What-If Organizations

Research artifact · proof of concept

Real firms, simulated under alternative theories.

2026-11-01T12:21:20.560192 image/svg+xml Matplotlib v3.9.4, https://matplotlib.org/ a Architecture Archive public snapshots of each firm's website, several a year, 2017 to 2026 Measurement a language model codes each firm-year under protocols; every statement keeps a quote What-if organization an agent built from the firm: anchors, a theory text, memory of its own decisions Simulate the same firm and record, decided year by year under no theory and each theory Calibrate, validate a mix of the conditions fitted before the pandemic, tested after it against the record What-if experiments another theory, no anchors, or its own history: outcomes the record cannot show Measure Simulate Compare LM record mix b to g: what the data look like at each component; values illustrative b Archive captures per year and page role 17 18 19 22 23 24 25 26 home products news about careers pandemic every page carries a Common Crawl pointer (file, offset, length, digest), so that anyone can fetch the same page and repeat the coding per firm and year, 15 to 50 pages are read, up to 45,000 characters, with quotas by page role c Measurement anchors of the firm offers what it sells customers to whom why it wins its theory of value situation industry, size, projects record of its new products 18 19 22 23 24 25 26 product A product B × product C home page own page mentioned × taken off the market a stated result, with its quote “… now in use at forty sites” — verbatim, or it is not kept d What-if organization what the agent reads in year t the firm stated strategy and situation, from its website the theory what management weighs when results come in the record to t−1 A: own page, result stated · B: mentioned only its own memory earlier decisions and reasons (multi-period) what it answers A keep p = 0.15 B drop p = 0.60 “B falls short of aspiration and ties up resources the core line needs …” p: probability of withdrawal e Simulate under theories probability of withdrawal in year t, five copies of one firm none BTF AC ESC RO record product A 15 20 12 10 18 kept product B 35 60 30 20 45 × product C 10 12 25 8 10 kept product D 25 30 40 15 20 × one row per product and decision year; the agent never sees the year it decides about every cell keeps the reason the agent gave, so that a theory can be read at work BTF behavioral theory · AC absorptive capacity ESC escalation · RO real options f Calibration, validation fit the mix: products of 2018, decided 2019 all firms by regime per firm regime and firm weights shrink toward the common mix test it: products of 2023–24, decided 2024–25 0% 15% 30% regime 1 regime 2 regime 3 regime 4 recorded share withdrawn predicted by the fitted mix scored by level, log loss and ranking against the record g What-if experiments the same firm, simulated otherwise another theory swap the text, keep the firm no anchors the firm reduced to industry, size, age own history its own first-year decisions as its past own history: share of the first-year cohort still on the market 100% 50% launch year 1 year 2 ESC record AC RO BTF the agent decides year 1, then carries its own decisions into year 2; the cohort is compared with the record escalation persists, the behavioral theory withdraws on weak results

How to read the figure. Panel a is the instrument in three stages: the archive and the measurement build the data, the what-if organization and its simulation produce decisions, and the calibration and the what-if experiments compare them with the record. Panels b to g show the data at each stage with illustrative values; the real data are on the three pages. b, each website is captured several times a year from 2017 to 2026, except for the pandemic years, and every page carries a Common Crawl pointer. c, a language model codes each firm-year under written protocols with a quote behind every statement: the anchors of the firm and the record of its new products, their prominence, their stated results, and the year they leave the market. d, the agent reads the firm, a theory text, and the record up to the year before the decision, and answers with a probability of withdrawal and a reason. e, the same firm is decided under no theory and under each theory. f, a mix of the conditions is fitted on the decisions of 2019 and tested on the held-out decisions of 2024 and 2025 against the record. g, the same firm is simulated otherwise, with another theory, without its anchors, or with its own earlier decisions as its past.

FirmsThe web data from which each what-if organization is built, the record of its products, and its decisions under every theory with the reasons given.Open → SimulatorThe results for any choice of firms, theories, and model, computed from the published decisions.Open → ValidationLevels and ranking on the held-out window, what the theories attend to, the simulation on the firm's own history, and the three models side by side.Open →

Firm database

All firms

The table sorts by any column; a row opens the firm above.

Simulator

The data, the theory, and the model of a what-if organization, with the result.

Computed in the browser from the published decisions; the address of the page carries the selection.

Web anchors

Validation

Does the reference organization behave like the firm?

External validation · against the record

Share withdrawn

Held-out window, the decisions of 2024 and 2025 about the products of 2023 and 2024, never used for fitting. Navy, the recorded share with its 95 percent interval from resampling firms; dots, the mean probability of each agent with the same interval.

Ranking of the withdrawn products

The three models side by side

Internal validation · is the agent built as designed

What each theory attends to

Mean probability of withdrawal over all decision years on the two product attributes the theory texts name. Filled circle, the agent with the firm's web anchors; hollow circle, without them; square, with the anchors of the theory's own constructs removed; navy line, the record. A theory that moves the filled circle away from no theory and loses the move when its own anchors are removed acts through the anchors its text names.

The agent layer by layer

Each row adds or removes one layer of the what-if organization, from the language model reading the product record alone to the full agent and back; the columns show what the layer changes. Navy lines, the record and the regression on the product record.

The firm's own history