Ask ChatGPT for the best running shoe under £120, or tell Perplexity to find you a reliable robot vacuum, and you will get back two or three named products — not ten blue links. Those answers are not random, and they are not advertising. In 2026 the AI agent has quietly become the most consequential product-recommendation surface since the Google shopping carousel, and the brands that get named are pulling away from the brands that do not.

The numbers behind that shift are no longer speculative. AI referral traffic to retail sites surged 693% year-over-year during the 2025 holiday season, and those AI referrals converted 31% higher than non-AI sources (Adobe, 2026). Klaviyo’s 2026 AI Consumer Trends Report found that 39% of consumers had bought a product on the strength of an AI recommendation in the previous six months. Perplexity’s product-recommendation queries grew roughly fivefold from late 2024, and 65% of consumers now expect AI-powered shopping to be a standard part of the experience. This is not an emerging channel any more. It is a primary one.

And yet almost everything written about it stops at the same place: optimise your product feed, add some schema, enrol in a merchant programme. That advice is correct and incomplete. It explains how to make your products eligible to be recommended. It does not explain why, when an agent has fifty eligible products to choose from, it names yours. That gap — between eligibility and selection — is where link building lives, and it is the gap this guide is built to close.

The thesis of this guide:  Feeds and schema get you into the room. Earned, off-site authority — citations, reviews at scale, brand entity strength, and third-party validation across the web — is what gets you picked once you are in it. The teams treating AI product recommendation as a pure feed-engineering problem are competing on a level playing field with everyone else who read the same tool-vendor blog post. The teams treating it as an authority problem are winning the recommendations those feeds make possible.

This is the foundational, hub-level resource for everything Link Building Journal publishes on AI commerce. We will cover how the agents actually work, the ranking signals that decide recommendations, a five-stage playbook you can execute this quarter, the measurement framework, and the mistakes that quietly keep brands invisible. Where a sub-topic deserves its own deep dive — feeds, schema, citations, marketplaces, decision tools — we link to the dedicated article rather than repeating it here.

What this guide covers

• How AI shopping agents work in 2026 — the discovery, reasoning and recommendation pipeline
• The agentic commerce protocols (ACP and UCP) and what the Instant Checkout retreat actually changed
• The seven ranking signals that decide which products get recommended — and which ones link builders own
• The five-stage playbook: eligibility, authority, citations, reviews velocity, and measurement
• A 30/60/90-day execution plan and the metrics that prove it is working
• The eight mistakes that keep otherwise-strong brands out of AI recommendations

How AI shopping agents actually work in 2026

An AI shopping agent is a system that takes a buyer’s request expressed in natural language — “a quiet dishwasher for a small kitchen, under £500, good warranty” — and returns a small set of specific product recommendations, often with reasoning, pricing, images and a path to purchase. The major surfaces in mid-2026 are ChatGPT Shopping (hundreds of millions of weekly users), Perplexity Shopping, Google AI Mode and Gemini, Microsoft Copilot, and Amazon’s Rufus inside the Amazon walled garden.

It helps to separate two things these agents do, because they draw on different signals. The first is discovery and eligibility: building the candidate set of products the agent will even consider. This is fed largely by structured product data — merchant feeds, schema, marketplace catalogues. The second is reasoning and selection: choosing which of those candidates to actually name, in what order, with what framing. This second step leans heavily on signals that live off your product page entirely — reviews, third-party reviews and “best of” lists, brand recognition, pricing consistency across the web, and the agent’s confidence that the brand is a real, trustworthy entity.

Why this distinction matters for link builders:  Feed optimisation maxes out the first step. But two competitors can have equally perfect feeds. The tie-breaker is the second step — and the second step is overwhelmingly an off-site, earned-authority game. That is the entire reason a link building publication has a strong, defensible position on a topic everyone else treats as a feed-management task.

