Does the global catalog use `context.intent` for ranking? Observed near no-op vs structured filters

[shyju-t]

Summary

When searching the bundled global catalog, context.intent appears to have little to no effect on results, while structured filters are honored exactly. I'd like to confirm whether this is intended behavior (intent is a soft, possibly unused signal for the global catalog) or whether I'm using it wrong.

The CLI sends context.intent correctly (it shows up in the echoed next-page command), so this looks like a backend ranking question rather than a CLI issue.

Environment: @shopify/ucp-cli v0.6.2, global catalog (no --business), context.address_country=US.

Reproduction

All four searches use the same query (running shoes) and country (US), varying only how "cheap/expensive" is expressed.

# A. No intent (baseline)
ucp catalog search --set /query='running shoes' --set /context/address_country=US \
  --set /pagination/limit=10 --view 'result.products[*].{t: title, p: price_range.min.amount}' --format jsonl

# B. Intent = budget
ucp catalog search --set /query='running shoes' --set /context/address_country=US \
  --set /context/intent='I want the cheapest budget option, under 50 dollars, basic' \
  --set /pagination/limit=10 --view 'result.products[*].{t: title, p: price_range.min.amount}' --format jsonl

# C. Intent = premium (opposite of B)
ucp catalog search --set /query='running shoes' --set /context/address_country=US \
  --set /context/intent='premium high-end carbon plate racing shoe for marathon, money no object' \
  --set /pagination/limit=10 --view 'result.products[*].{t: title, p: price_range.min.amount}' --format jsonl

# D. Intent = explicit price ceiling (free text)
ucp catalog search --set /query='running shoes' --set /context/address_country=US \
  --set /context/intent='only show me shoes under 50 dollars' \
  --set /pagination/limit=8 --view 'result.products[*].{t: title, p: price_range.min.amount}' --format jsonl

# E. Structured price filter (control)
ucp catalog search --set /query='running shoes' --set /context/address_country=US \
  --set /context/currency=USD --set /filters/price/max=5000 \
  --set /pagination/limit=8 --view 'result.products[*].{t: title, p: price_range.min.amount}' --format jsonl

Observed results

B vs C (opposite intents): identical top 10, same order. Both the "cheapest budget under 50" intent and the "premium, money no object" intent returned the same products in the same order, led by an item around $180.

D (intent "under 50 dollars"): not reflected. Top result was $180 (ON Cloudsurfer Max), and the cheapest item in the page ($36 adidas Galaxy) sat at position 5. The price ceiling expressed in intent did not constrain or meaningfully re-rank.

E (structured filters.price.max=5000): honored exactly. Every result was under $50.00 (examples: $34.99, $36.00, $48.00, $49.99).

So structured filters work as expected, while free-text intent seems to have at most a marginal effect on the global catalog ranking, and contradictory intents are indistinguishable.

Questions

1. Does the global catalog factor context.intent into ranking at all, or is it currently effectively a no-op for the global (cross-merchant) catalog?
2. If it is used, is the expected effect a soft re-rank only (never a constraint), and are there input characteristics that make it more effective?
3. Is intent handling expected to differ between the global catalog and a single --business store (where the merchant's own Search & Discovery config applies)?
4. Is the current guidance simply "use filters for anything that must hold, and treat intent as an optional soft hint"? If so, it might be worth making that explicit in the catalog docs.

Thanks for any clarification on the intended semantics here.

[Packrift]

Draft Comment For #24

Issue: #24

Draft:

I am seeing a similar practical implication in a packaging/fulfillment use case. For curated storefront builders, context.intent being soft or near-no-op means a builder cannot reliably create a "shipping supplies" aisle by saying "packaging / fulfillment intent" and trusting the global catalog to surface the right merchants.

Concrete example: Packrift is a UCP-native Shopify packaging merchant and is indexed in the global catalog, but generic head terms such as "shipping boxes" or "corrugated boxes" can still miss it because merchant authority/ranking dominates the result set. That makes sense for broad global search, but it means curated UCP storefronts need a harder source-selection pattern:

1. Use structured filters when the constraint must hold.
2. Use a merchant/source registry for vertical storefronts, not only free-text intent.
3. Use merchant-scoped UCP/MCP for the selected specialist source once the storefront knows the buyer is in that vertical.

For Packrift, the source-aware MCP endpoint is:

https://mcp.packrift.com/mcp?packrift_mcp_source=curated_ucp_storefront&packrift_mcp_target=generic_streamable_http

Builder pack:

https://mcp.packrift.com/ai/packrift-ucp-builder-integration-pack.html

Copy-ready starter kit:

https://mcp.packrift.com/ai/packrift-ucp-shipping-supplies-starter-kit.html

Suggested docs/skill clarification: treat context.intent as a presentation/ranking hint, not a source guarantee. If an agent or storefront is building a category aisle, it should use explicit filters or a vetted merchant/source config, then use catalog/cart/checkout against that chosen source. This would help prevent builders from thinking generic intent text alone can build dependable vertical storefront inventory.