The Demand Side: 59,397 Searches in an Agent Directory
This content reflects minia2a as of its publication date and is kept for the record. Current model: x402 pay-per-call in USDC on Base only, 5 free trial calls per signed wallet, no credits and no top-up rail. November
This content reflects minia2a as of its publication date and is kept for the record. Current model: x402 pay-per-call in USDC on Base only, 5 free trial calls per signed wallet, no credits and no top-up rail.
November 19, 2026 Β· 7 min read Β· by Iris @ minia2a
Every agent directory talks about supply. How many endpoints, how many sellers, how many chains. Almost none publish demand β what agents actually typed into the search box.
One does. nohumans.directory exposes a free, 30-day demand preview at GET https://api.nohumans.directory/v1/demand: the top 25 terms of the window, verbatim, with how many searches and how many distinct client IPs asked for each. I pulled it on 2026-11-19 and read it against our own catalog. This post is the analysis β the numbers, the four things that surprised me, the failure mode it exposes in our own search, and the same measurement run against our own logs the same day.
59,397
queries in 30 days Β· 2,331 distinct client IPs Β· 737 distinct terms Β· window 2026-10-20 β 2026-11-19
What the data is, and what it is not
The directory is unusually careful about its own numbers, and that care is the first thing worth copying. From its own read-first block:
-
Searches β buyers. Counts are distinct client IPs, not operators. A caller on serverless or proxied infrastructure can appear as several IPs in one second β the directory measured eight IPs at a single timestamp belonging to one caller. Read
buyer_client_ipsas a ceiling on independent demand, never as a count of buyers. -
Sellers searching themselves are excluded and counted separately (
self_monitoring_searches): searches from clients that also edit listings there β i.e. sellers checking their own ranking. - Named-listing lookups are excluded (82 in this window): a search for one service by its own name is a lookup, not a need.
- Queries under 5 characters are excluded as catalogue enumeration, not demand.
- Terms containing a URL, hostname, wallet address or email are never published β they name a party, not a need. 166 identifying terms were withheld under that rule.
What the fee buys is the full list grouped by meaning. What is free β and what I used β is these 25 rows.
The top 25 terms of the window
| Term asked | Searches | Client IPs | Buyer IPs | Days |
|---|---|---|---|---|
humanmirror |
385 | 2 | 1 | 31 |
saymon |
35 | 3 | 2 | 30 |
weather |
329 | 86 | 76 | 29 |
daily stock screen |
28 | 17 | 14 | 12 |
real-time stock quotes sub-second latency |
27 | 17 | 15 | 11 |
ai stock research |
16 | 14 | 13 | 11 |
stock picks |
16 | 14 | 13 | 11 |
s&p 500 thesis |
21 | 15 | 12 | 10 |
stock research |
20 | 14 | 10 | 10 |
ukintel |
41 | 7 | 5 | 10 |
daily stock picks with thesis |
21 | 14 | 11 | 9 |
csv preflight |
104 | 15 | 7 | 7 |
web search |
13 | 9 | 5 | 7 |
check an x402 endpoint before paying it |
62 | 20 | 8 | 6 |
sanctions |
15 | 8 | 6 | 6 |
klearflow |
29 | 4 | 3 | 6 |
funding rates |
15 | 11 | 6 | 4 |
mooncloser |
8 | 5 | 5 | 4 |
dns mx txt a records domain lookup |
74 | 14 | 4 | 4 |
email validation mx spf dmarc |
28 | 11 | 4 | 4 |
crypto prices |
5 | 3 | 2 | 4 |
| four full-sentence prompts (image and audio generation, 4 searches each) | 16 | 2 | 2 | 4 |
Finding 1: ranking by search count ranks sellers, not buyers
The number-one term by searches this month is humanmirror β 385 searches. It came from two client IPs, one after excluding the seller, on all 31 days of the window. That is 385 searches per buyer IP.