How the agents differ — and why you cannot optimise for just one

Each engine selects sources differently, which is why a brand can be recommended confidently by Perplexity and invisible in ChatGPT for the same query. Perplexity performs real-time web searches rather than relying on training data, pulling current pricing, availability and reviews, and it shows its sources — which makes earned third-party citations directly visible in the answer. ChatGPT blends merchant feed data, web content and brand entity recognition, surfacing product cards inside the conversation and (after its Instant Checkout retreat) redirecting buyers to the merchant to complete the purchase. Google AI Mode draws on the Merchant Center feed and the wider Google index, and Amazon’s Rufus operates almost entirely inside Amazon’s own review, Q&A and pricing ecosystem.

The practical implication is parity. You are not optimising for one agent; you are building the kind of broad, consistent, web-wide authority that every agent can find and trust. We cover the platform-by-platform specifics — including how each engine weights its inputs — in our companion guide on how ChatGPT, Perplexity and Gemini choose which products to recommend.

Agent / surface	Primary input it trusts	Where link builders have leverage
ChatGPT Shopping	Merchant feed + web content + brand entity recognition	Brand entity strength; cited “best of” coverage; review depth
Perplexity	Live web search with visible source citations	Earned citations in buying guides and reviews — directly visible
Google AI Mode / Gemini	Merchant Center feed + Google index + UCP	Editorial coverage that ranks; consistent NAP/brand data
Microsoft Copilot	Bing index + merchant data + enterprise sources	Authoritative third-party coverage indexed by Bing
Amazon Rufus	Amazon reviews, Q&A, pricing within Amazon	Review velocity and Q&A inside the marketplace

Because the marketplace surfaces (Rufus especially) play by different rules to your owned site, we treat that split in detail in marketplace vs owned-site visibility in AI shopping, including when to invest in Amazon authority versus your own domain.

The plumbing: ACP, UCP and the Instant Checkout retreat

Two open standards now define how agents transact with merchants. The Agentic Commerce Protocol (ACP), built by OpenAI and Stripe, defines how product feeds are formatted, what a checkout session looks like, and how payment is delegated. The Universal Commerce Protocol (UCP), co-developed by Google and Shopify and launched at NRF in January 2026, covers the full journey — search, discounts, payment, order status — across Google AI Mode and Gemini, and is backed by Walmart, Target, Etsy, Wayfair and 60-plus payment networks.

The single most misread event of the year is OpenAI’s retreat from Instant Checkout. Launched with Etsy sellers in September 2025, in-conversation Instant Checkout was scaled back in March 2026: only around 8% of US ChatGPT adults tried it, roughly a dozen Shopify merchants integrated, and OpenAI never built sales-tax collection or fraud prevention. ACP pivoted away from in-chat checkout toward product discovery, with purchases redirecting to the merchant.

The strategic takeaway:  Transaction infrastructure is a solved problem — both protocols work, and on Shopify the integration is close to trivial. Discovery is the unsolved problem. The war was never about checkout; it is about whether an agent can find and trust your products in the first place. That reframes the entire investment thesis: spend on being discoverable and recommendable, not on bleading-edge checkout integrations that may be deprecated within two quarters.

For link builders this is liberating. You do not need to bet on which protocol wins, and you do not need deep engineering to participate in the part that matters. The discovery layer is built from feeds (a one-time engineering task) plus authority (an ongoing earned-media task). The deep mechanics of feed formatting for both protocols sit in our companion piece on product-feed optimisation for LLM recommendations, which goes well beyond Google Merchant Center.

The seven signals that decide which products get recommended

Across the public research, the same signals appear again and again as the inputs agents weigh when choosing among eligible products. Here they are, ordered roughly by how much control a link building and digital-PR function has over each. The pattern is the point: the signals competitors obsess over (feed and schema) are the ones link builders influence least, and the signals competitors ignore are the ones link builders own outright.

