GSC Isn't Broken — Your Customers Moved: Five New Ledgers for the AI-Era Marketing Report

Google Search Console (GSC) isn't broken. Your customers moved.

Customers used to find a clinic or a service by walking into the street market — the search results page. They would browse, compare stalls, then click through to yours. Now a growing share of them skip the market entirely and go straight to the chef: they open ChatGPT or a similar AI and ask, "Which one should I pick?" The chef sources the ingredients, cooks the dish, and serves it straight to the table — the customer never sets foot in the market. And GSC is the market authority's foot-traffic report. It cannot measure what happens in the kitchen.

Several industry reports point the same way: search visibility is falling sharply, and more than half of clicks can no longer be attributed to a source (we could not trace the original statistics, so we cite no precise figures; treat this as a "time to pay attention" warning — not a market law you build decisions on). But the direction they point matches what your own dashboard has probably been telling you: the battleground the old reports measure is shrinking.

So what should you measure instead? Here is a set of new ledgers written in language marketers actually speak.

01 Start With Five Questions

Break the farm-to-table journey apart and there are really only five questions:

1. Can the data be cited — is your content organized to the point where an AI can actually lift it, and is willing to?
2. Did AI retrieve it — did the chef's buyers (the crawlers) actually come and stock up?
3. Did the answer adopt it — when the AI answers a customer, are you in the dish it serves?
4. Did real humans arrive — did anyone actually walk into your website because of an AI answer?
5. Did they act after arriving — did the people who came in look up your details, click through to the official site, contact you?

One question, one ledger — five ledgers, each settled on its own. The full set currently tracks 23 numbers — 23＝13＋10: 13 already being measured, 10 still being built (metric dictionary). Whatever isn't measured yet gets an NA (no data yet) on the report; nothing is padded. These five questions need no new jargon — you can put them to any vendor and they hold up. The cells a vendor can't answer aren't a sign you don't get it; they're a sign the vendor isn't measuring it.

02 One Iron Rule: A Purchase Order Is Not a Sales Receipt

Across the five ledgers there is exactly one iron rule: no ledger may impersonate another.

An AI crawler fetching your data is the chef stocking up on ingredients. Stocking up doesn't mean the dish reached the table (your name cited in the answer); reaching the table doesn't mean the customer walked in (a real human on your site); and a customer walking in still doesn't mean a sale (a next action).

Marketers already understand this — only the setting has changed. Sending a press release to a hundred outlets doesn't mean a hundred outlets ran it; tripling website traffic still earns the same question from the boss: where are the orders? Reporting AI crawler volume as results is reporting "we sent it out" as "they published it" — it wins you one pitch and costs you the renewal. Settling the five ledgers separately means you never have to write that report: each number vouches only for its own stage, and whichever stage the boss drills into, you open that ledger and every line reconciles.

03 How to Replace the Five Old Metrics

Impressions: What AI hands the customer is an answer, not a results page — inside a closed AI answer, there is no impressions report you can pull. Watch two things instead: whether AI has been coming to pick up your latest data (the stocking ledger), and — using a fixed set of test questions asked for real — whether the answers actually include you (the on-the-table ledger). Report them separately. Never blend them.

Clicks: industry reports indicate more than half of clicks have already lost their trail (original statistics untraceable, so no precise figure is cited); more fundamentally, AI often resolves the customer's question right inside the answer, so the customer never needs to click at all. Watch three things instead: the "cooked-to-order" count — real people asking live inside the AI while it fetches from your site on the spot; AI-referred human arrivals; and what those visitors do after arriving.

Queries: AI does not hand back what users actually asked — the keyword report is not coming back. The substitute is building your own question bank: distill a fixed test set from roughly 28,000 real user questions (corpus collected since April 2026) (it counts questions asked, not unique askers), then regularly test which questions produce answers that use you.

Rankings: An AI answer is a paragraph, not ten blue links — there is no "position 3." The questions worth asking now: when the answer cites you, how early does it place you; and does the AI ever mistake you for a same-named competitor.

CTR: CTR is clicks divided by impressions. With impressions gone entirely and a large share of clicks missing, the division no longer computes. The honest replacement is a visible behavior chain — cited, arrived, acted — logged segment by segment, with an NA in any cell that can't be computed. No forced division.

04 The Numbers Already Being Measured

Below are measured numbers from one Taiwanese clinic and the IDAEO platform. Every number comes with two sentences: what it means, and what it must never be spun into.