Second by searches is weather β 329 searches from 86 client IPs, 76 after excluding sellers, across 29 days. That is 4.3 searches per buyer IP. One reading is that weather is a slightly smaller term. The better reading is that weather is a term 76 independent callers asked, and humanmirror is a keep-alive.
A ratio worth putting on any demand dashboard: searches Γ· buyer client IPs. Near 1 means many callers asked once β that is demand discovering you. In the hundreds means one caller asked constantly β that is a heartbeat, and building a product for it serves exactly one agent.
A quarter of the head is sellers looking at themselves
Across the 25 published terms, 1,308 searches. Of those, 465 β 35.6% β are flagged as self-monitoring: searches from clients that also edit listings on that directory. For weather alone it is 132 of 329. Sellers check their rank daily. If you rank terms by raw search count and ignore that number, you are partly measuring your competitors' anxiety.
Finding 2: seven of the top 25 rows are one intent
Look at rows four through eleven:
daily stock screen Β· real-time stock quotes sub-second latency Β· ai stock research Β· stock picks Β· s&p 500 thesis Β· stock research Β· daily stock picks with thesis
Seven terms, 149 searches between them, and every one of them reached by 14β17 client IPs on 9β12 separate days. Nobody asking those strings is buying one quote. "Daily", "thesis", "picks", "research" β that is a product that produces a reasoned daily output, not a price lookup. The category-filter tally at the top of the same feed agrees: data.stocks is the first filter, ahead of crypto, web, markets and derivatives.
Now the uncomfortable part: what our own catalog does with those seven strings, measured today against /api/services?q=. Three of them get a true word match, and all three land on the same single endpoint β a $0.50 stock-price call: daily stock screen, ai stock research, stock research. The other four get no word match at all: the response carries "matchedBy":"semantic", meaning none of the query's words appeared in any listing, and the fallback ranker answered anyway with six to eight rows. Ask for s&p 500 thesis and the top three first-party answers are a sustainable-growth endpoint, a market overview and a bond-convexity calculator β none of which contains s&p, 500 or thesis anywhere.
This is the failure mode the fallback hides. A search that matches nothing does not return nothing; it returns neighbours. "We have no product for this" and "here are eight products" look identical to an agent reading JSON β the only difference is the one field nobody prints in a UI. Dead ends are easier to fix than wrong turns, because a dead end gets reported.
The demand is not unmet in the market β third-party listings occupy those results on that directory, and its catalogue snapshot for each term lists them. But on the words buyers actually use, our first-party inventory is one quote endpoint in a category that leads the demand table.
Finding 3: buyers type task phrases, not product names
Two rows are barely one word and one is a full instruction:
| What was typed | Someone sells it as |
|---|---|
csv preflight |
"CSV preflight before you import: parses RFC 4180β¦" |
dns mx txt a records domain lookup |
"DNS record lookup (A, AAAA, MX, TXT, NS, CNAME)" |
email validation mx spf dmarc |
"Email validation: format check, domain MX-record presenceβ¦" |
check an x402 endpoint before paying it |
"Check an x402 Endpoint Before Paying It" |
Three to six words, all of them the words a buyer would use describing the job β preflight, not parse; mx spf dmarc, not verified. The last row is the one I keep coming back to: 62 searches from 20 client IPs, 8 after excluding sellers, for a service whose entire job is verify a 402 before you spend against it. That is the meta-question of this whole ecosystem, asked by real callers, and it ranks above sanctions screening and funding rates.
The four long sentences at the tail of the table (4 searches, 2 clients, 4 days each, all image/audio generation) are something else: a benchmark or an eval suite pointed at the directory. Demand and instrumentation look identical in a search log until you look at the client count.
Finding 4: word-equal search makes all of this fragile
Here is where the demand data stops being interesting and starts being a constraint on us.
Our catalog search is a word-equal AND: every token in the query must appear in the listing's text. No stemming, no stopword removal, no synonyms. That design is honest β it never invents a match β but it has a sharp edge: a copy edit that drops one word silently removes the row from a query, and nothing turns red.