#	Signal	What it is	Link-builder control
1	Brand entity strength	Whether the agent recognises your brand as a real, distinct, trustworthy entity	High
2	Third-party citations	Mentions in “best of” lists, reviews and buying guides the agent reads	High
3	Review depth & velocity	Volume, rating and recency of reviews on- and off-site	High
4	Editorial coverage that ranks	Press and authoritative content that surfaces in the underlying index	High
5	Pricing & policy consistency	Consistent price, shipping and return data across the web	Medium
6	Structured product data	Feed completeness + Product/Offer/AggregateRating schema	Low
7	Crawlability	Whether AI crawlers can actually read the page (server-side rendering, robots.txt)	Low/Shared

1. Brand entity strength — the signal nobody is building deliberately

Before an agent recommends “Brand X running shoes,” it must understand that Brand X is a real entity, what it makes, and whether it is reputable. Where that understanding comes from is consistent naming, authoritative citations across the web, and clear brand information in knowledge graphs. This is precisely the entity layer that underpins every AI citation — and it is built through earned mentions, not feed fields. A brand that is named consistently across dozens of independent, authoritative sources reads to an agent as a confident, low-risk recommendation. A brand that exists only on its own domain and a marketplace listing reads as a gamble the agent would rather not take.

This is classic link building work pointed at a new target. The same digital-PR muscle that earns editorial coverage builds the entity. If you are new to the discipline, our hub on 15 link building strategies that actually work in 2026 is the right starting point; everything in this section is those strategies aimed at AI commerce rather than at rankings.

2. Third-party citations — getting named in the guides agents read

When Perplexity recommends three products, it cites where those recommendations came from — and those sources are overwhelmingly independent “best of” lists, expert reviews and comparison articles. ChatGPT and Copilot lean on the same body of content even when they do not display it. The corollary is blunt: if your product is not named in the buying guides the agent reads, it is hard for the agent to name it to a buyer. Earning placements in category round-ups, comparison pieces and reviewer content is therefore one of the highest-leverage activities in all of AI commerce.

This deserves its own playbook, and it has one: earning citations in AI buying guides and comparison answers covers the outreach, the formats reviewers favour, and how to become the default named option in your category.

3. Review depth and velocity — recency is a ranking factor

Reviews do double duty. Three in five shoppers hesitate to buy a product with no reviews (ChannelEngine), and agents inherit that caution — they weight review count and average rating heavily when deciding which product to recommend. But the subtler signal is velocity. AI surfaces reward freshness far more aggressively than classic Google does, and a steady stream of recent reviews tells an agent the product is currently relevant and currently good, not merely historically popular. A product with 400 reviews and none in six months can lose to a rival with 120 reviews and forty in the last month.

Feed-level fields matter here — the ChatGPT feed specification accepts product_review_count and average_rating, and free-shipping and return-policy fields feed the trust signals agents surface (91% of shoppers say free shipping influences purchase completion, per ChannelEngine). But the strategic work is engineering a sustainable review-generation cadence, which we treat as its own subject in the Cluster AI deep dive on review velocity and AI trust.

4–5. Editorial coverage and pricing consistency

Pages that already rank in Google’s top results have a strong head start in AI citations, though the relationship is not automatic. Authoritative editorial coverage that surfaces in the underlying index — Bing for Copilot, Google for AI Mode, the open web for Perplexity — keeps feeding the candidate set and the brand entity simultaneously. Pricing and policy consistency is the quiet disqualifier: agents increasingly cross-check price, shipping speed and return policy across sources, and a product whose price differs across its own site, a marketplace and a feed reads as unreliable. Consistency is cheap to fix and expensive to ignore.

6–7. Structured data and crawlability — necessary, not sufficient

These are the table stakes. AI crawlers parse JSON-LD as standalone data, and JSON-LD holds the dominant share of structured-data formats precisely because it aligns with how agents process information. Crucially, AI systems do not wait for JavaScript to execute — if your JSON-LD or product data loads client-side, crawlers may never see it, so server-side rendering is non-negotiable. And blocking the wrong crawler is fatal: disallowing OAI-SearchBot fully excludes you from ChatGPT recommendations regardless of content quality.

