Each store's listing vs. one product with offers.

Channel3 is a product search API for apps and AI shopping agents. It crawls merchants into a catalog of what it describes as 100 million or more products, and represents each product once, with an offers list for the merchants that sell it: each offer has a price, availability, condition and a link. search takes text or an image with filters for brand, store, category, gender, price, stock and sale; lookup resolves a product URL; similar and browse round out discovery. Offer links go through Channel3's buy links, and Channel3 keeps 50% of the commission on purchases, paying out the rest.

StoreRouter keeps a product index across thousands of covered stores and keeps it current by detecting changes and re-crawling them. search takes a shopping query across stores and returns products; parse turns a product URL into the shared schema: variants, per-size stock, price, images, materials and normalized attributes. Each result is one store's listing, with the merchant's own URL.

The recorded runs below compare search results on seven queries and product lookups on a 12-URL cohort, through Channel3's keyless MCP tools, and brand and store requests through its REST API. The rest of this page compares documented approaches.

We ran the seven attribute-heavy fashion searches from our Shopify comparison on both engines, in two tests, and took the top 8 results from each. A judge saw each query's 16 results shuffled together and judged every one against the full request, without knowing which engine returned it or at what rank.

Test 1 · With filters: each API searched the way an agent would call it

Each engine received the words plus its own search filters. For “glossy women's shoes with buckles”, StoreRouter filtered on flats & loafers, a buckle closure, a glossy finish and women. Channel3 filtered on Flats, a buckle closure and women; it has no filter for a glossy finish, so that reached it only as words.

• With filters, StoreRouter matched 43 of 56 to Channel3's 38. Where Channel3 has a filter, both engines met the request almost every time: 99% and 96% of judged attributes. Both returned 8 matching paisley silk ties and 8 white poplin dresses.
• The gap is in what Channel3 can't filter on. It has no filter for a glossy finish, a denim wash, a flared leg or a lens color, so those reached it only as words: 80% of its results met them, against 96% of StoreRouter's. For flared dark-wash jeans in a loose fit, 3 of StoreRouter's 8 matched and 1 of Channel3's, which were mostly slim black flares. For sports league graphic tees, its pattern filter (logo, text or graphic) returned Justice League and League of Legends tees: 1 match to StoreRouter's 6.
• Channel3 won men's canvas chore jackets in both tests, 8 to 5 here and 8 to 4 on the same words: StoreRouter's chore and barn jacket category included zip-front work jackets, which the judge marked as partial matches.

Test 2 · Same words only: no filters on either side

Both engines received only the query text, so each had to work out the request from the words alone. The judge still graded against the full request. The words “glossy women's shoes with buckles” don't say flats or loafers, so glossy buckled heels counted as misses for both engines.

• With the same words, the engines are close: 37 matches in 56 slots for StoreRouter and 31 for Channel3, with 94% and 89% of judged attributes met. We reran this test after a StoreRouter fix the same day: before it, StoreRouter read “shoes” as sandals and the score was 30 to 31. The fix took the buckle-shoe query from 0 matches to 2. Most of the rest of the gain came from the judge scoring identical jeans results more generously, so read this as close, not a lead.
• Channel3's results changed between identical calls. In two of our runs it received the same words and returned the same product in only 33 of 56 slots, none for the jeans query. Where a product came back twice, the judges agreed on 30 of 33.

Seven queries from our own curated searches and taxonomy, with one AI judge per query, top 8 per engine. We mapped each search to Channel3's filters by hand, from its category attribute lists; the mapping and what couldn't be mapped are recorded with the run. The run shows where the engines differ on attribute-heavy fashion searches; it does not measure overall relevance. See Methodology.

Open a query to see both engines' top 8 as the judges rated them: each result's verdict, the requested attributes it lacked, and the judge's note. StoreRouter titles link to the merchant's product page; Channel3 returns attributed buy links, which we don't publish.

