Research
What we measured, and what it does not prove yet.
Every Pulse product rests on a measurement. Here is each one: what we ran, on what data, the result, and the limit you would want to know before buying. Where a product has not met a real merchant's data yet, we say that too.
Findings, product by product
Decision Audit
MeasuredHow empty a field is and how much shoppers need it disagree.
Ranking attributes by the buying requirements shoppers state gave a different order from ranking by how empty each field is. Processor was missing on 162 products, so coverage puts it near the top; only 2 of 32 stated requirements mention a processor, so demand puts it last. Screen size came first under both.
- Data:
- A 216-product laptop shelf, 156 of them synthetic, and 32 written buying requirements. Criteria fixed before the run.
- Limit:
- Only the top position is stable; positions two to four move with the method. One shelf.
Catalogue and photographs
MeasuredHalf of what a catalogue records, the photograph already shows.
Across our 2,176-product demonstration catalogue, 52.6% of recorded attribute values repeat what the product photograph shows. On the nine facets a laptop buyer weighs, the catalogue and the photograph overlap on none.
- Data:
- 2,176 products, one image each, counted attribute by attribute.
- Limit:
- It measures the overlap, not lost sales. Whether a photograph shows a fact is our judgement.
Search, Recommendations and Pages
MeasuredKnowing the products is worth more than knowing the shopper.
On the laptop shelf, learning what a shopper needs narrowed 216 products to about 94, which is 1.20 bits. Knowing the products' facts narrowed those 94 to 5, which is 4.24 bits. That is why the catalogue products come first, and why live ranking needs the facts to rank on.
- Data:
- The same 216-product shelf, scored in bits of decision information.
- Limit:
- One shelf, partly synthetic. It compares two kinds of missing information; it is not a ranking result.
Search
MeasuredOur query parser misread most Thai searches. Now it does not.
A test set of shopping searches found the parser misreading 29% of Vietnamese, 94% of Thai, 78% of Bahasa Indonesia and 34% of English searches outside Vietnam. After the fix it misreads none of the four, with held-out cases kept back to check.
- Data:
- Searches phrased with help from a public set of synthetic Vietnamese personas, used for variety of wording only. English priced in six Southeast Asian currencies.
- Limit:
- Synthetic searches, not real search logs. It shows the parser reads the words, not that results improved.
Decision API for AI agents
MeasuredA fact missing from a listing's fields is usually still on the page.
On nine laptop listings from a large Vietnamese marketplace, 17 of 36 facet readings were missing from the structured fields, and all 17 were in the title or the description. An agent that reads can find them. The hard part is reading correctly: two values for one fact, units, and saying it cannot tell when the page is silent. That is what the Decision API answers.
- Data:
- Nine public listings, fetched once.
- Limit:
- We pre-registered 60 listings and reached 9, so there is no verdict. We stopped rather than get around the site's access limits.
Promotion Audit
Waiting on a first merchantBuilt, and tested only on a synthetic seller.
The Promotion Audit judges vouchers, price cuts and free shipping on orders that stayed bought, after returns and failed deliveries, against doing nothing. An offer that wins only within noise is marked as one to test, not one to run.
- Data:
- A synthetic seller built from published fee, delivery and payment figures, with assumptions where no source exists.
- Limit:
- No real seller yet, so no figure is published. A real result needs the seller's own order history.
Return Verification
Waiting on a first merchantReady to read a merchant's returns. It has not read one yet.
Return Verification reads order lines and return reasons against the catalogue, to show which missing facts go with returns and whether a repair moved them. The method and the free /outcomes endpoint are built and tested on fixtures.
- Data:
- Test fixtures only.
- Limit:
- We do not claim Pulse reduces returns. It measures your own history.
Decision Verification and Monitor
Waiting on a first merchantThe audit's measurement, run again.
Verification re-runs the audit after a repair and reports what moved. Monitor re-runs it on a schedule and says when a shelf slips. Both use the audit's measurement, so they carry its limits.
- Data:
- No real repaired catalogue yet.
- Limit:
- Neither has run on a merchant's repaired catalogue.
How we run research
- 01
Criteria first
What would count as success or failure is written down and committed before any data is fetched, so a result cannot be chosen after it is seen.
- 02
Failures stay up
A run that fails its own test is kept and published with the reason. So is a run stopped short, like the nine listings above.
- 03
Data is labelled
Synthetic data is called synthetic, and no number becomes a claim about results until it has run on a merchant's own data.
Questions about the research
Has Pulse been proven on a real merchant?
Not yet. The measurements here run on our own laptop shelf, our demonstration catalogue, synthetic data or a small public sample. A first merchant's own data is the next milestone, and results will be published here.
Why publish results that are unfavourable?
Because a buyer will find the limits eventually, and it is better they hear them from us. The photograph study exists because a result contradicted our own product.
Do you have an uplift figure?
No. An uplift figure needs a holdback on a merchant's live traffic, and none has run. We will not quote one before it has.
Can I see the method behind a number?
Yes. Ask us for the write-up behind any finding on this page. Each one has its pre-registered criteria, its data and its caveats.
Run the measurement on your own catalogue.
The Decision Audit is free. Send one category and get back which missing facts cost the most buying decisions, with the limits stated.
Request a Decision Audit