analogical-reasoning

Analogical Reasoning (in Strategy)

Confidence 0.80 · 3 sources · last confirmed 2026-09-09

The use of analogies between a known source case and an unknown target case to inform strategy formulation, communication, or evaluation. Logicians regard analogy as a weak inductive form, yet practitioners use it pervasively. The wiki’s anchor source (Carroll & Sørensen 2024) argues analogies should be disciplined, not banished.

Working definition

Two roles of analogy in strategy discourse (Gentner 1982):

Role	Use
Rhetorical / persuasive	Vivid metaphor for stakeholders (“Glassdoor is Tripadvisor for jobs”)
Generative / problem-solving	Surface candidate causal mechanisms; develop firm-specific theory of value

A predictive analogy has the structure: source A has features a₁, a₂, … and outcome a_c; target B has features b₁, b₂, … (where b_n is similar to a_n); therefore the unknown b_c is plausibly similar to a_c.

Key claims

Why analogies dominate strategy discourse despite weak logical foundations

1. Efficient communication — a one-phrase analogy carries vast detail.
2. Concrete and memorable — material analogies vivid; abstract theories aren’t.
3. Predictive even when conceptual understanding is limited.
4. Generative problem-solving tool — best solutions often generalize from familiar puzzles.
5. Intermediate-level abstraction — neither too specific nor too abstract; matches how busy executives reason.
6. Success-story anchoring — the source firm worked, providing existence proof.

Practical tools for disciplining strategy analogies (Carroll & Sørensen 2024)

1. Decompose the global analogy into atomic feature premises.
2. Add negative analogies (where source and target differ).
3. Distinguish horizontal vs. vertical relations (Hesse 1966):
   • Horizontal = surface similarity between source and target features.
   • Vertical = causal relationships within source that produced its outcome.
4. Test the vertical relations — does the source’s success theory plausibly transfer?
5. Build multiple analogies in parallel — increases predictive ability.

Connection to the theory-based view of strategy

The theory-based view (TBV) of strategy (Felin & Zenger 2009, 2017) asks executives to develop firm-specific theories of value:

1. What is your firm’s theory of value?
2. Is your theory novel, simple, elegant?
3. Is it falsifiable / generalizable / generative?
4. Who must you convince for your theory to be realized?

Analogies aid TBV in theory discovery, wider-team theory engagement, identifying unique aspects, and communication.

Worked example: Glassdoor / Tripadvisor

Tripadvisor (source)	Glassdoor (target)
experiential-good info (hotels)	experiential-good info (jobs)
free + open access	free + open access
user-generated content	user-generated content
five-star ratings + detailed reviews	same
collects data → reports + trend prediction	same
ad/referral revenue	same
market success	plausibly similar

Negative premises (where the analogy fails or weakens): travel events vs. ongoing employment; willingness to disclose negative experiences; user registration requirements.

AI-related analogy noted in passing

The paper’s footnote-level example: a leaked Google memo (Patel & Ahmad 2023) characterized the open-source AI threat to GPT-4 as “if GPT-4 is the Walmart you go to for apples, what happens when a fruit stand opens in the parking lot?” — illustrating the analogy form in a current AI strategy debate.

Analogy as a cognitive primitive, not just a strategy tool ( Stanford GSB 2026)

The wiki’s second source widens the concept beyond strategy discourse. Douglas Guilbeault (Stanford GSB) names metaphor and analogy as one of the non-statistical mechanisms by which humans reason — alongside an aesthetic “vibes” sense and intuition — and uses this to argue a limit of LLMs. His claim: LLMs learn by brute-force statistical optimization over massive data, whereas humans “do a lot with a little” partly because they reason analogically and metaphorically, achieving conceptual leaps rather than smooth, within-distribution steps. This reframes analogy not as a weak inductive form to be disciplined (the Carroll & Sørensen strategy framing above) but as a load-bearing primitive of human cognition that current AI architectures may be structurally unable to reproduce. The two sources are complementary: Carroll & Sørensen treat analogy as a strategy methodology to be made rigorous; Guilbeault treats it as evidence of a human cognitive capacity AI lacks. Confidence rises to 0.8 on this second, cross-domain corroboration.

