Industrial supplier AI search: methodology and source limitations
Study and source-inspection date: 16 October 2026

Question
Can an evidence-seeking Claude API workflow identify UK compressed-air supplier candidates while distinguishing published capabilities from unresolved buying requirements?

Sample and units
Five historical procurement processes were selected purposively during development, rather than through a representative market sample: CPI (2020), Capestone Organic Poultry (2021), Glass Futures (2022), Royal Mint (2023) and Penzance Dry Dock (2024). The initial pilot contained three cases. Two cases were added separately. Each case had four normalised requirements and three fresh requests. Fifteen completed requests produced 30 candidate appearances and 120 candidate-by-requirement rows. Repeated rows and candidate appearances are not independent suppliers or procurement processes.

Inputs and execution
Researcher-normalised briefs retained contract type, explicit requirements and missing information. Buyer names, notice URLs and historical awardees were omitted from the recommendation prompt. Distinctive wording prevents a claim of complete blinding. The prompt requested candidates worth contacting, primary evidence and uncertainty.

The API model was claude-sonnet-5-5, with between-tools thinking, no temperature override and 4,500 output tokens. Each request allowed up to two candidates, three web searches and three web fetches, with a 4,000-token fetch-content allowance. Requests had no shared conversation. The pilot used seeded order 20261016 and up to three concurrent requests. Inputs and execution records were retained. These settings describe this workflow, not ordinary consumer-app use.

Inspection and classification
Original model labels are preserved separately from subsequent machine-assisted source classifications. The first inspection covered 19 positive pilot rows, nine positive expansion rows and one contradicted expansion row. It found seven full, 18 partial, three generic and one unobserved requirement. All remaining 91 originally uncertain rows were then inspected: the pilot's 53 yielded zero full, 13 partial, 24 generic and 16 unobserved. The expansion's 38 yielded ten full, 12 partial, five generic and 11 unobserved. A separate later-source overlay changed one prior expansion generic classification to full.

Current classifications total 18 full, 43 partial, 31 generic and 28 unobserved. Full means explicit published coverage of the normalised capability. Partial means some requested components are explicit. Generic means related capability without the requested specifics. Unobserved means the inspected material doesn't establish the requirement. No classification certifies supplier fitness or project acceptance. Missing evidence doesn't establish incapability. Nonexclusive pipework alternatives didn't establish the original model's contradiction.

Source boundaries
Inspection distinguishes original run-bound tool bodies, later separate captures and fresh follow-up pages. Exact passages, source routes, dates and hashes were retained. Later pages can corroborate capability but cannot establish original model visibility. Truncated extracts, navigation-only content, unavailable pages and blocked collection routes constrain observation. The ten newly full uncertain rows concern simpler Royal Mint offerings, not fully matched compound pilot requirements.

Historical notices are compared retrospectively with current supplier content. Royal Mint's detailed portal specification wasn't retrieved. Penzance uses its original attachment only because a later clarification workbook was unavailable. The cases differ in detail and requirement grouping. Pooled accuracy, precision, recall and false-negative rates are not reported.

Disclosure and limits
This is machine-assisted textual inspection without human engineering validation, buyer contact or procurement-outcome testing. It tests neither ChatGPT nor consumer Claude, rankings, enquiry gains or formatting causality. SEO Yolo sells SEO services. Control Gear is an existing direct agency client and remains included. No private client analytics were used.