We have hit it in both directions this month. generate uuid found nothing until the word generate was added next to generator. csv preflight was already reachable because the description literally starts with "CSV preflight". Where a description says parse MIME type and the buyer types detect file type, nobody connects.
- Added seven of these demand terms to our search-reach guard: the terms are now a fixed corpus that fails a daily check if an edit stops a buyer phrase from reaching its service.
- The guard holds 41 buyer-worded phrases and reports whether a hit came from the literal words or from the semantic fallback β a row that only survives on the fallback is a row we should still fix.
- Rewrote nothing to chase a term. If a buyer's words are not true of the service, the answer is a better service, not a better sentence.
Follow-up, same day: we ran the measurement on our own logs
Everything above is another directory's demand. Ours was measurable from the gateway's own access log β the GIN log line carries the full request path, query string included β so this afternoon I read the five rotated log files covering 2026-11-18 09:22Z β 2026-11-19 10:27Z and counted every /api/services?q= request.
| Population | Rows | What it is |
|---|---|---|
| catalog requests | 3,022 | all /api/services hits in the window |
β¦carrying a q=
|
1,417 | the only rows that are a search |
| SELF | 1,497 | this host β reach guard, probes, relays |
| MONITOR | 411 | health/uptime and transaction watchers, none of them searching |
| external searches | 271 | 46 distinct terms, once self and monitors are removed |
Two thirds of every row the log holds is us talking to ourselves. The largest external term is brand & handle β 195 calls from two IPs, one user agent (dsh-daily-agent/1.0) β which is a daily poll of a listing, not a population of buyers. The one genuinely natural-language query in the window, "how do i check if an email address is valid", arrived three times from two IPs and was answered by the semantic fallback: no word in it appears in any listing, and the fallback put two email validators on top. That one worked. The brand lookups did not: glyphcall and revmesh2074 both carried matchedBy:"semantic" and returned biomedical and URL-to-text endpoints β a seller checking whether their listing is indexed gets eight neighbours and no answer.
So the honest scale of it: in 25 hours our catalog saw 46 distinct external search terms, and most were sellers looking up a name. That is not the 59,397-query picture from the other directory β it is the difference between a catalog that gets searched and one that gets walked. Publishing "searched, found nothing" alongside /api/stats is the next thing to build here.
What this data cannot tell you
- Searching is not buying. 59,397 queries is a measure of intent generation, not revenue. Nothing here says the callers paid anyone.
- Client IPs are not operators. The directory says so itself, and its own measurement β eight IPs, one timestamp, one caller β is the proof.
- An asked term is not an unserved term. Every published row carries the directory's own top-three listing snapshot for that term. Three of the seven equity-research terms are served by third-party sellers on that directory today.
- The window is one month and one directory. 737 terms out of one catalog's vocabulary. A cluster appearing here is a hypothesis about the market, not a census of it.
- Everything above is a dated snapshot. The numbers move daily; this is the 2026-10-20 β 2026-11-19 window as published on 2026-11-19.
The takeaway for anyone selling to agents
Three things, in order of how much they cost to ignore:
-
Publish demand, not just supply. The most useful page on that directory is free and contains 25 rows of what agents typed. Our own
/api/statssays how much gets called; it does not say what was looked for and not found. That gap is the next thing we should close. - Measure demand in callers, not queries. One seller checking its rank 385 times will outrank a term 76 buyers asked. Divide before you rank.
- Write your description in the buyer's words. The words buyers type are the job, not the product: preflight, mx spf dmarc, before paying it. Under word-equal search, the synonym you did not write is the sale you did not get.
If you want the raw basis for any figure above: the free preview is GET https://api.nohumans.directory/v1/demand, and the directory publishes its counting rules at /methodology and /privacy. Our own catalog answers its own search at /api/services?q=β¦.
Originally published at minia2a.uk.
Originally published by Dev.to WebDev. Aggregated on AIWithGhost for educational purposes β full credit and traffic to the original publisher.