The AI Examiner — verification-intent search, 2026-08-22

Headline

TrustRadius claims 94% of B2B buyers fact-check AI-generated claims before trusting them. Across a 1,735-keyword search corpus, 21 of 27 verification-intent term families returned zero searches.

Summary

A January 2026 survey of 1,862 technology buyers found most say they fact-check AI-generated output before acting on it. This issue tests whether that stated behavior shows up as search demand. Across fifteen Google Keyword Planner exports, deduplicated into 1,735 unique keywords, twenty-seven verification-and-skepticism term families — scam, hype, debunk, verify, benchmark, hallucination, vendor claim, evidence, audit, compare, lawsuit, and others — were tested as substrings. Twenty-one returned nothing. Of the six that returned something, the two largest are false positives: a single off-topic phrase, and a 158-keyword spike in "review" that is almost entirely generic software reviews and document-review-software terms, not searches evaluating an AI vendor's claim. Two low-volume terms — "agentic ai roi" and "ai failures in healthcare" — are the closest thing in the corpus to a buyer checking a claim.

The evidence

What buyers say they do

TrustRadius, now part of HG Insights, surveyed 1,862 technology buyers and 444 technology vendors in January 2026 for its "B2B Buying Disconnect" report. As reported by MarketScale on 2026-08-06, "94% of B2B buyers... fact-check what the AI told them" before trusting it (fetched, self-reported — MarketScale's own coverage; no direct link to a primary TrustRadius publication was located this run). This is a claim about buyers checking AI-generated research output during a purchase, not specifically about checking an AI vendor's own performance marketing — a distinction this issue does not blur.

What the search corpus shows

research/data/ holds five Keyword Planner exports pulled 2026-08-08; research/demand/keyword-planner/ holds ten pulled 2026-07-30, covering different seed terms (AI-agency and virtual-production clusters, US and Canada). Combined and deduplicated by keyword, the fifteen files hold 1,797 rows collapsing to 1,735 unique keywords. 1,688 carry a numeric average-monthly-searches figure; 47 are blank. Every distinct value present is one of six: 0, 50, 500, 5,000, 50,000, or 500,000 — Google's own buckets, not exact counts. The floor is 0, held by 28 keywords (1.61% of the corpus) — not 50, the floor this project's own prior notes had assumed without recomputing it. The ten highest-volume keywords are all generic AI-category or AI-plus-healthcare terms — agentic ai (500,000/mo) at the top, then ai agency, artificial intelligence agency, ai in healthcare, and medical ai, each at 50,000/mo. None involve checking a vendor's claim.

The verification-intent test

Twenty-seven term families were tested as case-insensitive substrings against all 1,735 keywords. Zero matches: worth it, scam, legit, hype, overhyped, debunk, misleading, snake oil, benchmark, accuracy, hallucination, vendor claim, due diligence, evidence, verify, verified, compare, alternative, complaint, lawsuit, failed — 21 of 27 families. One match each: independent (independent filmmaker ai tools, about filmmaking, not verification) and failure (ai failures in healthcare, 50/mo). Roi returned one match, agentic ai roi, 50/mo. Proof returned two matches, both substring hits inside unrelated product names rather than the word used as search intent. Audit returned six: one generic compliance phrase, five small-business "AI audit" searches — a buyer wanting their own AI usage assessed, not a vendor's claim checked, an intent this issue does not conflate with the rest.

Review is the outlier: 158 matches. Inspected individually, none represent a buyer checking whether an AI vendor's claim holds up. They split between generic consumer-software searches — pdf reader reviews, pdf editor reviews, onlyoffice review — and document-and-contract-review-software terms concentrated in the legal sector, which this publication's standing coverage rules exclude from citation by name here. Both groups are homonyms of the word this test searched for, not the search intent it was testing.

The two real hits

agentic ai roi and ai failures in healthcare are the only unambiguous verification-adjacent terms found, and both sit at the lowest populated volume bucket, 50 searches a month — against agentic ai itself at 500,000. Two terms at 50 each are not a market signal; they are close to the smallest thing Keyword Planner will report at all.

What this finding does not establish

It does not establish that buyers never verify a vendor's claims — TrustRadius's own survey says most report doing exactly that, through channels a keyword tool cannot observe: asking a vendor directly, a peer conversation, a trial. It does not establish that verification-intent search is zero everywhere, only that it is effectively zero in this specific corpus, built from this project's own prior seed terms rather than a neutral sample of the category. It does not explain the gap between TrustRadius's 94% and this corpus's near-total absence of verification terms — this issue measures the gap, not its cause. And it does not establish the 158 "review" matches are meaningless as a category, only that none of them, on inspection, are a buyer checking an AI vendor's claim.

What this means for a procurement lead evaluating an AI vendor

A vendor's performance claim is not something the market is searching to check, at any scale a keyword tool can detect. If verification is happening, per TrustRadius's own figure, it is happening through channels this kind of data cannot see — not through open search for a skeptical term. A procurement lead who assumes a public, searchable record of diligence on a given vendor's claim already exists should not expect to find one; the search record shows almost nobody looking.

Methodology & sample

This issue rests on operator-supplied data — fifteen raw Keyword Planner CSV exports already in this repository, converted from UTF-16LE to UTF-8 with a dynamically located header row — plus four fetched WebSearch queries for the named claim above. No figure was estimated, rounded, or carried over from prior write-ups; every count was recomputed from the export files in this run. A reader can reproduce the corpus figures by deduplicating the same fifteen files by keyword and testing the same twenty-seven substrings.

Gaps

• No independent measurement of "verification intent" search volume outside this project's own two Keyword Planner snapshots was located this run — there is no second source to check the zero-match pattern against.
• The TrustRadius figure came through one secondary source (MarketScale); the primary report was not directly fetched, so its methodology beyond sample size and survey month is unconfirmed here.
• "Verification intent" was operationalized as 27 given term families; a different word list could surface different results, and this list was not independently derived or tested for completeness.
• The corpus reflects this project's own prior seed terms (AI consulting, automation agencies, virtual production, healthcare, finance, insurance, customer service), not a neutral cross-section of AI-related search.
• The legal/eDiscovery share of the "review" matches was identified by inspection, not an exact programmatic count, consistent with this publication's standing exclusion of legal-industry AI vendors and eDiscovery from citation.