PowerUp Coaching — Living Knowledge Base

That AI Search That Came Back Empty? Run It Again.

5 min read

Topic

Why negative results from AI research have a shelf life, and why the improvement that expires them is in the tooling rather than the intelligence.

Target Reader

A knowledge entrepreneur, founder, or consultant who uses AI for research and due diligence — competitor scans, prior art, market checks, sourcing claims. They have, at some point, concluded “there’s nothing out there” on the basis of an AI search and made a decision on it.

The Fear / Frustration / Want / Aspiration

The fear is specific and worth naming early: what did I build on a null result? Everyone in this audience has cleared at least one decision with an AI search that found nothing, and none of them have gone back to check.

Before State

Treats an empty AI search as a finding — “this doesn’t exist.” Tracks AI progress through reasoning benchmarks and model releases. Doesn’t re-run old research, because re-running something you already did feels like the opposite of productivity.

After State

Reads a null result correctly: the retrieval path available on that date didn’t reach it. Watches the tooling layer as an independent axis of improvement — better hooks, better query construction, deeper source traversal — moving on a different schedule than reasoning. Keeps a null-results file and re-runs it before decisions that lean on old absences.

Narrative Arc

Scott Delinger runs a patent search across Gemini and Claude in January for the invention he’s designing. Nothing. He takes Q2 off, picks the project back up in late June, and decides to redo his early groundwork “just in case the tools are better.” The same query surfaces a patent filed in 2016 and awarded in 2018 — public and indexed the entire time. The turn is his diagnosis: the models probably didn’t get smarter about patents, “they’ve got better hooks in how to do queries.” The resolution is what he does next — asks whether the holder filed internationally, learns that multinational filings link from the same source, sends the model to check, and converts a six-month setback into a usable position.

Core Argument

AI research results decay, negative results decay fastest, and the decay is driven by tooling improvements that don’t show up in any benchmark you’re watching — so the correct question after a model release isn’t only “does it reason better” but “what can it reach now that it couldn’t reach before.”

Key Evidence / Examples

  • The dated before-and-after: empty in January, found in July, patent awarded 2018 and public throughout. The data never moved; the retrieval did.
  • Scott’s own read on the mechanism — better hooks, not better intelligence — from someone who had no incentive to be charitable about a finding that set his project back.
  • The follow-through as method: US-only filing, no international counterpart, held by Westinghouse Air Brake and apparently unused since award. He asked the second question, and the second question is where the value was.
  • Lou’s corroborating observation from the same session: with several research tools installed, Claude increasingly routes research through third-party APIs — sometimes when he’d prefer a plain web search. The retrieval layer is where the movement has been.
  • The uncomfortable corollary for the reader: due diligence performed with older tooling was performed with weaker reach, and nobody has re-checked.

Proposed Structure (5–7 beats)

  1. Open on the January null and the July hit — same query, same data, six months apart. Land the fact that the patent was public both times.
  2. The setback: this was bad news for Scott’s project. Establish that he had every reason to prefer the null result.
  3. The mechanism: tooling improves independently of intelligence, on its own schedule, largely invisibly. Name it as a second axis of progress.
  4. The implication that stings: your negative results were never findings, they were snapshots of what your tools could reach.
  5. The practice — the null-results file. Query, date, tool. Quarterly re-run. Fifteen minutes, no re-derivation.
  6. Scott’s second question, and why the follow-through matters more than the find. A hit is not a conclusion.
  7. Close on the reframe: stop asking whether the new model is smarter, and start asking what it can now reach.

Editorial Notes

The most immediately actionable piece from this session — protect that. The null-results file is a fifteen-minute practice with a clear payoff and should be impossible to miss in the draft. Scott’s story does the persuading; don’t editorialize over it, and keep the specifics (dates, the 2018 award, the company) because the specificity is what makes it land. One caution: do not let this become a piece about patent search — that narrows the audience to almost nobody. Patents are the illustration; the argument is about every category of research where “I checked and found nothing” ended an inquiry. Natural companion to Brief - Your Growing Trust in AI Is the New Attack Surface — that one is about over-trusting what AI finds, this is about over-trusting what it doesn’t.

Next Step

  • Approved for drafting
  • Needs revision
  • Deprioritised

Graph View