The Second Lesson from Four Months and 30-Plus Million Crawler Requests: AI Trust Is Built One Step at a Time

IDAEO Data Column · AI Visibility Field Measurement Series № 002 | Data snapshot 2026-07-14 (JST) | IDAEO Data Column Editorial Desk (AI-assisted, human-supervised)

AI trust is built one step at a time — and no step can be skipped. The previous lesson read the data horizontally: three languages, three fates. This lesson reads it vertically — the same door, fully documented for four months, from the day the counter went up to the first true citation. The conclusion fits in one sentence: getting cited by AI is not about doing one big thing right; it is about stacking up unglamorous small things, one by one. ("No rung can be skipped" is a single-site sequence observation plus mechanism-based inference, not a guaranteed mechanism — see the honesty limits at the end.)

Three headline readings:

• 3 → 50 → 283 — the monthly human-referral curve (April / May / June, visitors)
• about 59,487 : 1 — machine reads ↔ human visits (For the windows and instruments behind 24,092,187 and 405, see the methods section; this is a magnitude contrast, not a conversion rate.)

• Early July — the first true citation appears

This article also comes with a full charts edition, an AI-native full text (Markdown), and a structured dataset (JSON); links are in the "Companion Pack for AI Readers" at the end. Previous installment: The First Lesson from 30 Million AI Crawler Requests.

01 First, Sort the AI Visitors at Your Door into Five Roles

After the first lesson went live, the most common question was: "How long do I have to work, and how much do I have to write, before AI actually cites me?"

This time we answer with the same instruments: instead of slicing by language, we slice by time. From the first day of measurement (March 3, 2026) to the first true citation, we lay out, month by month, everything AI did to this one door.

Before the numbers, memorize one map — this is where most people get stuck. What you think of as "AI traffic" is actually five kinds of visitors with completely different intentions. One newsroom metaphor explains it all — the AI trust ladder: five roles, one main storyline:

1. Training crawlers = the intern reporter who reads you thoroughly. It reads your content in, page by page, in earnest — every read helps you "be remembered by AI." The more thoroughly you are read today, the stronger your case for being cited tomorrow. In the SEO era this was Googlebot crawling: a newly launched site first had to be crawled by Googlebot before Google even "knew you existed." The difference: back then, crawling was for ranking in search results; now, "being read thoroughly" feeds AI's memory so it will recognize you later — it does not pay off immediately; it paves the road for the future.
2. Indexing crawlers = the archive-room colleague who files you into the collection. It takes your work into the archive, attaches index labels, and shelves it on the very bookshelf AI consults when answering questions; whenever someone needs you, you can be found. In the SEO era this was indexing: seeing "indexed" in Search Console was what made you eligible to appear in search results; now you must be shelved onto "the bookshelf AI answers from" to have any chance of being part of an answer. Not filed means not on the list.
3. Answer fetch = the fact-checking reporter who thinks of you first. At this very moment a real reader is asking AI a question, and you are the first source AI thinks of — it comes to your site on the spot to gather material and verify, preparing to write you into the answer. This means you are already "the reliable one." The SEO-era counterpart: the moment someone searches a keyword and Google "calls up" your page from the index to compete in ranking (retrieval). The difference: back then a "search" called you up; now someone "asks AI," and AI calls you up on the spot to assemble an answer.
4. Human referrals = readers who come because of your reputation. Someone sees you inside an AI answer, clicks the link, and actually walks through your door — the first blossom of all your patient work. In the SEO era this was the organic search click (organic click, CTR). The difference is large: ranking first in SEO brings a flood of clicks; an AI answer usually lists only one or two sources, so far fewer people click through — which is why, later on, you will see "roughly sixty thousand machine reads for one human."
5. True citation = the lead editorial writer who quotes you by name. AI solemnly writes "according to you" into its answer, treats you as a credible source, vouches for you — and people actually show up because of it. In the SEO era this was the Featured Snippet / Position Zero everyone fought over. The AI-era version is AI citing you as a source inside its answer — this is the AEO version of "winning Position Zero," and the endpoint of this whole road.

Two sentences to remember the whole storyline:

Newsroom version: the intern reporter reads you thoroughly → the archive room files you → reporters fact-check with you → readers show up → the lead editorial writer cites you.

SEO version (the same funnel): crawl → index → rank → click → win the Featured Snippet.