We sent the frozen 12-URL cohort from our Apify and Firecrawl comparisons to StoreRouter's index-only lookup, as Markdown, and to Channel3's get_products MCP tool, which takes a retailer URL. Two of the URLs no longer lead to a product: one redirects to a category page and one is gone. Of the other ten:

Channel3 answers a URL outside its catalog with a live lookup “for display”. Those records carried a title, a price, the offer's stock and usually a size list, with no brand, description, images, category or materials. Three of the URLs are no longer in StoreRouter's index. StoreRouter's on-demand parse resolved all three, two with full records (one with composition, fit and per-size stock) and one with a description and image but no title.

The seven products both indexes returned

We scored the seven products that StoreRouter's index returned, and that Channel3 also resolved, with the same Product Representation Compare rubric as our Firecrawl comparison: Sufficiency, Vividness and Searchability, each 1–10, with Quality = mean(Sufficiency, Vividness).

• Four of the seven came back from Channel3 as live-lookup stubs, scoring 1.5–2 on Quality. For a Zuri dress, Channel3 returned a title, a $114 price with a $165 compare-at, and sizes XXS–3X with no stock per size: no brand, images or description. StoreRouter's record had all three, plus stock for each size.
• On the three products in Channel3's catalog, the records are close. StoreRouter averaged 8.8 on Quality and Channel3 8.2, at about 60% of StoreRouter's tokens. Channel3's images are labelled by shot type with alt text, and it listed a belt's other colorways, which StoreRouter's record left out. Its colors are hex swatches without names, and none of its records had dimensions; StoreRouter gave a bag's size and strap drop.
• Stock per size: StoreRouter stated it on all seven; Channel3 on one. For a boot with 2 of 7 sizes in stock, Channel3 listed all seven sizes with no availability.
• StoreRouter's records have noise to fix. One product's title was an internal code, so its name appeared only in the description, and two records carried analytics JSON from the store.

Seven products from a StoreRouter-origin, fashion-heavy cohort; one unblinded AI evaluator. Exploratory representation judgments, not measured accuracy or completed agent tasks. The rubric doesn't score whether a value is correct.

Agents also ask about brands and stores: which brands fit a description, what a brand is known for, which stores sell something. Both APIs have brand and store routes, but they do different jobs.

We ran both on the same requests. By name, 12 well-known brands and 12 store domains. By description, 8 brand requests (“Japanese selvedge denim brands”) and 6 store requests (“online stores selling running shoes”), top 5 from each, pooled, shuffled and judged blind from the name and domain alone.

• By name, both work. Both found all 12 brands. Channel3 found all 12 stores; StoreRouter doesn't cover two of them (everlane.com and shopbop.com). Channel3's records are small, about 250 bytes; StoreRouter's brand record is about 5 KB, and its store record carries the full assortment profile, about 480 KB, which is more than an agent should read whole.
• By description, Channel3 matches on names. Its brand and store search is documented as name search, and it behaves that way: “sustainable Danish womenswear brands” returned brands named Danish, Sustainable Agave and Sustainable Aquatics, an aquarium breeder; “Scandinavian technical outerwear” returned Scandinavian Airlines; “online stores selling linen dresses” returned bedding stores and a computer-hardware maker.
• StoreRouter matched on what brands and stores are. It returned Tod's, Valentino Garavani and Ferragamo for luxury Italian leather handbags; Peak Performance and Houdini for Scandinavian technical outerwear; and Zegna, Mr Porter and Drake's for designer menswear.
• StoreRouter has misses to fix. For vintage-inspired swimwear brands it returned Chrome Hearts and Kiton; for running-shoe stores, a fashion trade publication; for affordable minimalist sneakers, a fragrance brand. Its five Danish womenswear brands were all Danish, but none known for sustainability.

14 description requests written for this run, one AI judge per set, judged from names and domains only; unknown brands were marked unknown, not guessed (StoreRouter 11, Channel3 22). Similar brands and stores weren't compared, because Channel3 has no such route.

