Claim
Between 65% and 85% of the prompts that drive AI-search activity have no keyword equivalent in standard SEO tools. Keyword tools see Google's autocomplete and search-volume data; they don't see what people type into ChatGPT, Claude, or Perplexity, and the shape of those prompts is materially different — longer, more conversational, more often loaded with specifics, more often follow-ups in a session. Operators relying on keyword tools to plan AI-search strategy are looking through the wrong instrument; they're seeing a small biased slice of the actual demand surface.
Mechanism
A keyword tool reads search-engine logs (or scrapes autocomplete) and aggregates query strings. AI-search interfaces don't expose query logs in any equivalent way. The query patterns also diverge structurally: a Google query is optimised for one-shot keyword matching ("crm comparison"), an AI-search prompt is optimised for natural-language back-and-forth ("which CRM has the best integration with X if my team is mostly remote and we already use Y"). The longer, specifier-loaded prompt produces different content needs (specifier-rich pages, comparison-shaped pages, scenario-specific guides) than the keyword approach surfaces. Planning AI search by extrapolating from keyword tools systematically over-rotates to head-term-shaped content and under-rotates to specifier-rich content the AI prompts are actually using.
Conditions
Holds when:
- The category has meaningful AI-search volume (B2B SaaS, anything where buyers use ChatGPT/Claude as a research tool).
- The team can run prompt-coverage measurement instruments (Profound, Otterly, internal scripts) to see what queries actually fire.
- The work has any specifier richness — the gap matters less in genuinely commodity categories where head terms still describe the demand.
Fails when:
- The category is tiny enough that AI-search volume isn't yet meaningful; keyword tools remain the dominant signal.
- The team has no instrument other than keyword tools and can't acquire one; the gap is real but unactionable.
- Prompt patterns shift faster than the team can track; over-fitting to last quarter's prompt-coverage report misses the new shape.
Evidence
"65 to 85 percent of ChatGPT prompts are invisible to keyword tools."
— Kevin Indig, Growth Intelligence Brief #17, https://www.growth-memo.com/p/growth-intelligence-brief-17, 2026-04-17.
The brief includes the underlying analysis: prompt samples from third-party data, side-by-side comparison against keyword-tool coverage of the same intent, and the structural shape difference (length, specifier density, follow-up frequency).
Signals
- The team's content briefs cite prompt samples, not just keyword volumes.
- Prompt-coverage measurement runs on a regular cadence and feeds the editorial calendar.
- Specifier-rich and scenario-shaped content is a clear category in the publishing plan, not an afterthought.
- "What ChatGPT actually returns for this query" is a routine review item before publishing.
Counter-evidence
- The 65-85% figure is a snapshot; the gap will close partially as keyword tools begin to integrate AI-search data sources. Don't quote the number as if it were stable.
- For some commodity categories the keyword-tool signal is genuinely sufficient; the AI-prompt gap doesn't manifest as content-strategy difference. Use the framing where it actually changes the work.
Cross-references
- Citation rate and mention rate are different metrics; comparative content closes the gap — same author, complementary finding on which content types AI surfaces actually cite.
- Measure AI search on three layers: Presence, Readiness, Business Impact — Solis's measurement framework operationalises Presence layer using prompt sweeps that close exactly this visibility gap.
- Owned-brand authority is now the only defensible organic asset — middleman content layers erode regardless of quality — Lily Ray's parallel on how the demand shift is reweighting content economics.