The crawler behaviours are not uniform, and the differences are operationally useful. ChatGPT’s crawler places heavy emphasis on HTML content — by one analysis roughly 58% of its fetch requests — which tells you that clean, server-rendered text matters more than clever interactivity. Anthropic’s ClaudeBot devotes a large share of its requests to visual content, a reminder that product imagery and alt text are part of the signal, not decoration. GPTBot and ClaudeBot can download JavaScript files but cannot execute them, so anything rendered only on the client is invisible to them. Gemini is the exception: because it crawls on Google’s infrastructure, it can fully render JavaScript — but designing for the lowest common denominator (server-side rendering) is the only safe strategy when you are optimising for every agent at once. A further wrinkle: AI crawlers fetch proportionally more 404s than Googlebot and appear to operate on tighter time budgets, so a clean URL structure and shallow crawl depth for key product pages directly affect whether they get discovered.

Get these right once and move on; they will not differentiate you because your competitors will get them right too. The full schema specification for AI commerce — Product, Offer, AggregateRating, Brand — is in structured product data for AI commerce, and the crawler-access and server-side-rendering mechanics live in our guide to AI bot crawl optimisation. Treat both as prerequisites to everything else in this article — and confirm the underlying technical SEO foundation is sound, because an uncrawlable page cannot be recommended at all.

A worked example: how two identical products get different outcomes

Consider two mid-market UK coffee-equipment brands selling a near-identical £180 burr grinder. Both run Shopify, both enabled Agentic Storefronts, both have complete feeds and valid schema. On the eligibility layer they are indistinguishable. Yet ask ChatGPT, Perplexity or Gemini for “the best home coffee grinder under £200” and one is named in the answer while the other is not. The difference is entirely in the selection layer, and it is instructive to trace exactly where it comes from.

Brand A exists almost entirely on its own domain and Amazon. It has 90 reviews, the most recent from four months ago. No independent UK coffee blog has reviewed it, no “best grinders” round-up names it, and its brand has no presence in any knowledge graph. To an agent, Brand A is a plausible but uncorroborated option — eligible, but a risk to recommend, because nothing outside the brand’s own marketing vouches for it.

Brand B looks identical on paper but has spent two quarters on earned authority. It is named in three independent “best home grinder” guides that already rank in Google, two of which Perplexity cites directly. It generates 15–25 fresh reviews a month through a simple post-purchase system. Its founder has been quoted in a national lifestyle piece, and its brand entity is consistently represented across the web. When an agent assembles its candidate set, Brand B is not just eligible — it is corroborated by exactly the independent sources the agent trusts. It gets named, and because AI referrals convert around 31% higher than non-AI traffic, that recommendation is worth disproportionately more than its volume suggests.

The lesson:  The feed got both brands to the same starting line. Everything that decided the race happened off the product page — in citations, reviews, brand entity and editorial coverage. That is the work, and it is link building work.

Building the brand entity AI agents trust

Brand entity strength is the highest-leverage and least-understood signal, so it is worth treating in its own right. An agent’s willingness to recommend you is, in large part, a function of how confidently it can answer the question “is this a real, reputable brand that makes this kind of product?” That confidence is assembled from signals scattered across the web, not declared on your homepage. Three things build it.

Consistency: one name, one identity, everywhere

Entity confusion is the silent killer. If your brand appears as “Acme Coffee”, “Acme Coffee Co.” and “AcmeCoffee” across your site, feeds, marketplaces and press, an agent may treat these as weakly-linked or even separate entities, diluting every signal you earn. Standardise the brand name, the canonical description, the category language and the core facts (founding, location, what you make) across every surface you control, and push that same consistency into the surfaces you influence. Consistency is free and compounds: it makes every future mention reinforce a single, strong entity rather than scatter across several weak ones.

Corroboration: independent sources that confirm who you are

A brand that exists only on its own properties gives an agent nothing to cross-check. The fix is the same earned-media work that has always built authority — editorial coverage, expert commentary, data-led PR, directory and association listings, and reclamation of unlinked brand mentions — now aimed at building corroboration rather than just passing PageRank. Each authoritative, independent source that names your brand and describes it consistently raises the agent’s confidence on every query in your category. This is why a steady digital-PR programme is not a nice-to-have for AI commerce; it is the mechanism by which the entity gets built. The full tactic set is in the 15 strategies hub, and reactive digital PR that produces timely, citable coverage is one of the fastest ways to generate the corroboration agents look for.

