggregator scoring. A slow storefront produces buyer frustration the agent absorbs as recommendation-quality cost. A storefront that goes down during a recommendation impression produces worse frustration. Retail Bulletin's reporting on agent ranking flags merchant reliability as one of the operational inputs the largest agent platforms now factor into recommendation eligibility, though specific thresholds are not yet public.

Audit: Test product-page time-to-first-paint on a throttled mid-tier mobile connection. Pull uptime data for the last ninety days from your monitoring tool (or set one up if you do not have one). Improve: CDN coverage for static assets, edge rendering for product pages, uptime monitoring with at least 99.9% target. What good looks like: sub-2-second first paint on a 4G mobile, 99.95% rolling 90-day uptime, no incidents during high-traffic moments traceable to under-provisioned infrastructure.

What to audit this week

If you skip every other piece of this post, do the 30-minute audit below. It surfaces where your brand stands across the nine signals without requiring engineering work, tooling, or budget.

What this changes about marketing in 2026

Marketing in the agent era is closer to SEO in 2010 than to social ads in 2020. Both are optimization games against a ranking algorithm. The difference is volume: agent-driven shopping queries grew 4,700% through 2025 per Retail Bulletin's reporting, while the number of brands actively optimizing for agent rank is still measured in the low thousands. The competitive-density-to-volume ratio in 2026 looks like Google in 2005. The 2026 brands-stop-advertising-start-answering post covered this as a thesis; this post is the operator playbook.

The window for asymmetric advantage is twelve to twenty-four months. Brands that ship the nine-signal work now compound across the next ten years the way brands that figured out backlinks in 2008 to 2010 compounded across the decade after. The signals consolidate slowly. The compounding compounds. The brand that audits this week and ships against the audit by next quarter has a structural lead that the brand starting in 2028 cannot close without paying a meaningfully larger acquisition tax.