AI agents in B2B buying find, read, extract, compare and verify public facts, then exclude providers with missing or conflicting answers. To be shortlisted, publish certifications, regions, service levels and pricing models as open, explicit, dated facts that read identically across every source.
Your next shortlist may be assembled by software that ignores adjectives, keeps only what it can verify and never tells you it left you out.
Booking a hotel once meant phoning an agent who knew the owners. Now an app compares price, location and cancellation terms in seconds, and the brochure stays unopened. AI agents in B2B buying are beginning to do the same to software shortlists. This guide explains how an agent researches your product and how to audit your evidence so it is found, read and trusted.
How are AI agents changing B2B buying?
Gartner, as reported by Digital Commerce 360, projects that AI agents will intermediate more than USD 15 trillion of B2B purchases by 2028. Your buyer may still sign the contract. The first shortlist, however, may be built by software that reads public pages and drops anything it cannot confirm.
Agent research follows a recognisable sequence: find, read, extract, compare, verify and exclude. An agent has no use for "enterprise-grade". It looks for facts it can line up against competitors, such as certifications, regions, service levels and pricing model.
Most failures share one cause. The evidence exists, but sits behind a demo form, inside a gated PDF or in a trust portal that needs a login. Nobody tells you an agent left you off a shortlist; the opportunity simply never reaches your pipeline.
How to run an agent-readiness audit
Begin with the facts buyers compare. Audit your homepage, pricing page, trust page, documentation and the third-party pages that rank for your name. In practice, a first pass takes a working day and finds the largest problems.
- Certifications: the standard, type, scope and period, such as a SOC 2 Type II report on a named product.
- Data location: hosting and processing regions by name, including backups and logs.
- Availability: the committed percentage, how it is measured and a link to a status page.
- Pricing model: the unit you charge on, even where exact prices need a quote.
- AI data use: whether customer inputs train any model, and how long prompts are retained.
A SOC 2 Type II report is an independent auditor's opinion on how controls operated over a period. The full report is usually shared under NDA, which is fine. Its existence, type, scope and period can still be stated publicly.
Next, check access. Your robots.txt file, defined by the IETF in RFC 9309, tells crawlers which paths they may fetch. It is advisory and is not access control, so keep confidential material behind authentication and public trust facts open.
Then check structure. Put key facts in HTML text and tables, not images. Structured data using the schema.org vocabulary, such as the SoftwareApplication and Offer types, can help. An llms.txt file, a community proposal from September 2024, is a low-cost addition rather than a fix.
Finally, score each fact on four checks, one point per pass. A fact scoring 4 is agent-ready; anything below 3 needs work this quarter.
| Check | Pass looks like | Fail looks like |
|---|---|---|
| Reachable | Open page within two clicks of the homepage | Behind a form, a login or a blocked path |
| Explicit | A number, scope or named term | An adjective |
| Dated | Issue, validity or last-reviewed date shown | No date anywhere |
| Consistent | Same value on website, documentation, listings and contracts | Different values in different places |
Precision matters for numbers. A 99.9% availability commitment allows about 43 minutes of downtime in a 30-day month, but only if you also state how it is measured.
A worked example: three providers, one shortlist
In a hypothetical case, a procurement analyst asks a research agent to shortlist document-processing tools. The criteria are EU data residency, single sign-on, a SOC 2 Type II report and at least 99.9% availability.
| Requirement | Provider A | Provider B | Provider C |
|---|---|---|---|
| EU data residency | Regions named on pricing page | Only in a gated PDF | "Global infrastructure" |
| SOC 2 Type II | Type, scope and period stated | "SOC 2 compliant", no type or date | Type and period stated |
| Availability | 99.9%, measured monthly, status page linked | 99.95% on website, 99.5% in contract | Not published |
| Agent outcome | Shortlisted | Excluded: unverifiable and conflicting | Excluded: residency and availability unknown |
Provider B may well have the best product. It lost because its evidence was gated, vague and inconsistent. Winning required no larger budget, only facts published where a machine could check them.
Common mistakes and why this matters now
- Blocking everything: a blanket block on AI crawlers hides public trust facts, so decide path by path.
- Gating basic facts: a fact that appears in every proposal probably belongs on a public page.
- Leaving old pages live: retired plan pages keep feeding wrong facts into answers.
- Letting sources drift: two pages naming different regions read as risk, so keep one owned source per fact.
Agents reward checkable evidence, and the advantage compounds. Independent, audited evidence suits this shift because it is verifiable by design; the Trust-IQ Score™ from Smartreview AI is one example. Test the result monthly by asking general AI assistants a buyer's questions and tracing each wrong answer to its source.
The bottom line
Conclusion
The hotel app never reads the brochure; it compares what it can confirm. Software shortlists are heading the same way. An agent will reach what is open, trust what is dated and drop what conflicts, without telling you.
Publishing for that reader is disciplined housekeeping rather than a new campaign. This week, choose the five facts buyers in your category compare most, score each on the four checks, and fix every fact that scores below three.
Frequently asked questions
How do AI agents decide which software to shortlist?
They compare extractable facts against the buyer's criteria, such as certifications, data residency and service levels. Providers whose facts are missing, gated, undated or contradictory tend to be dropped.
Should I block AI crawlers in robots.txt?
Decide path by path rather than blocking everything. Robots.txt is advisory, so keep confidential material behind authentication and keep public trust facts crawlable.
What is llms.txt and do I need one?
It is a community proposal from September 2024 for a file that points AI agents to a site's most useful pages. It is not an adopted standard and cannot fix vague content.
Does structured data help AI agents read my site?
Structured data using the schema.org vocabulary, such as the SoftwareApplication and Offer types, can make product and pricing facts easier to extract.
Will AI agents replace human software buyers?
People still set criteria and sign contracts. Agents, which Gartner projects will intermediate more than USD 15 trillion of B2B purchases by 2028, increasingly shape which options people see.
References
- Digital Commerce 360, Gartner: AI agents and $15 trillion in B2B purchases by 2028
- IETF, RFC 9309: Robots Exclusion Protocol (September 2022)
- Schema.org, SoftwareApplication type
- llms.txt, A proposal to standardise on using an /llms.txt file (proposed September 2024)
- Secureframe, SOC 2 Trust Services Criteria explained
- Smartreview AI, Trust-IQ Score™ methodology
Disclaimer: General information only, not legal, security, compliance or financial advice. Check the cited sources and seek professional advice. Names belong to their owners; no endorsement implied.