General Assistant or Retailer Assistant: Two Markets Inside One Answer Layer

Two things are true about the same product at the same moment. In ChatGPT, it is recommended because the retrieval layer found credible sources supporting it. In the retailer’s own assistant — the one embedded in the shop the buyer already trusts — it may not appear at all, because that assistant answers from a catalogue, a margin structure and a merchandising logic that no amount of citation quality reaches.

Criteo’s Michael Komasinski expects exactly this split to persist: specialised retail assistants outperforming general ones inside their categories, with both coexisting. If that holds, “AI visibility” is not one market with one scoreboard. It is two, and a brand can lead one while being absent from the other.

This is the ninth report in our research series, following engine divergence, the citation supply chain, brand misinformation, the B2B shortlist, the long-tail thesis, enterprise agents, challenger versus incumbent and the shopping source map.

Methodology

A synthesis, openly sourced

This report is a research synthesis, not a Brantial panel measurement. Every figure comes from published third-party research, executive interviews and 2026 conference presentations, cited inline; interpretation is labelled as interpretation. The Turkish measurement — how brand presence differs between general assistants and Turkish marketplace assistants, category by category — is planned as a follow-up on Brantial’s prompt panel.

The frame

Three questions decide whether a brand should treat these as one market or two: do the two assistant types select differently, do buyers use them differently, and can a brand influence both with the same work?

Finding 1: The two types select on different inputs

A general assistant assembles its answer from retrieval: web sources, community threads, comparison content and structured pages, weighted by whatever its engine trusts. Our source map report documented that mix in detail — around 16% community citations on ChatGPT against 0.9% on Claude, roughly 36% of Claude citations from listicles, and fewer than 3% of citations to tier-1 media (Profound; G2 — Zero Click NY 2026).

A retailer assistant starts from a different premise: it answers from the catalogue it sells, with the availability, pricing and promotion logic of that retailer. Alex Chriss’s framing of the merchant problem at PayPal makes the tension explicit — exposing a catalogue to agents while not losing the customer relationship is the open question of this phase, and over 80% of e-commerce sits in catalogues that could be opened to agents.

Interpretation: the selection inputs barely overlap. Citation quality moves the general assistant; catalogue presence, feed quality and commercial terms move the retailer one. A brand optimising only for the first can be structurally absent from the second — not outranked, simply not stocked in the answer.

Finding 2: Buyers use them at different moments — and trust them differently

The trust asymmetry is the most striking published number in this area: 55% of consumers are hesitant to give payment details to a general AI assistant, while retailer-domain assistants are trusted roughly three times more (Criteo consumer research, 2026).

Read alongside the referral behaviour — about 70% of AI referrals to retailers land directly on a product detail page, up from 50% a year earlier, converting at roughly 1.5x traditional search (Criteo/OpenAI data) — a plausible division of labour emerges: the general assistant does discovery and comparison, then hands the buyer to a trusted commercial surface where the transaction closes.

Interpretation, labelled as such: this is a reading of two datasets, not a measured funnel. What the data supports is that discovery and transaction are trusted to different surfaces; how often the handoff happens is not established in the published work.

Finding 3: The influence levers are not interchangeable

For the general assistant, the work is the one this series has documented repeatedly: retrievability, structured facts, corroboration in the sources the engine reads, currency. It is editorial and technical, and a brand can do all of it unilaterally.

For the retailer assistant, the levers are commercial and operational: product feed completeness, attribute quality, imagery, stock reliability, review volume on that platform, and the trading relationship itself. A brand cannot publish its way into a marketplace assistant’s answer; it has to be a well-described, well-performing item in that marketplace’s catalogue.

The infrastructure work of 2026 — Shopify and Google standardising catalogue exposure through UCP, PayPal’s remote MCP server opening catalogues to agents, Mastercard’s agent payment framework — is mostly about the second lever. It makes catalogues machine-consumable; it does not make a brand’s own site more citable.

What this means for a brand: two workstreams, two owners. Content and SEO/GEO teams move general-assistant presence; commercial and marketplace teams move retailer-assistant presence. Measuring them with one number hides which one is failing.

Finding 4: The strategic risk is being recommended and not credited

The two markets interact in one uncomfortable way. When a general assistant recommends a product, the answer may name the brand, the marketplace listing, or both — and where marketplace listings are better structured than brand pages, the listing becomes the cited source, as our source map report sets out.

The brand still gets the sale, and loses the relationship, the data and the attribution. Komasinski’s gatekeeper argument points the same way: as organic discovery erodes, platforms that own the discovery layer capture more of the value in it.

Practical implication: “is my brand recommended?” is the wrong single question. The measurable pair is: does the answer name my brand, and does it link to a surface I control?

What this means in Türkiye

The vertical half is concentrated. Turkish e-commerce concentrates in a small number of large marketplaces, so the retailer-assistant market here is effectively a handful of platforms rather than a long list. That makes it unusually tractable to audit — and unusually consequential, because presence in those catalogues is close to a binary.

The general half is unusually open. As our long-tail report documented, Turkish prompts face fewer optimised competitors, so the editorial and structural work buys more presence per hour than the same effort in English. The two halves therefore reward different sized brands differently: challengers gain fastest in general assistants, incumbents defend more easily inside marketplace catalogues.

The attribution question is sharper here. Where marketplace listings dominate product data in Turkish, the risk of being recommended-but-not-credited is higher than in markets with strong direct-to-consumer web presence. Measuring whether Turkish answers name the brand or the listing is the specific thing worth knowing before allocating budget between the two workstreams.

The takeaway

The answer layer is splitting into a discovery market decided by retrieval quality and a transaction market decided by catalogue and commercial terms. They select on different inputs, are trusted differently by buyers, and respond to different work — which makes a single “AI visibility score” a blend of two things that should be managed separately.

The sequence: measure presence in general assistants and in the retailer assistants that matter in your category as two distinct numbers; fix retrievability and structure for the first; fix feed quality and catalogue presence for the second; and track, in both, whether the answer names your brand or the platform carrying it.

Sources

• Michael Komasinski, Criteo — vertical versus general assistants, referral conversion, product-page landings and paid-discovery economics, 2026
• Criteo consumer research — payment-trust distribution across assistant types, 2026
• Alex Chriss, PayPal — catalogue exposure to agents and merchant relationship risk, 2026
• Josh Blyskal, Profound — citation distribution by source type and format, Zero Click NY 2026
• Tim Sanders, G2 — citation ecosystems and tier-1 media share, Zero Click NY 2026
• Morgan Stanley — agentic contribution projection for US e-commerce (via Komasinski), 2026
• Brantial research series (2026): Source Ecosystem, Long-Tail Discovery, Shopping Source Map

Figures are as presented in the cited interviews, talks and publications and primarily describe US-market behaviour. Brantial’s follow-up study will measure brand presence across general assistants and Turkish marketplace assistants on our own prompt panel.