Debates and supersession

• Weak inductive form vs. load-bearing cognitive primitive. Logicians treat analogy as a weak inductive inference (the Carroll & Sørensen framing: useful but to be disciplined with negative analogies and vertical-relation tests). Guilbeault inverts the valence — analogy/metaphor is a strength, a primitive that lets humans leap where statistical optimizers cannot. Not a contradiction so much as two altitudes: normative (how to use analogy rigorously in strategy) vs. descriptive (what analogy reveals about human cognition).
• Scope drift (2026-06-25). The page began strategy-specific (“Analogical Reasoning (in Strategy)”) on a single source; the Guilbeault ingest widens it toward analogy-as-general-cognition. If a third, cognitive-science source lands (Gentner’s structure-mapping; Hofstadter), the page may warrant splitting the strategy-methodology layer from the cognition layer.
• Open contradiction with LLM-analogy optimism. Whether LLMs can do generative/causal analogy (not just rhetorical) is unsettled — see Open questions. Guilbeault is the pessimist anchor; a source demonstrating strong LLM analogical transfer would create a genuine contradicts edge.

Related concepts

• Stanford GSB — analogy/metaphor as a non-statistical human cognitive mechanism and a proposed limit of LLM optimization.
• theory-based-view — the framework analogical reasoning supports in Carroll & Sørensen 2024’s framing. Disciplined analogy is the practical methodology for TBV theory-discovery; TBV is the falsifiability layer that disciplines what counts as a good analogy.
• strategic-foresight — analogies between historical convergence cycles and the current period are explicitly invoked in FTSG 2026 (industrial revolution, post-WWII, late-1990s).
• dynamic-capabilities — sensing involves analogizing across industries, contexts, time periods.
• Anand-Wu’s 2×2 — itself a kind of generative analogical structure for matching deployment context to AI use case.

Open questions

• Two sources now (Carroll & Sørensen on strategy methodology; Guilbeault on analogy as cognition). Related Gavetti & Rivkin work (2005, 2014) and Felin & Zenger TBV writings would deepen the strategy side; a cognitive-science source on analogy-making (e.g. Hofstadter, Gentner’s structure-mapping) would deepen the new cognitive-primitive side.
• Open question whether LLMs can serve as effective analogy-generation tools — the jagged-frontier suggests they may be good at the rhetorical role and weaker at the generative/causal role, and Guilbeault argues their statistical-optimization approach is structurally weak at exactly the leaping/insight that analogy enables.

Isolating transfer from recognition: the masking protocol (added 2026-09-09)

[[2026-03-11-allen-mcdonald-how-well-can-ai-do-strategy-simulation-benchmark|Allen & McDonald (Strategy Science, 2026)]] contributes a method this page can use: an experimental design that separates reasoning from general principles (analogical transfer) from recognising a case you have already seen (retrieval).

Their benchmark runs on a published MBA simulation whose teaching materials and solutions are on the open web, so a model might simply recall the winning playbook. Their defence is a deterministic masking function that rewrites every case-specific identifier into a synthetic equivalent — the firm, both technologies, all three customer segments, every performance dimension — plus a from-scratch rewrite of the background brief. The design principle is stated exactly right for this page’s purposes: mask recognition, not knowledge. General strategic concepts (disruption, exploration versus exploitation) stay fully available, “because leveraging such general knowledge is precisely what we aim to evaluate.” What is removed is only the lexical fingerprint that would license retrieval.

The validation makes the distinction concrete. Asked about the simulation by its real name, four of six models recite the published playbook. Asked about the masked version, none can help: four invent a fictitious simulation and two admit no knowledge. The same underlying problem, presented twice, produces confident retrieval in one condition and nothing in the other — which is about as clean a demonstration as the corpus has that surface cues, not structural similarity, were doing the work.

This is a reusable instrument. Any claim that a model “reasoned by analogy” to a known case can be tested the same way: strip the identifiers, keep the structure, and see whether the reasoning survives.