# AIARF GEO project workbook

Version: 2026-09-10. For a limited, single-product pilot. Use this workbook to connect requirements, facts, pages, sampling and acceptance. Each stage has a concrete deliverable reviewed under consistent definitions.

## 1. Project card

- Business / product:
- Intended and unsuitable audiences:
- Evidence of real questions: enquiries / interviews / other
- Intended customer action:
- Pilot scope, budget ceiling and final review date:
- Business-fact approver:
- Editing / website / monitoring / sales owners:
- Existing marketing activity:
- Deliverables directly under your control:
- External outcomes that can only be observed, not promised:

## 2. Fact card

| Field | Entry |
|---|---|
| Fact ID | F-001 |
| Approved statement | One complete, accurate fact |
| Conditions and limits | Region, version, audience, validity period |
| Internal evidence | Document, owner and verification date |
| Public evidence | URL supporting this exact claim |
| Status | Public / unconfirmed / permission pending / obsolete |
| Related pages | Pages and accounts requiring synchronized updates |
| Owner | Name and review triggers |

Supply only approved, public material to content production. Never invent an unverifiable number.

## 3. Question pool

| ID | Exact customer question | Real source | Stage | Includes brand? | Facts | Target page | Version |
|---|---|---|---|---|---|---|---|
| Q-001 | To complete | Anonymized interview or enquiry | Shortlist / compare / verify / risk / act | Yes / no | F-ID | URL | v1 |

Start with 20 questions as a planning suggestion, not a platform threshold. Prioritize observed frequency, business fit and decision impact. Ten thousand generated questions do not demonstrate customer demand.

## 4. Baseline and retest records

Retain these fields for each request:

```text
run_id / question_id / question_version
platform / consumer_app_or_api / model_if_visible
entry_url / web_search_setting / language / region_if_known
account_state / conversation_mode / timestamp
exact_prompt / original_answer / visible_citation_urls
request_status / failure_reason
brand_mentioned / explicitly_recommended / ordered_list_position
brand_fact_errors / source_supports_claim / reviewer
```

Three repeats per question can be an initial design. They reveal some variability but do not guarantee statistical conclusions. Keep planned, valid and failed counts separately for each platform.

### Counting rules

- Mention rate = valid answers containing the brand in their body / all valid answers. Record mentions only in citation titles separately.
- Recommendation rate = valid answers explicitly offering the brand as an option / all valid answers.
- Website citation rate = valid answers containing a clickable official-site source / all valid answers. A visible source is not necessarily the sole basis of the answer.
- Brand-audit factual error rate = valid answers containing at least one verified error / valid answers in that group.
- Failure rate = failed requests / planned requests. A network failure is not a normal brand absence, but must remain visible in the report.
- Normal answers without web search or without any brand recommendation still belong in the predefined valid sample. Do not exclude disappointing answers.
- Separate identity and non-branded buying questions, single turns and multi-turn chains, and consumer interfaces and APIs.

Example: 60 planned requests, 54 valid answers, 18 mentions and 10 website citations yield 18/54 ≈ 33.3% mentions, 10/54 ≈ 18.5% citations and 6/60 = 10% failures. These figures do not represent real user exposure. If failures concentrate in one question category, discuss the bias before comparing against a round with few failures.

Repeated samples share questions, time windows and platform conditions; they are not independent users. For small samples, raw counts and per-question changes are more useful than spurious decimal precision.

## 5. Content task

```text
Task ID:
One question to answer:
Reader and intended action:
Fact IDs and versions:
Structure: direct answer → conditions → steps / comparison → evidence → limits → next action
Missing material:
Fact reviewer / outcome / date:
Editorial reviewer / outcome / date:
Published URL / date / status:
Update triggers:
```

## 6. Five page roles

- Entity relationships: distinguish the brand, operating entity and product; do not present an individual practice as a company.
- Capabilities: explain currently deliverable features, prerequisites, limits and versions.
- Use cases: define audience, inputs, process, outputs and exclusions.
- Case evidence: match sources to claims, with clear permissions and limits.
- FAQ and updates: address comparison, verification, risk and action; retain revisions to obsolete statements.

Technical checks: public access, successful response, readable text, discoverable navigation, consistent canonicals and sitemap, and crawl/index settings matching publication intent. Schema and llms.txt do not universally guarantee recommendations.

## 7. Multi-turn testing

Natural discovery: begin with a real scenario without naming the brand.
Brand due diligence: name the brand and check follow-up information; do not merge these results into discovery data.

1. Which solutions fit this scenario?
2. Compare them using agreed dimensions.
3. Which evidence supports these capabilities?
4. When are they unsuitable, and how are risks handled?
5. What material is needed next, and where is the official entry point?

Preserve context within a chain and begin a new session for each repeat. Record any deliberate brand introduction. Prompted recommendations are not natural discovery. Record absent citations as no visible citation; do not infer hidden sources.

## 8. Distribution and maintenance

| Original page | Channel page | Purpose | Fact version | Publication date | Status | Owner |
|---|---|---|---|---|---|---|
| URL | URL | Full guide / demonstration / specific answer | Version | Actual date | Accessible / update needed / removed | Owner |

Reuse approved facts while addressing each channel audience. Ten reposts of one item remain one original piece of evidence. Do not fabricate independent endorsements, cases, reviews or instructions forcing model recommendations.

## 9. Weekly review

```text
Work actually completed:
Unfinished work and reasons:
Were baseline and retest conditions consistent?
Planned, valid and failed requests by platform:
Brand factual errors and repaired pages:
Buyer-question mention, recommendation and citation changes:
Two most important multi-turn failures:
Deduplicated enquiries / qualified leads / actual orders:
Differences between self-reported and observable sources:
Concurrent advertising, events, product and price changes:
Next round: continue / adjust / pause expansion
Specific tasks, owners, deadlines and acceptance outputs:
```

Direct visits, branded search and sales impressions alone do not establish GEO causation. With long sales cycles, report early indicators and observation limits rather than premature sales-growth claims.

## 10. Vendor acceptance

- Are exact prompts, environments, original answers and failed requests supplied?
- Are fact cards, sources and review records delivered?
- Does each fee correspond to explained work?
- Are case versions, scope and verification methods clear?
- Does the company control domains, accounts and exportable material?
- Are controlled processes separated from external platform outcomes?
- Who resolves unfinished items, by when, and when is the next stage decided?

## Related guides

- https://aiarf.cn/en/guides/brand-evidence-and-leads
- https://aiarf.cn/en/guides/geo-content-operations
- https://aiarf.cn/en/guides/multi-turn-decision-conversion
- https://aiarf.cn/en/guides/geo-investment-and-vendor-selection
- Shared methodology: https://aiarf.cn/en/methodology

Usage: examples and budget calculations illustrate methods, not real client results. We assess changes using verifiable facts and consistent sample records, without promising fixed rankings or time-to-results.