Knowledge-graph presence: being a recognised node

Agents lean on knowledge graphs to resolve entities. When a user asks for “Nike running shoes,” the system must already understand what Nike is and what it makes; the same logic scales down to your brand. Clear, consistent brand information that knowledge graphs can ingest — and corroborating citations that justify your inclusion — turn your brand from an unknown string into a recognised node the agent can reason about. The deeper mechanics of entity and knowledge-graph work are a discipline of their own, which is why Phase 7 devotes an entire cluster to it; for AI commerce specifically, the priority is simply to ensure your brand is a resolved, well-described entity before you expect agents to recommend its products confidently.

How this differs by vertical and channel

The playbook is universal but the emphasis shifts by where you sell. Three patterns matter.

Channel / vertical	Dominant agent surface	Where to concentrate effort
Amazon-first sellers	Rufus, inside Amazon	Review velocity and Q&A within Amazon; pricing competitiveness. External citations matter less inside the walled garden
Owned-site DTC brands	ChatGPT, Perplexity, Google AI Mode	Brand entity, third-party citations and on-site reviews — the full earned-authority stack
Multi-channel retailers	All of the above	Maintain product data across ACP and UCP ecosystems; build authority once, and ensure pricing/policy consistency across every surface
Considered / high-ticket purchases	Perplexity, Copilot	Comparison-guide citations and linkable decision tools — buyers ask longer, research-heavy questions agents answer with sources
Impulse / low-ticket	ChatGPT, Gemini	Review depth and feed completeness; fast trust signals (free shipping, returns) carry disproportionate weight

The Amazon split is the one that trips brands up most often. Inside Amazon, Rufus reasons over Amazon’s own reviews, Q&A and pricing — your external citation programme barely touches it. Outside Amazon, your owned-site recommendations depend almost entirely on the web-wide authority Rufus ignores. Most brands need both engines running in parallel, with separate review-generation systems for each surface. The full treatment of that trade-off, including when a marketplace-first strategy beats an owned-site one, is in marketplace vs owned-site visibility in AI shopping.

One more channel nuance worth internalising: AI shopping queries are long and personal. Klaviyo’s 2026 data found 30% of AI searches run to eight words or more, and most carry emotional or personal context at least some of the time. That favours brands whose web presence — reviews, guides, editorial — speaks to specific use-cases and buyer situations rather than generic category terms. The more your earned coverage mirrors how real buyers describe their need, the more often an agent can match a long-tail query to your product.

The five-stage playbook for earning AI recommendations

Here is the operating model, sequenced. Stages 1 and 2 are foundational and largely one-time; stages 3, 4 and 5 are the ongoing earned-authority engine that actually wins recommendations. The mistake most teams make is spending 90% of their effort on stage 1 and wondering why recommendations do not follow.

Stage 1 — Eligibility (get into the candidate set)

1. Confirm AI crawler access. Audit robots.txt: allow OAI-SearchBot, PerplexityBot and the Google/Bing AI crawlers. Confirm server-side rendering of product data.
2. Complete the feed. Populate titles, detailed descriptions, specifications, materials, GTINs, price, availability, product_review_count, average_rating, free-shipping and return-policy fields.
3. On Shopify, enable Agentic Storefronts (active by default for all stores as of 2026) — this exposes your catalogue to ChatGPT, Copilot, Gemini and Google AI Mode in one step. Enrol separately in the Perplexity Merchant Program.
4. Deploy Product + Offer + AggregateRating + Brand schema as server-rendered JSON-LD.

Reality check:  Stage 1 is necessary and finite. A competent developer completes it in days. If a vendor is selling you stage 1 as an ongoing retainer, you are overpaying for the easy part.

Stage 2 — Brand entity foundation (become a recognised entity)

• Standardise brand naming everywhere — site, feeds, marketplaces, social, press. Inconsistency fractures the entity.
• Build the canonical brand information agents rely on: a clear, fact-dense brand/about presence and consistent representation in knowledge graphs.
• Earn a baseline of authoritative, independent mentions so the agent has corroborating sources for who you are.

