How should each of the five layers be defined?

I write the protocol as field definitions rather than drawing a single funnel. Mentions, shortlist appearances, and citations are properties of an answer; clicks and inquiries belong to later visits and business records. Different properties within an answer may overlap, and they are not guaranteed to occur in any particular order.

This is the operating protocol I recommend teams use, not an official standard shared by all platforms. You can adjust the fields, but write the definitions down in advance; don’t wait until the monthly report to choose the calculation that favors you.

For an owner, the most useful thing is not a composite score but seeing which layer has evidence and which has not been measured yet. You can pair this with how monitoring duties are divided across GEO mentions, citations, and visits to confirm what kind of material a tool can actually read.

How do you count identity inside an answer accurately?

Mentions need a brand-alias table. Our strict definition recognizes 天问, 天問, and tianwen; the broad definition includes either a strict brand hit or a citation of our own domain. Strict hits also count toward the broad definition; a URL appearing on its own does not automatically count as a strict hit, and the two are not added together. This way you can both see whether the name appeared and avoid missing the case where a URL appears only in a citation list.

Shortlist appearances need context. When the company appears alongside other providers on a list, it can be marked as a list appearance; but if the same passage is warning buyers to avoid a certain kind of company, you also need to keep the direction of the assessment. The appearance of a name by itself cannot show that the answer is recommending it.

Citations must be verified against the actual URL. Whichever site the answer’s link points to is the site you record; similar brand text inside a URL path does not count as your own domain. The false positives in T1 came precisely from this confusion, and only after the audit correction did we have a final count.

I also keep the specific content that a citation supports. When an answer cites a selection guide, it shows that the model treated the page as a source; you cannot jump from that step to “it endorses the author’s company’s services.” For checking source position, see how to track AI citation URLs against the original text.

Several tags can belong to the same answer. If the brand appears on a list and a guide link is attached at the same time, you mark shortlist and citation separately within the same record. When aggregating, you cannot add the two tags together and call them two independent answers, let alone two buyers.

Why can’t the numbers from a real panel be pooled?

Start with the dates and the sampling unit. In our fixed-prompt panel, T0 is 2026-09-17 and T1 is 2026-10-12; each round has 24 prompts, each run in 3 independent sessions, for 72 answers in total. These are research probe outputs, not real customer counts.

The shortlist layer also needs a stated prompt scope. In T1, the T6 provider-naming prompt produced 12 answers, in which 2 provider identities were observed. If that sub-scope is what you are reporting, write the sub-scope denominator clearly; don’t switch back and forth between denominators from different prompt types.

These items are not a chain from which a conversion rate can be computed. Strict mentions are not the full parent set of shortlist appearances, and a citation does not necessarily pass through a brand mention in the body text; site clicks are even less a buyer behavior naturally measured in this batch of probe answers. So you cannot divide list appearances by mentions and name it a “recommendation conversion rate.”

Changes between rounds should also keep the opposite direction. Strict mentions rose while the broad definition fell; the total number of cited domains went from 322 to 245, with only 90 overlapping. These results contain sampling noise, so you cannot take only the upward items and attribute them to your work. You can check your reporting definitions with a vanity-metric check for GEO monthly reports.

What evidence is missing between a citation and a click?

The existence of a link does not mean a user clicked it. A probe can save the citation URL in an answer, but judging a real visit requires looking separately at the actual records on the site side. Opening a link yourself to double-check should not automatically count as a buyer visit either.

Visit records also need source-identification rules. If the site records can identify a visit coming from a particular AI product, save it under the agreed fields; when the source is unknown, you cannot assign the visit just because the landing page happened to be cited by an AI.

I list “identifiable source,” “customer’s own account,” and “unknown source” separately. A customer saying they saw the company in an AI is a useful clue, but it differs in evidential strength from a link supported by a visit record. The monthly report should preserve that distinction.

Technical setup must also be distinguished from results. Installing an analytics tag only establishes the conditions for collection; you then have to check whether visits, sources, and landing pages are actually being recorded. For how the site’s data tools divide responsibilities, see how GSC and GA4 divide the data work; don’t let a configuration screenshot stand in for actual data.

