GEO Service Standards Arrive: What T/CAAAD 046–048—2026 Require

Short answer: on 14 October 2026 the China Advertising Association published three GEO group standards — T/CAAAD 046—2026 (service requirements), 047—2026 (effect monitoring and verification) and 048—2026 (content-source classification and selection) — effective 23 October 2026. They are voluntary industry standards, not law, and they bind nobody outside China. They are still the first attempt anywhere to write down what a GEO engagement must contain and how its results may be verified: three sampling windows, 30 valid observations per metric, a sealed intent-phrase set, a clean testing environment, independent verification with at least 20% resampling, and a hard separation between the party selling optimisation and the party certifying the outcome. For anyone buying GEO in any market, that list is now the cheapest audit checklist available.

What was published, exactly

The CAA proposed and centrally managed the three standards, and they took effect on the opening day of the 33rd China International Advertising Festival. Their stated purpose is narrow and unglamorous: the GEO market is growing faster than its behaviour is being governed, and the association wanted a baseline that separates compliant service from the practices that had already produced complaints. Each standard addresses one link in the chain, and the three are designed to support each other rather than stand alone.

Source: China Advertising Association, announcement of the three GEO group standards, 14 October 2026 (published in Chinese; standard numbers per the association's own release and the participant announcements). Full texts are filed with the National Group Standard Information Platform.

The measurement standard is the real news

A service code tells you who to hire. A verification code tells you whether you were told the truth, which is the scarcer commodity. The GEO industry's trust deficit is not a rumour; it is structural. The party that performs the optimisation usually also produces the report that grades it, the prompts are chosen after the results are known, sample sizes are omitted, unfavourable runs are dropped, and screenshots arrive without a model version or a timestamp. Any one of those habits is enough to make a 40% improvement appear and disappear on demand.

T/CAAAD 047—2026 attacks exactly that. It does not set a target for citation rate, because a regulator that invents a target invites gaming; it sets the conditions under which a number may be reported at all. Six of those conditions are portable to any market:

1. Three time windows, not one. Effect verification should use repeat sampling across no fewer than three separate windows. A single measurement is explicitly not acceptable as proof.
2. Thirty valid observations per metric. Any single key indicator needs at least 30 valid observations before it can be reported — the difference between a reading and an anecdote.
3. A sealed intent-phrase set. The prompts are agreed and frozen before measurement starts and may not be added to, removed or reworded mid-programme. This kills the oldest trick in the category: quietly swapping a hard question for an easy one.
4. A clean environment. No login state, no conversation history, no personalisation, no local cache, no device fingerprint, no geo drift. Two people asking the same question in the same city can otherwise see different source sets, which is a property of the engines, not of your optimization.
5. A rebuilt baseline on model change. When a model or product version changes materially, the baseline is rebuilt and pre-change data is not compared directly with post-change data. Every practitioner who has watched a citation count jump on a silent model update knows why this clause exists.
6. Independent verification with ≥20% resampling. The verifying body should be independent of the party delivering the work, must resample at least 20% of the sealed phrase set, and must fail the verification outright if the deviation traces to data fabrication.

The measurement standard also insists that reports carry unfavourable results and volatility, not just the wins, and that validators avoid words like "guaranteed", "controlled", "permanent" or "exclusive". Those two clauses are worth more than any KPI in the document, because they attack the incentive that produces bad GEO reporting in the first place.

Whom the service standard excludes, and why the list is short

T/CAAAD 046—2026 sets four principles — truthfulness and accuracy, user value and fair competition, transparency with reviewability and cautious claims, and disclosure of conflicts of interest — and then converts them into a negative list. A provider cannot claim conformity if it has no real operating subject, has been penalised in the previous twelve months for false advertising or data fabrication, or has been found to have run corpus poisoning, corpus bombing, forged trustworthy sources, forged monitoring results, prompt injection or hidden text.

That list is short because the associated harms are already documented rather than hypothetical. A Fudan University study of 17,211 citations across 10 AI-search platforms found that 8 of 10 platforms cited a deliberately fabricated concept within seven days, and a $14 purchase produced 13 posts that one platform cited within an hour. A standard that names corpus poisoning is responding to a measured event, not a theoretical risk.