• Ledger 2 (Did AI retrieve it) — AI visit score 82.0. This is a ranking score that blends the frequency and breadth of actual AI visits over the past 30 days. The raw value means nothing on its own; the rank is the point: #1/358 among the 358 peers indexed by the IDAEO ranking ("peers" = pages in the same IDAEO ranking category that already hold an AI visit score) (2026-08-05 snapshot, recomputable input pack — the score moves daily with its rolling window; formula and method), ahead of 99.7% of peers, and still climbing: +41.4 (16 days). This means AI is stocking up from this source more often and more broadly than from its peers; it does not mean the business is the best, and it must never be spun into revenue or customer volume.
• Ledger 2, the real-time version (the "cooked to order" count from the previous section) — 42,824 (two sites, 7 days; aggregation pack): over the past 7 days, real people asked live inside ChatGPT and the AI fetched from these two websites on the spot, 42,824 times in total (counting basis: requests carrying OpenAI's official ChatGPT-User agent, which marks fetches triggered by a user's live question; this is a request count). This means the data is being pulled into live Q&A; it cannot be converted into an equal number of visits or clicks, and it does not mean the final answer actually cited you.
• Ledger 1 (Can the data be cited) — 15/1,036/3,131/0: every claim on this clinic's website, checked sentence by sentence against the medical literature, lands in a four-cell ledger — 15 proven, 1,036 partially supported, 3,131 pending review, 0 contradicted. This means the evidence status of every sentence is on the books; "pending" means not yet reviewed, not wrong. Deleting the hard-to-prove sentences would also make the score look better — which is exactly why the four cells must be read together.
• Ledger 1, the platform's base layer — 13,593/1,928: across the whole platform, 13,593 claims have been checked against 1,928 publications (a different scope from the dental census's 18,731 claims / 745 papers: that counted claims extracted from clinic websites, this counts the platform's full verification pipeline; same-named metrics, different statistical windows — see the methodology page). This measures how thick the platform's evidence base is; it must never be passed off as any single clinic's result.
• Ledger 4 (Did real humans arrive) — which language of page AI referrals land on (counted as referrer-identified requests): English 74.7%, Japanese 22.0%, Chinese 3.3% — a 22× gap between English and Chinese (background in Lesson 1; the earlier piece counted five months of citation clicks, this one counts the current period's AI-referred arrival requests — different numerators and windows, so the figures cannot be compared). This means that for local Taiwanese content, AI-referred visitors land overwhelmingly on the foreign-language versions; it measures the language of the page the visitor lands on, not the language the asker speaks. And the 22× must not be read as a 22× gap in business opportunity. The control: in the same month, Chinese far outweighed English in Google Search clicks — ZH 31.6% vs EN 8.4% (Japanese led at 60.0%), Chinese beating English 3.8×. The two channels point in opposite directions, so the 22× gap is specific to the AI channel, not the site's underlying mix (baseline control).

05 The Part Still Being Built: The Mystery Diner

Ledger 3 — did the answer adopt it — is the heart of the whole system, and the part still under construction. The method works like sending in a mystery diner: fix a batch of test questions, fix a handful of AI models, ask the full round every week, and check whether the answers cite you, where they place you, whether what they say is accurate, and whether they mistake you for someone else. The question bank draws from the same roughly 28,000 real user questions above.

Why must the questions and the models stay fixed? Because if you change the ruler, you lose the comparison. Use one ruler this week and a different one next week, and nobody can say whether the swings belong to the ruler or to you; only the same ruler, applied every week, produces a trend worth trusting.

Until these numbers go live, the monthly report writes NA — every time — and never back-fills "AI came to crawl" where "the answer adopted it" should be. The first mystery-diner round ran on 2026-08-10: a fixed panel of 20 real questions across two model surfaces, and IDAEO embassy pages were cited 0/20 — an honest zero baseline that also captured who IS being cited today (directories and hospital sites). The curve starts now (full round-1 record). And the final referee is a two-field experiment: one field runs this playbook, the other is a holdout left alone; after enough time, compare the harvests. That gap, and nothing else, is the real proof.

06 The Monthly Report Keeps Only 7 Numbers

The monthly report to the boss condenses into four sentences and 7 numbers:

1. Can the data be cited — 3 numbers: evidence support, data completeness, official identity verification.
2. Did AI retrieve it — 1 number: the AI visit score.
3. Did the answer adopt it — 1 number: the mystery-diner citation rate (NA until it goes live).
4. Did they arrive, and did they act — 2 numbers: AI-referred arrivals, and the post-arrival action rate.

Do not roll them up into one total score. Composite "AI visibility scores" will flood the market soon, and the filter is simple: trust the reports that dare to write NA and report stage by stage; the ones that don't are mostly passing off "AI came to crawl" as "the answer adopted it" — or even "it closed a deal." A thinner report isn't laziness; it's escorting every number that can't explain itself out of the room. Every cell that remains can survive the boss asking, "Where did this number come from?"

Finally: Open Your Monthly Report and Audit It Cell by Cell

Now open the monthly report in front of you and put the five questions to it, cell by cell: Can the data be cited? Did AI retrieve it? Did the answer adopt it? Did real humans arrive? Did they act after arriving? The cells you can answer are your current position; the cells you can't are next quarter's work list.

And state this article's conclusion ceiling plainly: what can be said is that GSC cannot represent the AI battleground on its own; what cannot yet be said is that this playbook has been proven to drive business — because Ledger 3 (answer adoption) and Ledger 5 (post-arrival action) are still being built. A report that dares to print that sentence is a report worth trusting with your budget.

Method and Honest Limitations

Every number in this article was recorded and reported by our own pipeline; by construction it can undercount but not overcount. Measurement windows differ across numbers (7 days, 16 days, 30 days, and so on), so they must not be added together or converted into one another. The "answer adoption" and "post-arrival action" stages are still under construction; until they go live, the corresponding cells are reported as NA. The opening external warnings are industry-circulated claims whose original sources do not disclose windows or samples and which we could not trace, so this article cites no precise figures for them; they serve as a caution only. The definitions and downloadable assets for every number in this article are collected on the methodology page and in the evidence pack.