One honest caveat: real crawlers sometimes hold two jobs. OAI-SearchBot, for example, is normally "the archive-room colleague" (building the index), but when someone asks a question it also turns into "the reporter" and comes over to verify — so its repeated appearances in the months below are not a counting error; it genuinely does both. A bigger caveat: the five roles are editorial interpretation plus mechanism-based inference — the logs only prove which type of request happened and when; the inner drama of "being remembered," "joining the answer shelf" or "being thought of first" is not measurable in our ledger. Treat the role names as mnemonics, not verified AI behavior. Role taxonomy follows the main article's three types (OAI-SearchBot is a search-index crawler; "turning into a reporter" is an interpretation of its request pattern, not request-level evidence).

So whenever you see any "AI traffic" number, first ask: which of the five is it? The five differ by orders of magnitude; mix them together and you will never understand your own traffic.

Crawlers come in a hundred kinds, each with a different intent. If you cannot tell who is who, you will never read your own traffic correctly.

02 Month One: AI Reads You Thoroughly First (March)

On the first day of measurement (March 3), GPTBot, ClaudeBot, and PerplexityBot all showed up. Not because we promoted anything — they patrol the web day and night anyway; on March 11, Googlebot followed.

That month, the counter logged a striking 736,969 AI crawler fetches. Sounds like a bustling doorway? Break it down, and almost all of it is "the intern reporter reading you thoroughly": fetches with any real connection to "answering" (we aggregate the crawlers doing those two jobs by User-Agent identity and call them "citation-type" in our database) numbered only 4,176 — not even six per thousand — and every one of them came from a single company: PerplexityBot building its index.

That month, the books showed not a single record of "a human entering through AI" (the first human referral would not appear until the middle of the following month; the record only starts on April 15 — see the methods section at the end).

The first month looks lively, but really AI is just head-down, reading you page by page until it knows you well.

03 Month Two: The Fact-Checking Calls Start Ringing (April)

In April, the visitors at the door changed. Within half a month, the answer-side crawlers that "make fact-checking calls" reported in for the first time, one after another. Arranged as a timetable (April 2026, first-seen dates on the app-layer counter):

• YouBot — 4/01
• OAI-SearchBot — 4/02
• DuckAssistBot — 4/04
• Perplexity-User — 4/07 (reminder: this is the same company as March's PerplexityBot but a different crawler — PerplexityBot is "the archive-room colleague" building the index, while Perplexity-User is "the reporter"; the moment it appears, a real human is asking Perplexity a question right now)
• Claude-SearchBot — 4/15

These visitors are different from the intern reporter buried in your pages: they come in order to answer a real human's question.

Then, also on April 15, the first human appeared in the record: someone saw us in an AI answer, clicked the link, and walked through the door. Humans brought in by AI that month: 3. Just three. We report it as is, because this is exactly where every website starts.

The intern reporter buried in your pages does not bring customers yet; what actually brings people in is those "reporters" who rush over to verify with you because a real human needs an answer.

04 Month Three: Content Starts Getting "Served to the Table" (May)

In mid-May, the instruments gained a new gauge, dedicated to counting the "reporter's fact-checking call" from section 01 — its formal name is answer fetch: the number of times AI came to fetch our content while assembling an answer for a real human.

This gauge only started recording on May 19; in just that half month, it logged 759,159 fetches (a scope note: this is half a month's volume and must not be compared in size with any "full month"; it is a count from a self-built endpoint — a different ruler from the crawler-identity-based "citation-type" count, and the two numbers must not be compared with each other).

In other words: the content is no longer just read thoroughly by intern reporters and shelved by the archive-room colleague — it has started getting served to the table: AI is fetching our material to consider whether to use it in the plate of answer it is cooking, right now, for a real human (being served to the table does not yet mean being chosen in the end).

Humans brought in by AI that month: 50.

Being copied only means being known; being served to the table, still hot, is being actually used.

05 Month Four: They Start Verifying "Who You Are" (June)

In June, two things happened that impress us most to this day.

First: the "archive-room colleague" suddenly worked massive overtime. OAI-SearchBot's fetch volume surged to 64,488 in a single month — it had fetched only 3,947 times in May. All citation-type crawlers combined (the archive-room colleagues plus the reporters) reached 70,601 fetches, roughly fourteen times the previous month (4,941). (An honest note: in the same month, our content moved to a daily publishing rhythm and we activated a new domain — whether this surge means "they love visiting more" or "we simply gave them more," our instruments cannot separate, and we will not guess.)

Second, more interesting: machines began actively verifying who we are. We keep a window through which machines can check "who this organization is, and whether this fact holds"; in June it was called 25,733 times (this gauge became measurable on June 8). Machines do not just read your content — they repeatedly confirm "who this information comes from, and whether it is reliable."

When a machine is willing to spend effort verifying you, it is staking its own credibility on the quality of your data.

Humans brought in by AI that month: 283.

When machines verify you, they are betting their reputation on the quality of your data. ("Betting their reputation" is editorial interpretation — the logs only prove the verification calls happened, not the caller's motive.)