This is the layer that most directly separates recommended brands from invisible ones, and it is built with the link building and digital-PR toolkit rather than with feed fields. It compounds: every authoritative mention strengthens the entity, which raises the agent’s confidence on every future query.

Stage 3 — Citation acquisition (get named in the guides)

This is the core earned-media engine. The objective is simple to state and hard to execute: become a product that independent buying guides, comparison articles and reviewer content name in your category. Tactically that means identifying the guides agents actually cite for your category queries, then earning inclusion through the established outreach disciplines — expert contribution, product seeding to reviewers, data-led PR that makes you the citable source, and reclamation of unlinked mentions.

• Run probe queries weekly across ChatGPT, Perplexity, Gemini and Google AI Mode in your buyers’ language; log which sources get cited and which competitors get named.
• Reverse-engineer the cited sources into an outreach target list — these are your highest-value placements.
• Prioritise the guides that already rank in classic search; they have a head start in AI citations.

The complete tactical breakdown — formats, pitch angles, reviewer relationships — is in earning citations in AI buying guides, and the broader tactic catalogue sits in the 15 strategies hub.

Two execution principles separate effective citation work from busy work. First, prioritise the sources agents actually cite over the sources with the highest domain rating. A mid-authority niche guide that Perplexity quotes for your exact category query is worth more than a high-DR placement the agents never surface — relevance to the query beats raw authority here. Your weekly probe queries are the map: they tell you precisely which publications feed the answers in your category, and those become your target list. Second, give reviewers a reason to include you that survives editorial scrutiny. Product seeding works, but it works far better paired with a genuine differentiator — a data point, an independent test result, a feature no competitor offers — that gives the writer something specific to say. Generic outreach asking to be “considered for your round-up” converts poorly; outreach that hands the writer a ready-made, citable reason converts.

Stage 4 — Review velocity (sustain the trust signal)

Stand up a repeatable system that generates recent, genuine reviews on the surfaces that matter for your channel mix — your own site for owned-site recommendations, the marketplace for Rufus. Velocity beats volume: a consistent monthly inflow signals current relevance in a way a large but stale review count does not. Pipe review count and average rating into your feed so the trust signal is legible to every agent.

Stage 5 — Linkable decision tools (earn citations passively)

The most durable citation asset is a genuinely useful buying-decision tool — a sizing calculator, a comparison engine, a fit finder — that reviewers and agents reference because it helps buyers decide. These earn links and citations continuously without per-placement outreach, and they position your brand as the category’s reference point. We cover how to design tools agents actually cite in building linkable buying-decision tools that AI agents cite.

A 30/60/90-day execution plan

Window	Focus	Concrete actions	Success signal
Days 1–30	Eligibility + baseline visibility audit	Fix crawler access and SSR; complete feed; enable Agentic Storefronts; enrol Perplexity; run baseline probe queries and log competitor citations	You appear in the candidate set; baseline share-of-voice recorded
Days 31–60	Entity + citation engine	Standardise brand data; launch outreach to the buying guides agents cite; ship one data-led PR asset; start review-velocity system	First new citations land; brand recognised as an entity
Days 61–90	Compounding + measurement	Scale citation outreach; ship a linkable decision tool; formalise weekly probe-query tracking and AI-attributed traffic segmentation	Measurable lift in recommendation frequency and AI-referred conversions

Notice the shape: eligibility is a 30-day project, but authority is a 90-day-and-beyond engine. The brands winning in 2026 front-loaded the easy work and then committed to the hard, compounding work that competitors keep deferring.

Measuring whether it is working

Both agentic protocols ship with primitive analytics, so you will not get clean, deterministic attribution for why an agent recommended a competitor over you. Move on velocity, not certainty. The practical measurement stack has three layers, and you need all three because a brand mention is not the same as a product recommendation, which is not the same as captured revenue.