The service standard also adds vertical-specific obligations: medical, pharmaceutical, financial, franchise and education clients face stricter qualification checks and dedicated review, and technical adaptation must aim at accessibility and machine-readability rather than at deceiving crawlers or bypassing platform safety mechanisms. Note what that implies for the tactic list most GEO guides still lead with — hidden text, doorway pages and bulk synthetic placements are now formally outside the definition of the service, in the largest AI-search market by user count, whatever they do for a short-term citation count.

Content sources get grades — and the grades expire

T/CAAAD 048—2026 is the least glamorous and most consequential of the three, because it governs supply. It scores a content-source subject on six dimensions — subject qualification, record of content truthfulness, traceability and machine parseability, behavioural compliance, stability and continuity, and independent external verification — and assigns one of four grades:

Grading is not permanent. The standard calls for a full review every six months, immediate review on major events, downgrade-and-exit paths for serious violations, confirmed black-hat GEO or content-driven reputational incidents, and a stability marker derived from the standard deviation of scores across consecutive assessment cycles — "stable", "volatile" or "oscillating". For a brand that buys third-party placements to earn citations, the practical shift is that the domain's identity no longer carries its trustworthiness; its qualification record, traceability and external corroboration do.

This is the clause most likely to be copied outside China, because it solves a real procurement problem. Buyers currently have no way to answer "is the site you placed me on a legitimate publication or a synthetic container?" The six-dimension model is a defensible answer, and it is checkable after the fact.

What the standards do not do

Three limits deserve to be stated plainly, because any coverage that treats this as a solved problem is overselling it.

• ▸ They are voluntary and national. Group standards are industry self-regulation. No regulator outside China is bound by them, and inside China compliance is enforced through procurement, reputation and the association's assessment programme rather than penalties.
• ▸ They standardise procedure, not engine behaviour. Following every clause still yields a number that may not reproduce next week, because engines reroll their retrieval sets constantly. A compliant measurement is a defensible measurement, not a stable one.
• ▸ No public audit registry yet. The association has signalled compliance assessment and an AI-marketing review platform, but until grades and verification reports are publicly queryable, buyers remain dependent on disclosure from the party they are paying.

Where the standards stop, and the four gates we run instead

A verification standard tells you how to measure a result. It does not tell you why a page failed, which is the question a content team actually needs answered. Our own audit reduces that to four sequential gates, each of which can be failed independently of the others:

The order matters and cannot be shuffled. A page that fails gate 1 has no measurable performance in gate 3, and a benchmark-grade statistic inside an unfetchable page is invisible. That is also why the standards and the gates are complementary: the CAA documents tell you whether a reported result was measured honestly; the gates tell you where in the pipeline it was lost.

"The release of the standards is not the endpoint, but the starting point of industry self-regulation." — Huo Yan, secretary-general of the China Advertising Association, in the association's release of 14 October 2026 (translated from Chinese).

Five clauses to put into your next GEO contract

Whether or not your vendor operates in China, the verification standard gives you language you did not have before. These five clauses capture most of its value:

1. 1.
   Sealed prompt list, agreed before work starts. No additions, deletions or rewordings mid-programme. Any change requires a written amendment and a new baseline.
2. 2.
   ≥30 valid observations per reported metric, across ≥3 time windows. Anything below that threshold is labelled a preliminary reading, not a result.
3. 3.
   Environment and version disclosure on every observation. Model name, version, date, entry point and whether personalisation was disabled — the four fields that make a screenshot auditable.
4. 4.
   Independent verification, or at minimum delivery/reporting separation. Resampling of ≥20% of the prompt list by a party that does not invoice for the optimisation.
5. 5.
   Unfavourable results are contractually reportable. Volatility, regressions and lost citations appear in the same report as the gains, and "guaranteed top placement" language is excluded from the agreement entirely.

None of these clauses requires a regulator. Each one removes a way for a bad result to be disguised as a good one — which, for a discipline whose main output is a claim about someone else's algorithm, is close to the whole job.

Frequently asked questions

Related reading

• ▸ GEO Evidence Audit 2026
• ▸ Black-Hat GEO: Corpus Poisoning — the practice the 046 negative list names
• ▸ How to Measure GEO Visibility
• ▸ The 27-Point GEO Audit