When Enterprise Agents Pick Suppliers: Inside the Machine-Led Vendor Search

Enterprise procurement has always been a research problem wearing a relationship costume. Someone had to read the vendor sites, collect the specifications, reconcile the pricing pages and build the comparison grid the committee argues over. That work is now being delegated — and the delegate does not browse, does not remember your brand campaign and does not call a rep.

Gartner’s projection puts the scale plainly: by 2028 AI agents will outnumber human sellers roughly ten to one, and by 2029 around 80% of customer service interactions will be handled by AI (via Andy Hoare, 2026). Satya Nadella’s version at Build 2026 is architectural: the web is increasingly read by agents rather than people. Brex’s Pedro Franceschi describes the buy side already building for it — an agent for every persona in the purchasing process, against $1.8 trillion in US corporate card spend growing 12% a year.

This is the sixth report in our research series, following engine divergence, the citation supply chain, brand misinformation, the B2B shortlist and the long-tail thesis.

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 — which suppliers Turkish-language procurement prompts actually surface, sector by sector — is planned as a follow-up on Brantial’s prompt panel.

The frame

We follow the mechanism rather than the hype: what an agent can retrieve, what it can verify, where it stops, and which supplier-side properties decide whether it can include you at all.

Finding 1: The committee did not disappear — its research layer was automated

The published data resists the simple story. Enterprise purchases still involve 11 to 20 stakeholders and roughly 27 touchpoints before sales contact (Searchable customer data), and 94% of B2B buyers consult an LLM before speaking to a representative (Paolo Provincali, LinkedIn — Zero Click NY 2026). Both are true at once: the human committee still decides, but the file it decides from is machine-assembled.

That is the practical definition of machine-led vendor selection in 2026 — not autonomous purchasing, but autonomous shortlisting. John Collison and Criteo’s Michael Komasinski both push back on full autonomy for exactly this reason: the agent researches and executes, the human decides.

What this means for a supplier: the artefact that decides your fate is a comparison you never see, produced by a reader you cannot email. Influencing it means being legible to retrieval, not persuasive to a person.

Finding 2: Agents retrieve differently from buyers — and the difference is measurable

An agent building a vendor comparison inherits the retrieval behaviour of its engine, which our divergence report documented: only 8% domain-level citation overlap between ChatGPT and Claude, Claude drawing 79.2% of citations from Brave’s top-10, ChatGPT resolving around 95% through its own index (Josh Blyskal, Profound — Zero Click NY 2026).

Two consequences follow for procurement specifically. First, a supplier’s presence in an agent-built comparison is engine-dependent — a category leader on one stack can be missing from another’s grid entirely. Second, because fewer than 3% of citations go to tier-1 media and the rest to niche and community sources (Tim Sanders/G2), the documents that decide enterprise comparisons are review platforms, specialist sites and forum threads, not the coverage a corporate communications budget produces.

There is a controlled data point on the supplier side: adding concise context summaries lifted citation rates by 44% (G2). Machine legibility is not a metaphor; it is an editable property of your pages.

Finding 3: Agent trust is replacing domain authority

The most instructive published experiment came from Ramp’s George Bonaci at Zero Click NY 2026: an incentive offer published for AI agents only. The result was 1,300+ bot visits across seven platforms in three weeks, with Claude traffic up 180% and DeepSeek up 845%, and around 370 “agent relays” — instances of an agent passing the offer to its user. Model behaviour diverged wildly: Claude relayed it, ChatGPT never mentioned it.

Bonaci’s framing is the finding: agent trust is becoming the successor to domain authority. Whether an agent surfaces your offer depends on how the model weighs your source, not on how many backlinks the domain carries.

Interpretation, labelled as such: this is early and experimental — one company, one offer, one three-week window. It is not a playbook. It is evidence that agent-directed publishing produces measurable, engine-specific responses, which is the premise the rest of this report rests on.

Finding 4: What makes a supplier machine-selectable

Reading the published findings together, four supplier-side properties decide inclusion — and all four are auditable rather than negotiable:

Retrievability. If AI crawlers cannot reach the pages holding your specifications and pricing logic, no agent can include you. Blocked bots and missing llms.txt directives are exclusion, not caution.

Structured facts. Agents extract claims they can attribute: named specifications, plain pricing statements, clear category language, schema markup. Prose that requires interpretation gets skipped in favour of a competitor’s table.

Third-party corroboration. Since answers are built mostly from niche and community sources, a supplier corroborated across review platforms and specialist coverage is easier to include than one whose only source is its own site.

Currency. Dated, maintained pages beat stale ones — a point our misinformation report covers in detail, where outdated sources become confidently repeated wrong facts.

What this means in Türkiye

Machine-led selection arrives before local optimisation does. Turkish enterprises are adopting the same assistants on the same timelines, but Turkish supplier pages have rarely been prepared for agent retrieval. The gap between demand and readiness is the opportunity — and it closes as competitors notice.

Export procurement is judged in English sources. A Turkish supplier selling abroad is assembled into comparisons from English-language review platforms and specialist sites. Local reputation does not travel into that corpus; corroboration must be built where the agent reads.

Procurement language is a retrieval problem. Turkish enterprise buyers prompt in Turkish, in domain vocabulary that often differs from a vendor’s marketing language. Publishing specifications in the words buyers actually use is the cheapest step toward being selectable — and it is measurable per prompt.

The takeaway

Machine-led vendor selection is not a future scenario waiting for autonomous purchasing; it is the current state of the research phase, with the human committee still holding the decision. The supplier-side work is unglamorous and specific: let the crawlers in, publish facts in extractable form, earn corroboration in the sources agents actually read, and keep the dates current.

The sequence: audit retrievability first, structure the facts second, build third-party corroboration third — then measure whether agent-built comparisons in your category name you at all.

Sources

• Andy Hoare — Gartner projections on agent volume and service interactions, 2026
• Satya Nadella, Microsoft — agent-read web, Build 2026
• Pedro Franceschi, Brex — persona-level purchasing agents and corporate card spend, 2026
• Paolo Provincali, LinkedIn — pre-sales LLM consultation, Zero Click NY 2026
• Searchable customer data — stakeholder counts and pre-contact touchpoints, 2026
• Josh Blyskal, Profound — retrieval architecture and citation overlap, Zero Click NY 2026
• Tim Sanders, G2 — citation ecosystems and the context-summary experiment, Zero Click NY 2026
• George Bonaci, Ramp — agent-directed incentive experiment and agent trust, Zero Click NY 2026
• Brantial research series (2026): Divergence, Source Ecosystem, Brand Misinformation, First Two Slots, Long-Tail Discovery

Figures are as presented in the cited talks, interviews and publications and primarily describe English-language, US-market behaviour. Brantial’s follow-up study will measure supplier presence in Turkish procurement prompts across the engines in our panel.