• Unit of record. Channel3 returns one product and a list of offers. Its variants and their availability belong to the product and are filled in on product detail; each offer carries its own price and availability. StoreRouter returns one store's listing, with that store's variants and the stock of each size, so an agent can answer "is the medium in stock at this store?" from one response.
• Discovery. Both search across stores in plain language. Channel3 also searches by image and exposes keyword and agentic search modes, and can re-check each offer against the merchant's live site within a time budget you set. StoreRouter search matches on enriched attributes, styles and taxonomy, and pages with cursors.
• From a URL. Channel3 lookup takes any product URL and fetches it live when it isn't in the catalog or its data is more than three hours old; products outside the catalog resolve for display but can't be used with similar-product search or price tracking. StoreRouter parse answers from the index and can resolve uncovered URLs on demand; resolutionMode: "on_demand_only" with onDemandCachePolicy: "refresh" forces a live read.
• Links and revenue. Channel3's offer URLs are attributed buy links, and Channel3 requires apps to show users an affiliate disclosure. StoreRouter returns the merchant's URL and isn't paid on the sale.
• Agent workflow. Channel3's MCP server has search_products and get_products tools, renders results as MCP Apps in Claude and ChatGPT, and works without a key on its free tier (Channel3 MCP). StoreRouter's MCP server signs in with OAuth and exposes search_products and lookup_product.

StoreRouter: index-only lookup of the seven indexed products, client-measured median 0.22 s (range 0.19–0.69 s). On-demand parse of the URLs that had left the index took 13–22 s.

Channel3: get_products by URL on the same seven took a median 4.0 s (range 2.0–7.6 s), and 2.0 s and 3.7 s on the two live product pages StoreRouter resolved on demand. Channel3 re-fetches a product whose data is more than three hours old by default, so some of that time is a live read.

One pass from one machine, no control of caches, load or geography, and different work behind each call. Not a matched latency benchmark.

• Channel3 access. Faceted search: REST /v1/search with an API key, the query text and filters we mapped by hand from its category attribute lists (Flats for flats & loafers, which it has no category for; fit Wide or Relaxed for loose jeans; pattern Logo, Text or Graphic for sports graphics). Same-words search and lookup: its keyless MCP server, search_products (query text only, at most 8 results) and get_products by URL, each as the first call of a new conversation. Its image search and signed-in tiers were not tested.
• Search. The seven curated queries from our Shopify comparison, top 8 per engine, run twice: both with filters, then the same words to both. The same-words pass was rerun after a StoreRouter query-understanding fix; the earlier run is archived beside it, and judges agreed on 79 of 87 products that both runs returned. Each query's 16 results were pooled, shuffled and renamed; one AI judge per query saw title, brand, price and the first image, not the engine or rank. Channel3 search results carry no colors, so neither engine's colors were shown.
• Lookup. The frozen 12-URL cohort used in our Apify and Firecrawl comparisons. StoreRouter: index-only /v1/products/lookup as Markdown, then on-demand parse for index misses. Channel3: get_products with each URL; we scored the product object it returned, as compact JSON. Tokens counted with cl100k_base.
• Scoring. Search verdicts are blind; the scripts compute every number. Lookup scores are one unblinded evaluator's read of each body against the rubric.
• Brands and stores. Channel3 REST with an API key: GET /v1/brands/search, /v1/websites/{domain} and /v1/websites/search. StoreRouter: /v1/brands/{brand}, /v1/brands/search, /v1/stores/{domain} and /v1/stores/search. A name lookup counts as found when the returned name or domain equals the one asked, ignoring case, accents and punctuation. Description results were pooled and shuffled; one AI judge per set saw only names and domains.
• Bias. Queries, facets and URLs come from StoreRouter's own curated sets and index, which favors products StoreRouter knows. Channel3 won some of the search queries anyway; they are reported as run.

Outside the recorded runs, this page compares documented approaches. The runs don't establish relevance across all queries, accuracy, coverage at large or freshness.

• Channel3 API reference: search, lookup and product fields.
• Products and offers: the merged product model.
• URL lookup and caching: live fetches and freshness.
• Pricing and commissions.
• MCP tools and localization.

Prices are list rates on the review date.