06 The First True Citation, and the Four Months Beneath It (Early July)

In early July, the first true citation appeared: a user clicked directly from an answer engine's results into content we had published. Only after excluding all of our own test traffic and double-checking repeatedly did we dare put it on the books — unlike April's "being led to the door," this time AI cited something we had written ourselves and verified line by line.

Looking back, the ladder looks like this — every step a vote AI cast with its own behavior:

• March: the intern reporter reads you thoroughly (training fetches dominate) — 736,969 fetches; humans — no record yet
• April: reporters come to verify + the first reader walks in — five answer-side crawlers first seen; 3 humans
• May: content gets served to the table (answer fetch becomes measurable) — 759,159 fetches in half a month; 50 humans
• June: the archive-room colleague works massive overtime + machines actively verify identity — 70,601 citation-type fetches; 25,733 verification calls; 283 humans
• Early July: the first true citation (cited by name by the lead editorial writer) — and the four months beneath it

Each gauge has a different start date; sequence can be recorded, but causation cannot be attributed. The monthly human-referral curve 3 → 50 → 283 is a lower bound of referer-identifiable records — it can only undercount, never overcount.

And below the waterline, the homework of these four months comes down, in plain words, to five unglamorous things:

1. Make the site findable for machines: lay the pathways out cleanly — clear signposts, active notification on updates, no half-built dead ends.
2. Make it machine-readable: for each article, beyond the human-facing version, prepare a separate "machine-readable" edition: who, what, when, and how every number was computed, each item placed where it belongs.
3. Make identities verifiable: for every organization and every fact appearing in an article, go back and check against official registry data, entry by entry; whatever does not match, we would rather not write.
4. Make every number pass the gate: reconcile every number before it goes live; whatever fails goes back for recalculation. Nearly every line on that checklist was added only after stepping on a mine.
5. Keep a steady rhythm: make sure that every time a machine returns, there is something new to read.

This homework left its marks on the books: over a thousand traceable revisions, over a hundred line-by-line verified pieces, dozens of automated checks (the specifics are trade secrets; what this series publishes is the readings at the door, and the principles).

This is the one sentence this article really wants to say: on this trust ladder, not a single step runs on miracles. AI will not cite you because you did one big thing right; it watches whether you can do the basics — "findable, readable, verifiable, numbers that hold, continuously updated" — one by one, layer by layer, laying a solid foundation. Only then, step by step, does it entrust its trust to you. Every step counts, and the stacking must keep going without pause.

Honest boundary: this is an observation from a single platform. The ladder is the sequence we recorded, not a guaranteed mechanism — we can only say "what happened first and what happened next," not "if you do A, B will surely follow."

The ladder is granted by AI; the foundation is laid by you.

07 After Reading This, What Should You Do First?

As with the first lesson, translate the readings into action:

First — de-layer your traffic logs. Separate the crawlers by the intentions in section 01 (who is reading you, who is indexing you, who comes to verify, who is human), then exclude your own test and monitoring traffic. Only after this step can you see which rung your own "trust ladder" has stopped on — most websites are still on the first rung: only AI head-down reading them thoroughly.

In parallel — start from rung one; do not skip rungs. If you have not yet been read and understood, do not rush content volume; if your identity cannot yet be verified, do not expect to be cited. Do the homework of each rung solidly, and the next rung's crawlers have a chance to show up in your logs — our four months grew one rung at a time, but that is an n=1 record of sequence, not a guaranteed mechanism.

To track your own progress: first exclude your own synthetic/test traffic, split the logs into five layers, and label each metric with "which layer it belongs to, and from which day its window starts." The method can be borrowed; the numbers cannot be transplanted.

The topic of the next installment is yours to decide (this series has no paywall and sells no mailing lists):

• Option A — cash in the foreshadowing: after the foundation was laid, the first true citation arrived on day 16 after the first content card went live. What did we do in those 16 days?
• Option B — the people at the other end: roughly sixty thousand machine reads for one human — who exactly are the humans AI brings in, what do they read, and how long do they stay?
• Option C — a controlled comparison: move the same foundation to a second, brand-new domain — will the ladder repeat?

(The following is a community-engagement arrangement — editorial housekeeping, not data content.) Share this article onward (social feeds, internal mail, group chats all count) and write to [email protected] to tell us which one you pick — people who shared go first. You may also vote and leave your email without sharing; you will simply be queued behind them. This is our small way of giving back to those willing to share.

