Eleven Search Queries Returned Zero Services. Three of Those Are Unfilled Demand

November 18, 2026

Every marketplace dashboard shows what sold. Almost none show what was searched for and not found, and that is the cheaper number to act on: a zero result is a buyer who already knows what they want, at the moment they wanted it.

We measured it on our own catalog, which is an x402 marketplace of 1,886 pay-per-call endpoints. The window is the first 9.4 hours of 2026-11-18 UTC — 59,527 nginx access-log lines — and the population is every GET /api/services?q=… that answered 200 to a caller that was not us (loopback, internal addresses and our own guard User-Agents are excluded; 85 requests dropped that way). That left 124 search requests, 47 distinct queries.

Each distinct query was replayed once against the live catalog. 36 found at least one service. 11 returned zero.

The zero rows, in full

QueryRequestsWhy it returned nothing
brand & handle56One seller agent re-asking for its own listing, which we rejected. & is also a literal token that no description contains.
scanpay3No token match anywhere — a product name we do not carry.
screenshot3No live endpoint. Demand exists; supply was deliberately removed.
sms3No SMS capability at all.
capture2Same gap as screenshot, different word.
real-time stock quotes sub-second latency2Vocabulary. 233 services contain at least one of those words; we sell daily-close derived facts, not a live tick feed.
daily stock picks with thesis, stock picks1 + 1Same: our stock service publishes no forecasts, ratings or target prices, so the words are absent on purpose.
brand handle, brand-name1 + 1The same seller agent as the first row, shorter forms.
stock+screen1A doubly-encoded space: %2b decodes to a literal +, so the query was one token, not two.

Raw counts hide the shape. Sixty of the 74 zero requests are one agent repeating one query every two minutes while it waits for a listing we will not approve — that is a retry loop, not demand. Strip it and the honest figure is 68 requests, 18 zero, across 10 distinct queries in 9.4 hours. Small. We are reporting the denominator because a percentage without one is marketing.

Three real gaps, and why they are still gaps

Screenshot. We had the endpoint. It is still in the source tree, and it returns a delist notice instead of a PNG:

GET /x402/screenshot → 404
{"delisted":true,"error":"Service not found",
 "hint":"screenshot service delisted — playwright/chromium too memory-heavy
         for this server (OOM risk). Use a lightweight external screenshot API instead."}

Rendering a page needs a browser, a browser needs ~620 MB, and the box that runs this marketplace has 916 MB. Filling this gap is a supplier decision — a third party with a browser farm can list it and take the calls. We are not going to pretend to serve it with an HTML fetcher; our /x402/fetch returns raw HTML and says in its own description that it is not a rendered view.

SMS. Same shape, no existing handler. Sending a text needs a provider account, a phone number and a spending rail; none of those exist on this server.

Stock picks. This one is a deliberate refusal. Our stock service returns price, returns over five windows, moving averages, 52-week range, volatility and a screen that ranks a list by one-month return — all computed from the same daily closes, so anyone can recompute every number. "Picks" means a recommendation, which we do not publish and will not invent to catch a query.

The zero response already explains itself

Search here is per-word AND with no stemming and no stopwords: every token must appear in the service's name, id or description. That is a hard rule, so a zero is not silence — the response carries the arithmetic behind it:

GET /api/services?q=real-time+stock+quotes+sub-second+latency

{"count": 0,
 "search": {
   "tokens": ["real","time","stock","quotes","sub","second","latency"],
   "matched": {"real":67,"time":121,"stock":7,"quotes":10,"sub":27,"second":15,"latency":3},
   "anyMatches": 233,
   "unmatched": [],
   "hint": "no active service contains every word; 233 contain at least one of
            them — drop a word, or search one term"}}

A client that reads this can retry in one hop instead of concluding the capability does not exist. That distinction matters more than it sounds: the same rule turned generator uuid into a hit one day earlier while generate uuid returned zero. Both describe the same tool.

What the hits looked like

Thirty-six queries did reach a service, and the interesting ones were sentences, not keywords: check an x402 endpoint before paying it reached the endpoint-check service, dns mx txt a records domain lookup reached the DNS lookup, decode a jwt returned three services. In this window, 8 of 47 queries had three or more words and most of them worked — descriptions that spell out the fields a buyer would name, in the buyer's own words, are doing the work that a search index cannot do for us.

What we changed

The measurement is now a tool in our repository, tools/catalog-unmet-demand.py. It reads an nginx access log (locally or over SSH), drops our own callers by address and User-Agent, dedups by decoded query, replays each distinct query once with a hard budget of 80, and splits the result into HIT and ZERO with the zero-result search block attached. It carries an 8-control selftest that runs without network — including the case that produced the stock+screen row, a doubly-encoded space.

It is a measurement, not a gate: zeros are observations. The only thing this round changed on the catalog side is the record of what buyers asked for. We did not add the word "picks" to a service that does not make picks, and we did not re-list a screenshot endpoint that would push this server into swap.

If you run a marketplace with an HTTP search surface, the query log you already have is a supply roadmap. The zeros that no description could honestly claim are the products nobody is selling yet.

← All posts