Tilltrend

The AI analyst for your store. Every answer shows its work.

Ask questions about your store in plain language. Every figure Ask gives you carries a receipt: the governed metric definition, the exact SQL, and a hash you can replay without the AI in the loop.

On the Shopify App Store · built in public

What an answer actually looks like

The same question, answered twice. The chat answer is an illustration of a number with nothing behind it. The Tilltrend answer is not a mockup: it comes from one real run of the product answering "What was total revenue?", executed against the demo warehouse.

A chat answer (illustration)

What was total revenue?

Your total revenue was about $29.4 million. Sales look strong overall.

  • Which orders does it count?
  • Are refunds taken out?
  • Would it say the same tomorrow?

The answer cannot tell you.

Tilltrend (real output, demo dataset)
Total revenue no date filter
$29,356,250.00
  • Revenue after discounts: refunds are not subtracted and cancelled orders stay in.
What this means written by the AI, not part of the receipt

Total revenue was 29,356,250. This represents revenue after discounts; the figure includes cancelled orders and does not deduct refunds.

Two renderings of one stored value: the card formats it as money at the display layer, the receipt keeps the exact number.

RECEIPTrcp_7998c101b59b4301a29132c719f414d0
metric
total_revenue
scope
no date filter
value
29,356,250
hash
sha256:26ff1aad7b2136ec
The governed SQL behind 29,356,250
SELECT SUM(f.sales_amount) AS total_revenue
FROM gold.fact_sales f
replay, no model
MATCH
post-check
PASS

How it works

Ask in plain language

"What was total revenue?" "Compare June and July." "Which countries drive revenue?" The model picks the right governed metric and writes the prose. It never writes the SQL and never touches your data directly.

Every figure Ask gives you carries a receipt

Answers come in two visible sections: receipted facts (numbers produced by governed SQL, each with a receipt id and hash) and interpretation (model prose, clearly labeled as not receipted). Any receipt can be replayed model-free; a doctored one fails.

Your warehouse is yours

Tilltrend builds a real bronze/silver/gold warehouse from your store data: on PostgreSQL you control, exportable, inspectable. Built so that leaving is easy, which is exactly why you won't have to.

Check it yourself

Reproduce the answer above with:

python app\scripts\ask.py "What was total revenue?"

The exact audited text stays one click away in the app, unchanged to the character. Here it is in full: the verbatim terminal output of that same run.

Every number traces to one query. The SQL on the receipt is the governed statement the deterministic layer executes for total_revenue and stores in the receipt's execution block. The model only chose the metric; it never writes SQL. Tilltrend's Ask page shows this beneath every figure it gives you.

The 29,356,250 is a property of the demo dataset, not a revenue claim about this product or anyone's business. Replaying that receipt re-executes the same governed SQL with no model involved and grades the result. Observed verdict: MATCH. Doctor the receipt file and it fails.

See every claim with its reproduction command →

Real output · demo dataset (public training data, 2010–2014 bike sales, not real merchants)
question: What was total revenue?
turn:     turn_fab1d16b8a8c   mode: model   model: claude-haiku-4-5-20251001

RECEIPTED FACTS
===============
(1) [run_metric] total_revenue (no date filter)  <rcp_7998c101b59b4301a29132c719f414d0 | sha256:26ff1aad7b2136ec>
  total_revenue = 29,356,250
  caveat: Revenue after discounts: refunds are not subtracted and cancelled orders stay in.

INTERPRETATION (model, not receipted)
=====================================
| Total revenue was 29,356,250. This represents revenue after discounts; the figure includes cancelled orders and does not deduct refunds.

receipts:
  rcp_7998c101b59b4301a29132c719f414d0 | sha256:26ff1aad7b2136ec
replay:   python app\scripts\replay.py rcp_7998c101b59b4301a29132c719f414d0
postcheck: PASS (attempts=1)

Where this is honest by construction

Tilltrend is new. No merchant counts, no testimonials, no revenue-lift percentages appear on this site. What we do have is a working product whose every claim is reproducible: the evidence page lists each one with the exact command and the observed result, and the changelog records only shipped, verified work.

Install Tilltrend

Tilltrend is on the Shopify App Store. If "an analyst that shows its work" sounds like what your store is missing, install it from the listing. We're also looking for a small number of merchants to shape the product: if that's you, write to us. A human reads every mail.

Install on Shopify