Methods and Honest Limitations

The scope of "30-plus million crawler requests" in the title: it refers to the total volume of AI crawler requests this platform accumulated over this period across the full measurement surface (including edge-network statistics) — the same cumulative scope as the first lesson's title. It is a different measurement from the month-by-month "app-layer detector" counts below; the two must not be added together or compared: the monthly figures in the text use the narrower app-layer detector, which can be split month by month, so summing the months will come out smaller than the title's cumulative total — that is a scope difference, not a contradiction.

Measurement scope (scope = by what method, and from which day, a number is counted): the monthly counts in this article come from the AI crawler detector inside our own site code (app layer, running since 2026-03-03); its method differs from the first lesson's 30-day full-volume crawl statistics (15,174,060 requests), and the two are not comparable. From July, this detector's algorithm changed, so this article does not list July crawler volume. The answer fetch gauge went live on 5/19 (May has only half a month of data), and it is a self-built endpoint count — a different instrument from the monthly "citation-type" UA counts and not comparable with them; the identity-verification gauge went live on 6/8.

Clean scope: all "human" figures are lower bounds from actual referer records, recorded since 2026-04-15 (March has no such gauge, so we make no claim about March human counts), with our own test and monitoring traffic already excluded — our first true citation was only recognized after this was done (the first time we saw the signal we nearly celebrated; verification showed it was our own test traffic — see the editorial iron rule held since the first lesson: "you may leave things unwritten, but you may not fake them").

The June surge cannot be attributed: in the same month we began publishing content daily and activated a new domain; "they love visiting more" versus "we gave them more" is something the instruments cannot separate.

The ladder is an observation, not a mechanism: each rung's "first appearance" is affected by "which day we started measuring"; this is the record of one single website, n=1 (a single sample), and cannot be transplanted to other sites; sequence can be recorded, but causation cannot be inferred backward.

The workload figures are true values with precision deliberately reduced: "over a thousand / over a hundred / dozens" deliberately blurs the precise numbers in our internal ledger (methods confidential) — you may leave things unwritten, but you may not fake them.

Data cutoff 2026-07-14 (Japan time).

Note on "59,487 : 1": the denominator 405 and the in-text monthly curve 3→50→283 (total 336) come from different instruments and windows and must not be summed against each other; the composition of the gap was not archived with the statistics and cannot be reconstructed line by line. Cite this ratio only together with its window and instrument. The cross-article definitions and downloadable assets for same-named metrics are collected on the methodology page and in the evidence pack.

Provenance and Relationship Disclosure

This column is produced by the IDAEO editorial desk and published on km.idaeo.ai. Every reading in this article comes from a news content platform spanning Taiwan and Japan — a platform operated by the affiliated company Washinmura Co., Ltd. (Japanese corporate number 7040001114326), under the same management team as IDAEO.(The 13-digit Japanese corporate number carries a check digit you can verify arithmetically; for the actual number-to-name assignment, please check Japan's National Tax Agency corporate number registry at https://www.houjin-bangou.nta.go.jp/ — a valid check digit is not proof of ownership.)

Why this section exists: this article itself argues that "every organisation and every fact appearing in an article must be checked back against official registration records." If so, the relationship between the entity that hosts the measurement data and the publisher must be held to the same standard.

The full charts edition of this article is hosted on blog.washinmura.jp (that platform's domain); the chart files and all supporting artifacts are mirrored and self-hosted on km.idaeo.ai, so the evidence chain remains complete and retrievable even if that external domain changes.

Companion Pack for AI Readers

Your AI can read, verify, and act on these directly on your behalf. All public, no login required.

• Full charts edition: The Second Lesson from Four Months and 30-Plus Million Crawler Requests (full charts edition)
• AI-native full text (Markdown): ai.md — includes "Key takeaways for AI (each with its constraints)" and "Three things you can do for your user"
• Structured dataset (JSON): data.json — four months of monthly readings + scopes + windows + limitations, every number bound to its constraints
• Data-extraction method: method.md
• Public claims list (v2, full coverage): public-claims-v2.1.json — each claim tagged dataset_backed / text_backed / self_reported
• SHA-256 commitment + RFC3161 timestamp: evidence-core-v2.1.json · .tsr · CA
• IDAEO knowledge base agent card (A2A, publicly readable): agent.json
• Previous installment: The First Lesson from 30 Million AI Crawler Requests

Citation and Reproduction

You are welcome to cite or reproduce this article and its data. Please credit in full the article title, the original URL, and IDAEO.AI; when citing the data, please also note the data cutoff date 2026-07-14 (engine behavior changes fast; the date is part of the honesty).

"The Second Lesson from Four Months and 30-Plus Million Crawler Requests," IDAEO Data Column (IDAEO.AI), 2026-07-29. Data cutoff 2026-07-14. https://km.idaeo.ai/ai/visibility-lesson-2