An overseas marketing team working with a Chinese manufacturer may commission product demonstrations, publish technical explanations and then receive an impressive “AI adoption” score. Before using that score to renew the content programme, ask what happened: was a page fetched, loaded into a generation context, cited, presented or acted on?
Buy the reporting capability that actually exists. A crawler report can be useful without proving that a product explanation shaped an answer. Combining those claims makes a content budget harder to defend, not easier.
A reporting vocabulary is not universal platform coverage
On October 22, 2026, the Prebid LLM Taskforce published “Provenance and Reporting: Who Made Your Content, and How Did AI Use It?”. It separates content provenance from reporting subsequent use. Its authors include a solution provider and a specification maintainer; read it as a mechanism proposal, not an independent adoption survey.
The SPUR Content Telemetry Specification, currently labelled version 1.0 and published September 2, 2026, supplies a vocabulary for usage events. Its existence does not establish that a particular AI operator emits those events or that your reporting vendor receives them.
That distinction changes the procurement question. Alongside “Which platforms do you monitor?”, ask “Which events do you actually receive from each platform, and which findings come from your own observations?”
Do not turn different events into one adoption score
In brief, the specification distinguishes retrieval, entry into generation context, source association, recipient-facing presentation and observable interaction. Grounding does not quantify a source’s contribution; presentation does not establish attention. These are not a compulsory funnel: a citation can lack matching grounding, and content can be presented without grounding. The specification does not report that a model trained on a work.
The observer matters. Origin-side traffic records cannot see inside generation. Finding a link in a public answer is a useful external observation, but it is not an operator’s internal usage report. Keep those evidence types separately labelled. Sections 4.3 and 4.4 of the specification explain the distinction.
ONCE’s five-column usage receipt
We propose a practical reporting acceptance sheet—not an industry standard or a claim that ONCE has deployed a telemetry product. Every record used to justify a content decision should name:
| Column | Question the buyer should be able to answer |
|---|---|
| Content version | Which approved page, transcript or revision is this about? |
| Event | What exactly was observed or reported? |
| Observer | Who could see that event, and where did the record originate? |
| Coverage | What period and population does this report actually cover? |
| Action | What can this finding reasonably change? |
Consider a hypothetical appliance brand whose China-based product team has revised an installation guide. The overseas team receives a claim that “AI use increased,” supported only by server logs. Its receipt should identify the guide’s URL and revision, label the observation as retrieval, name the server or CDN as observer, describe the recorded scope and propose checking access to that version. It should not recommend cutting other creative investment on that basis.
A separate answer sample might show a citation. Preserve the question, platform, date, complete answer and actual destination. Does it reference the current guide, an older document or a distributor’s page? Record “citation observed in this sample,” not “all users received our explanation.” Do not silently relabel a public-output check as operator telemetry.
Where an operator report is genuinely available, examine its definitions and declared coverage before comparing periods. Unknown coverage is not a minor footnote. An absent report is neither zero use nor evidence of invisible success.
Commission a sample report before a retainer
Choose one approved public content item and fix its version. Ask the prospective supplier for a small, inspectable report. They do not need to manufacture every event type; they need to substantiate every capability they sell.
Compare available source-side access records with any claimed operator reports. Keep unlinked records separate rather than inventing a continuous journey. Then inspect public-answer samples and their cited destinations. Do not collect private conversations or confidential customer prompts merely to populate this exercise.
Ask the supplier to explain how duplicate observations are handled, whether a coverage change invalidates a month-to-month comparison, and how unsupported platforms appear in the overall score. A defensible limitation belongs beside the number, not behind it.
For production decisions, add a causal check. A change in observed citations does not, by itself, prove that a new edit or reshoot caused it. Questions, platform behaviour and other sources may also have changed. Production Soup’s October 1, 2026 guide to checking video content for AI search likewise separates publishing readiness, observed citations and attributable visits. That is a practitioner’s method, not a performance guarantee.
What the production partner should deliver next
For brands coordinating an overseas campaign with a China-based production team, add this to the brief: “Connect the approved film, transcript and supporting-page revisions to the observations the report can actually support.” A useful partner can explain which asset needs maintenance and which decision still lacks information, without promising control over AI answers.
ONCE’s commercial interpretation is that repeatable reporting services should accumulate comparable asset histories and observation records, not inflated activity totals. That is a proposed operating model, not a valuation forecast or a measured client result.
Before approving the next retainer, request one usage receipt with all five columns and its limitations intact. If it cannot support a specific next action, a larger dashboard will not solve the purchasing problem.
