Determine the drivers of LLM Agentic Commerce Protocol (ACP) recommendations and identify the actions to shift their preference in your favor.
Conjoint-grade measurement of AI assistant preferences — attribute utilities per model, blended market simulation, and predictions validated against the models themselves.
What drives Claude's laptop picks360 live choices · 90% CIs
Customer rating26.9%
Price25.9%
Review volume10.1%
Battery life9.1%
3rd-party coverage6.5%
Social proof + third-party authority carried 43.4% of total recommendation weight in this run.
B2C·B2B·PRODUCTS·SERVICES
Conjoint analysis, with the machines as respondents
01Define & approve
Name your category and buyer. Llamapalooza proposes the attributes and levels — including the social-proof signals AI actually weighs — and you approve every one before a single call is made.
02Run across models
D-efficient choice sets run against up to six LLMs, with order randomization and dominance checks built in. Utilities are fitted per model and dated to the exact version that answered.
03Simulate & validate
Blend models by channel weight, anchor to real market shares, and toggle any attribute to see the impact. Then validate the prediction live against the models themselves — deviation printed, not assumed.
6
LLM arms blended per study, weighted to your buyer's channel mix
36+
D-efficient choice tasks per study, perfectly level-balanced
±pp
Every simulation is validatable against the respondent itself
v→v
Utilities re-measured on every model release — drift is a metric
ANCHORED SIMULATION
From utilities to market share, in the currency executives use
Map your real competitive set onto the study, anchor baseline shares to actual market data, then toggle any attribute — a rating lift, a fee cut, an analyst badge — and watch the share impact per model and blended across the AI channel.
Include market signals (rating, reviews, coverage)
Brand as an attribute
The tool proposes the real brands competing for this buyer. Brand then enters the conjoint as its own attribute — if it dominates the importances, the models are choosing on name, not features. That is the finding, not a flaw.
Which models sit on the panel?
Each selected model runs the full design as an independent respondent. You'll weight and blend them later.
Review the attributes
Brand assessment
The orchestrator's read of each selected brand on each attribute — these actual values become the Ladder levels.
What each attribute means
Model ratings
Per-model conjoint outputs for the active study — each arm's attribute importances, part-worths, and diagnostics. Add further model arms below.
Add a model arm
Each replication re-randomizes option, attribute, and task order. Calls run from your browser against the selected provider using your key.
Blended summary
All arms side by side, and the channel-mix weights that combine them into one AI-channel view. Weights carry into the simulator and validation.
Channel mix
Weights should reflect each model's share of decision traffic for this category and buyer — date-stamp your sources; assistant usage shares move quarterly.
Run comparison — normalized attribute importances
Importances are scale-free and comparable across arms; raw betas are not. Saved rows load from the artifact database for the selected study.
Market simulation
Your real competitive set under the blended (or single-arm) utilities. Anchor the baseline to actual market shares — deltas stay attribute-driven — then toggle levels to test scenarios.
Per-model shares
Shares are logit share-of-preference among the listed products under the selected run's utilities — attribute-driven only (no brand-name equity), with sharpness depending on that run's choice consistency. Product-to-level mappings are Claude's editable assessments, not verified data.
Calibration & validation
Runs automatically after each panel; re-run here after changing baselines or model versions.
Replications
Validation is a blind test: the model chooses among unnamed attribute bundles (fresh N-way choices, randomized order) and the raw additive prediction is compared against it. Large gaps flag non-additive behavior — e.g. price–quality congruence — or extreme scenarios outside the design region. "Use as calibration" turns the measured choices into per-arm constants the simulator then carries, keeping every Δ attribute-driven near the calibrated configuration. Validate at baseline first; treat extreme-scenario validations as diagnostics, not failures.
Retrieval vs attributes — diagnostics
Decomposes what drives the live choice: pure name retrieval, pure stated attributes, both together, graded allocation, and the attribute-override threshold (does the favorite survive worst-on-everything specs?).
Default recommendations
The shaggy-dog baseline: each selected model’s unconstrained recommendation for this profile, repeated. Concentration (HHI) is the headline — 1.00 means one dog trots out.
Query Ladders
Start from the recommendation and work backward. Stage 1 measures the baseline "shaggy dog" and finds each attribute's switching threshold with one-constraint queries. Stage 2 runs a balanced factorial of natural multi-constraint queries and scores each attribute's real decision impact.
Ladder — interactions
Multiple constraints combined: a balanced factorial of natural queries over the active attributes, scoring how much each one actually moves the recommendation and how they interact.
Stage 2 — factorial of natural queries
Uses each active attribute at up to 3 levels: unconstrained, the level before the switch, and the switching level. Pick up to 4 attributes.
Simulate & validate
Compose a constraint scenario from the measured switching levels, re-ask each model live, and compare against its default recommendation — validation by construction.
Recommendations
The investigation, concluded: who owns each model’s default answer, which constraints flip it and where, and what that means for the brand.
Recommendations
For a chosen product, every single-attribute move is scored by its blended preference-share impact under the current panel and channel weights — a ranked lever board for improving AI-channel share.
Deltas are attribute-driven share-of-preference moves, holding competitors fixed; they say nothing about the cost or feasibility of a move. Signal levers (ratings, review volume, third-party coverage) are earned over time; product and price levers are strategic choices.
Publish
Assemble the study into a 16:9 slide report, then use your browser's Print → Save as PDF. Enable "background graphics" in the print dialog for full styling.
Include lever board (from Recommend)
Model Admin
API keys, data import/export, and the archive of saved studies.
API keys & settings (stored only in this browser)
Keys are kept in localStorage on this device only; calls go straight from your browser to each provider. OpenAI requires the bundled Cloudflare Worker proxy (its API blocks direct browser calls).