peterbamuhigire AI Generative Search Optimisation (GEO)
Acknowledgement: Shared by Peter Bamuhigire, techguypeter.com, +256 784 464178.
Use this skill to plan evidence-bounded discoverability across AI answers, search, social profiles, communities, and the destination a customer uses next. It is a planning and audit route, not a promise of inclusion, ranking, or sales.
Use When
- The deliverable is an AI-search visibility plan, audit, content system, or measurement loop for a social or digital-marketing engagement.
- The work must connect social profiles, native content, off-site reputation, website destinations, and customer action.
Do Not Use When
- Use
ai-readiness-diagnosticfor a general AI maturity, data, team, or deployment assessment. - Use
seo-discovery/seo-geo-optimisationfor one page or article only. - Do not publish, send, spend, alter a live account, collect personal data, or claim a certification without explicit authority and the relevant release gate.
Required Inputs
| Artefact | Source/provider | Required? | If absent |
|---|---|---|---|
| Business name, offer, audience, market, goal, channels, and intended decision | Approved brief and client fact sheet | yes | Stop the affected recommendation; state a narrow assumption only where safe |
| Existing profiles, content, destinations, analytics, referral data, and customer questions | Supplied exports, URLs, CRM or platform evidence | conditional | Mark the check not assessed; do not infer visibility or performance |
| Current platform, market, legal, privacy, rights, and AI-search claims | Social source register and Digital Research verification | yes for material claims | Quarantine the claim and narrow the deliverable |
| Approval, access, budget, language, accessibility, and moderation constraints | Accountable owner | conditional | Stop publication, spend, collection, or live changes |
Capability and Permission Boundaries
Read and search are the minimum capabilities. Planning and audit are read-only. Edits to repository guidance are in scope for maintainers; live publishing, outreach, spend, personal-data processing, production changes, and certification claims require separate explicit authority.
Degraded Mode
If evidence, network, platform access, native-language review, rights review,
or measurement data is unavailable, return the narrowest useful plan and label
each affected item not assessed. Never convert a missing check into a pass.
Decision Rules
| Condition | Action | Failure or risk avoided |
|---|---|---|
| The claim is current, material, and supported by the source register | Cite the source at the point of use and record scope, dates, freshness, and limit | Stale platform or market advice |
| The observation is a mention, citation, referral, sentiment, or conversion | Name that exact outcome; keep it separate from the others | False “AI rank” or attribution certainty |
| The profile or post contains a factual, regulated, sensitive, or rights-bearing claim | Require owner evidence and the relevant legal/rights/market gate | Harm, rights breach, or fabricated proof |
| A destination is useful to people and agents | Improve clear facts, accessible text, consent-safe CTA, and failure path | Optimising a surface that cannot complete the job |
| Evidence is partial or contradictory | Narrow, quarantine, or mark NOT_ASSESSED; preserve the contradiction |
Confident synthesis from a weak source |
Workflow
- Frame one audience, channel, customer job, business outcome, market, and approval boundary. Record the consequence of getting it wrong.
- Establish the baseline: customer questions, profile/entity consistency, content and source quality, canonical destinations, available referrals, self-report, prompt observations, platform data, and (where authorised) logs.
- Apply the three-mode planning lens from the Carter synthesis: evergreen brand/offer facts for remembered knowledge; current sourced updates for retrieval; deep evidence, trade-offs, and working for reasoning. This is a durable planning lens, not a fixed platform taxonomy.
- Select one content or profile slice. State its hypothesis, primary outcome, trust/cultural/accessibility guardrail, owner, time-box, stop rule, and rollback path.
- Make the slice legible: who the brand is, what it does, for whom, where, under what limits, with a clear next action and an accurate canonical link. Use native channel conventions without forcing slang, hashtags, or claims.
- Verify every current claim, source, statistic, quote, rights assertion, and platform rule. Run anti-slop, creative, legal/market, language, and accessibility reviews that apply to the asset.
- Measure separately: representation, retrieval/citation observation, referral, qualified action, and revenue. Record sample, date, denominator, consent, and attribution limits.
- Check normal and failure paths, including inaccurate AI descriptions, negative or misleading UGC, broken destinations, opt-out, moderation, and no-data states. Correct, quarantine, or rerun the affected check.
- Standardise only a demonstrated improvement in the skill, reference, template, source register, fixture, or gate. Record the next re-audit.
Use the Garner, Woolley, and Bishop/Starkey independent synthesis to add three checks to the slice: the outside-in customer journey, an intent/customer-language map, and a recognisable human voice. Retain native adaptation, source/rights/approval handoffs, moderation, and a correction path; the supplied books are historical or editorial inputs, not current platform authority.
Outputs
| Artefact | Consumer | Observable acceptance condition |
|---|---|---|
| AI-search visibility audit or plan | Strategist, client reviewer, or delivery team | Audience, channel job, evidence boundary, outcome definitions, sequence, owners, and gaps are explicit |
| Content/profile action brief | Content or community operator | One real slice has channel-native copy guidance, source/rights status, CTA, moderation path, and acceptance checks |
| Measurement and learning record | Analyst and accountable owner | Prompt observations, platform data, referrals, self-report, and qualified outcomes are not conflated |
| Decision and gap note | Approver or next workflow | Unsupported, unauthorised, stale, and NOT_ASSESSED items are visible with a recovery action |
Evidence Produced
| Evidence | Format | Acceptance condition |
|---|---|---|
| Source and claim register | Inline table or linked JSON/Markdown record | Every material current claim has a verified source, scope, date, freshness, support state, uncertainty, and owner |
| Content/profile fact map | Table | Identity, offer, audience, location, proof, limits, rights, and canonical destination are traceable |
| Experiment record | Markdown or tracker row | Hypothesis, baseline, measure, guardrail, stop rule, result, rollback, and standardisation decision exist |
| Release review | Completed gates | Anti-slop, rights, legal/market, cultural, language, accessibility, and approval status are explicit |
Quality Standards
- Use British English and Uganda/East Africa defaults only where they apply; record any different market, language, currency, timezone, or channel reality.
- Keep the human job primary. Clear answers, evidence, honest limits, and a usable destination matter more than AI-facing formatting.
- Make social and community presence part of the discoverability surface without buying, seeding, manufacturing, or suppressing mentions or reviews.
- Use
llms.txt, markdown mirrors, APIs, MCP, or agent integrations only as a named, reversible experiment for a real consumer or task; never as a default ranking lever. - Run
ai-marketing/anti-ai-slopduring drafting andai-marketing/ai-slop-auditafter major iterations; an F blocks progression until fixed.
Anti-Patterns
- Unsupported benchmark or adoption number. Fix: verify the primary source or remove it.
- “AI rank” reported as a metric. Fix: name mention, citation, referral, action, or revenue.
- FAQ, 50-word opening, monthly cadence, or speed target treated as universal. Fix: make it a tested local acceptance choice or remove it.
- Inauthentic mentions, seeded comments, or manufactured reviews. Fix: use authentic, rights-cleared evidence and moderation.
- A profile optimised without an accurate destination. Fix: trace the click, consent, form/WhatsApp path, and failure recovery.
- A calendar presented as learning. Fix: add a hypothesis, guardrail, stop rule, and result.
- A current platform claim copied from a book or AI answer. Fix: route it through Digital Research and mark it
NOT_ASSESSEDuntil verified.