• Visibility — run structured probe queries weekly and track how often you are named, in what position, and against which competitors. This is your AI share of voice.
• Traffic — segment AI-referred traffic from day one. AI-driven visits often land as direct or referral from chat domains in GA4; UTM-tag where you can and isolate the segment.
• Revenue — connect AI-referred sessions to conversions. Recall the quality signal: AI referrals convert markedly higher and produce fewer returns, so even modest AI traffic can outperform its volume.

The attribution gap to plan around:  Brand mentioned ≠ product recommended ≠ revenue captured. Most monitoring tools track only the first. Build visibility into all three, accept that the data will be imperfect, and let directional trends — not perfect causality — guide spend. For the broader benchmark context, our living dataset in the link building statistics hub is the reference.

For category-level benchmarks and the underlying data points cited throughout this guide, see our continuously updated link building statistics for 2026. And because crawlability quietly underwrites every measurement here — an uncrawlable page cannot be recommended, measured or attributed — keep the technical SEO foundation in good order as a precondition, not an afterthought.

Running probe queries that actually tell you something

The visibility layer is only as good as the queries you test, so build the probe set deliberately. Start from how your buyers describe their need, not from your product names: pull the actual language from reviews, support tickets and search data, and turn it into 15–25 representative queries spanning broad discovery (“best X for Y”), specific comparison (“X vs competitor”), and constrained requests (“X under £200 with good warranty”). Run them across ChatGPT, Perplexity, Gemini and Google AI Mode on a fixed weekly cadence, and log three things each time: whether you are named, in what position relative to competitors, and — on Perplexity especially — which sources the answer cites. The cited-source column is gold, because it converts directly into your citation-acquisition target list.

Over a few weeks this produces a share-of-voice trend per engine, which is the closest thing to a deterministic scoreboard the current tooling allows. Watch the direction, not the absolute number: a rising trend across engines as your citation and review work compounds is the signal that the playbook is working, even before the traffic and revenue layers catch up. Because the engines update their source selection frequently, treat the probe set as living — refresh queries quarterly and add new competitor names as they start appearing in your category’s answers.

Eight mistakes that keep strong brands invisible

• Treating it as a pure feed problem. Perfect feeds get you eligible; they do not get you picked. Authority is the tie-breaker.
• Blocking the wrong crawler. A single line in robots.txt disallowing OAI-SearchBot removes you from ChatGPT recommendations entirely.
• Client-side-only product data. If schema loads via JavaScript, AI crawlers may never read it. Render server-side.
• Optimising for one agent. Each engine weights inputs differently; the goal is web-wide authority every agent can find, not a single-surface hack.
• Ignoring review velocity. A large but stale review base loses to a smaller, fresher one. AI surfaces reward recency.
• Inconsistent pricing and policy data. Cross-source contradictions read as unreliability and quietly suppress recommendations.
• Chasing checkout integrations over discovery. Discovery is the unsolved, high-leverage problem; bleeding-edge checkout may be deprecated within quarters.
• Never being cited anywhere independent. A brand that exists only on its own domain gives the agent no corroboration and reads as a risk.

Where this leaves you

AI shopping agents have collapsed product discovery into a named-recommendation surface, and the mechanics of that surface reward a specific kind of work. The eligibility layer — feeds, schema, crawler access — is real, finite, and identical to what every competitor can do. The selection layer — brand entity strength, third-party citations, review velocity, web-wide authority — is open, compounding, and overwhelmingly an earned-media discipline. That second layer is where recommendations are actually won, and it is where link builders have a structural advantage that feed-management vendors do not.

The brands that will own AI recommendations through 2027 are not the ones with the most perfectly formatted feed. They are the ones the rest of the web already vouches for — the brands that buying guides name, that reviewers reference, that knowledge graphs recognise, and that fresh reviews keep validating. Build that authority deliberately, measure it honestly, and let the feeds do the easy job of carrying it into every agent’s answer.

Start with eligibility this week, stand up the citation engine this month, and treat the whole thing as the authority programme it actually is. The companion guides linked throughout — on feeds, ranking factors, buying-guide citations, schema, marketplaces and decision tools — turn each stage of this playbook into an executable project.