If collection has not been set up yet, write “not measured”; if collection exists but the source cannot be determined, write “source unknown.” Only when the scope, time window, and collection validity are all clear, and there genuinely are no records matching the rules, may you write observed zero within that scope. These states are not interchangeable.

From click to inquiry, what records should be kept?

I first check whether there is real contact content. Clicking an inquiry button, opening WhatsApp, or reaching a thank-you page is a technical or behavioral record; whether the buyer submitted valid information and left a contactable purchasing request still requires checking the contact record.

The contact record and delivery should also be separated. A form saved on the site does not automatically prove the recipient received the email; an email being received does not mean sales has followed up. When writing a report, state how far you actually verified; leave later states blank if there is no evidence.

Deduplication rules must be set in advance. Whether the same contact repeatedly sending the same request through different channels counts as one inquiry should be judged from the actual content, and the rule should be recorded. Test submissions, spam, and clearly irrelevant contacts are each labeled separately and must not be mixed into business inquiries.

Inquiry quality also needs follow-up feedback. Customer identity, purchasing relevance, contactability, and the sales outcome can be appended as statuses on the same record. A real inquiry, a qualified lead, and a closed deal each have their own supporting material; you cannot upgrade the labels all the way down the line just to make the presentation look good.

If visits rise but purchasing contacts do not improve, I go back to source and demand. You can check for a mismatch using how to diagnose rising traffic with falling inquiries; you cannot automatically explain it as GEO acquisition working just because the growth happened on a page that was cited.

How does the team execute the same measurement protocol?

I first freeze a definitions table. Brand aliases, owned domains, shortlist identity, visit-source rules, inquiry deduplication methods, and the valid time range are all written down before collection begins. If a rule later turns out to be wrong, keep a revision record and recompute the comparable older data under the new rule.

Answer records and business records each use traceable identifiers. Prompt number, language, sample ID, and round locate an answer; visit and contact records use their own identifiers. Connect them only when linkage evidence exists; don’t force a match just to draw a complete funnel.

For every report, I require an evidence location next to each result. A list identity in an answer comes with the original sentence, a citation with the full URL, a visit with the source determination, and an inquiry with the real record. Only when aggregate numbers can be traced back to the source documents can different colleagues check them under the same rules.

Rule checks also need counterexamples. A similar brand on a different domain, a negative assessment inside a list, a button click with no submission, an unknown-source contact — all of these can help expose the boundaries. Examples are for validating rules and should not be disguised as new customer cases or statistical samples.

The report ends by writing observation, work delivered, and business linkage separately. A change in the fixed-prompt results can support an observation; an actual page change can support a delivery; without click and inquiry evidence, you cannot claim business linkage. For a concrete report structure, see what a GEO monthly report should include and how to read a GEO diagnosis report.

Our research materials still explicitly retain the item “citation-to-inquiry lacks a metric.” Writing the gap down does not detract from the existing observations; it shows exactly what evidence the next round still needs to add.

What other common questions come up about counting and attribution?

If one answer both mentions and cites the brand, does that count as two successes?

Just tag the mention and citation fields separately. They describe different properties of the same answer; they cannot be added together to look like more independent answers, nor counted as more buyers.

Does the AI listing the company name prove it endorses the company’s capability?

No. Appearing on a shortlist is only an identity observation in the answer; you still need to keep the context and the direction of the assessment. A model naming the company does not mean its capability, delivery quality, or customer satisfaction has been verified.

If the report has no AI-source click data, should it be filled in as zero?

When there is no collection or the source cannot be identified, mark it as not measured or source unknown. Only when the scope, definition, and collection validity are clear, and no records match the criteria, may you report zero within that scope.

How can an inquiry be attributed to results brought by an AI citation?

You need source evidence that can link an actual visit to a real contact record, and you must explain the linkage rules and the parts that cannot be identified. If you have only a customer’s own account or an isolated screenshot, list it separately by evidential strength; you cannot write it as a fully verified causal chain.