1. Direct answer: the 68% does not hold
It fails on three counts, any one of which disqualifies it as evidence:
- No primary source. Follow the chain of citations upward and it always ends at “a survey shows”, with no openable report, sample size or fielding date.
- Inverted meaning. In TrustRadius's annual buyer research, “AI had no impact on our buying process” has long been the majority answer — that is where the 68% comes from. Read as “68% are influenced”, it points the opposite way.
- Mixed denominators. 68%, 78% and 94% are routinely listed side by side although they come from different firms, years, questions and even different reports — suggesting a trend line that does not exist.
2. Four figures you can open and check
Every number below maps to a paragraph in a primary document. Note the last column: what a figure cannot support matters more than the figure itself.
| Figure | Primary source | Sample & timing | What was actually asked | Supports | Does not support |
|---|---|---|---|---|---|
| 45% | Gartner press release, 2026-05-20 | n=645, fielded Aug–Sep 2025 | Did you use generative AI to gather vendor and product information in your most recent purchase? | Nearly half of B2B buyers already research with AI | “45% of orders are decided by AI” |
| 38% (30% occasional + 8% frequent) |
TrustRadius, Bridging the Trust Gap, 2025 | n=2,058 buyers / 490 vendors, fielded 2025-01 | How often do you use AI in your buying process? | AI use is still a minority behaviour, but no longer marginal | Extrapolating to every category and region |
| 72% / 90% | Same study | Same | Did you encounter AI Overviews during research; did you click a cited source? | AI summaries are a common entry point, and cited links do get clicked | “90% of people will click your link” |
| 79% | G2, 2025 Buyer Behavior Report | n=1,169, fielded Apr 2025 | Has AI search changed how you research? (software & services buyers only) | Research behaviour in software buying has shifted | Extrapolating to cross-border physical goods |
| −25% | Gartner press release, 2024-02-19 | Analyst forecast | Change in traditional search engine query volume by 2026 | Search entry points are shrinking; you need another visibility channel | “Traffic halved”, or anything about this site's own organic traffic |
3. Three failure modes, and how to spot them
3.1 Inversion: from “no impact” to “directly influenced”
TrustRadius's annual buyer study has asked the same question for several years: how much of a role does AI play in your buying process? The distribution has been stable — “no impact” is the majority answer. The retelling flips the direction of that answer entirely. The tell-tale sign is a very specific number with no question text.
Spot it: ask for the question wording. Any serious study can produce it; if it cannot, the number is unusable however precise it looks.
3.2 Relabelling: 94% moves from satisfaction to adoption
The 94% does exist, in a different TrustRadius AI report published in 2025 — but its denominator is buyers who already use AI, and the question is whether they find it helpful. In the same report, 45% of software buyers use AI in purchasing; measured by frequency in the other study, “at least occasionally” is 38%. Quoting 94% as an adoption rate more than doubles the apparent scale.
Spot it: ask for the denominator. Before accepting any percentage, ask “a percentage of whom?”
3.3 Second-hand drift: how “traffic halves by 2028” was born
Gartner's official release of 2024-02-19 contains exactly one forecast: a 25% drop in traditional search engine query volume by 2026. “Down 50% by 2028” does not appear there; it circulates through secondary articles citing a paywalled report, with no readable original. Our site quoted that second-hand version and replaced it with the official wording on 2026-11-09.
Spot it: look at where the link points. A press release or report PDF is usable; another article means someone else's retelling is being passed off as evidence.
4. The five-step check
5. Our own rules for using data
After this audit we turned these into hard constraints enforced by a pre-publish check:
- Only cite primary sources you can open. If it cannot be opened, it does not go into a stat card, a headline or JSON-LD.
- Every external figure carries sample size and fielding date. The three homepage figures each have an evidence card with scope and limitations.
- Our own monitoring data needs a published definition first. Until then we describe direction, not percentages.
- Correct errors and keep the changelog. Section 3.3 above and the homepage evidence card are products of this rule.
- No invented report titles, no correlation written as causation, no single-year finding written as a trend.
6. Common misuses
- “AI-assisted research” written as “AI decides purchases” — four gates sit in between: validation, price comparison, negotiation, sample approval.
- “User satisfaction” quoted as “adoption rate” — swap the denominator and the conclusion flips.
- Software-category findings applied to physical exports — category variance exceeds regional variance.
- An analyst forecast (−25%) presented as a measurement, then layered with interpretation (“so traffic halves”).
- Different years and different firms listed together, implying a trend line that does not exist.
- Ranking vendors with metrics that have no public baseline (“citation share”, “AI visibility score”). Without a baseline these are internal comparisons only.