GEO Operating System, Research → Sense → Decide → Act → Verify

Off-page earned-media & PR execution (loop step 4b, Act / Off-page; owner: the brand's GEO lead). ~85% of AI citations come from third-party pages, so this is the highest-leverage lever. Work the gap-ranked off-page targets from GEO Decide across the source categories the models actually cite:

- Editorial / digital PR: pitch data, contrarian takes and named frameworks to publications the models already cite; earn mentions in 'best [category]' and '[competitor] alternatives' listicles (Barnacle SEO). Keep a brand kit ready. Human approval the owner before any outreach.

- Reddit & community (via Reddit): genuine participation in threads that already rank for target queries, answer real questions, never spam (Reddit punishes self-promotion; manufactured seeding backfires).

- Reference / directories: complete and correct Crunchbase, G2, Wikidata and industry directories, fastest wins, can move within 24h.

- Source correction (sentiment): for weak themes, politely ask the most-cited inaccurate/outdated sources to update (templated outreach) and solicit reviews on the platforms the models cite.

- Guardrails: white-hat / low-risk grey-hat only, no fake reviews, sock-puppets, hidden text or manufactured consensus.

Feeds GEO Verify.

Targeting & tracking standard.

- Rank off-page targets by how many tracked answers each source feeds, not by domain authority or traffic.

- Match the ask to the source type: directory/listing → submit or claim the profile; comparison or review page → inclusion backed by a real differentiator; editorial → a data point, a take, or a correction.

- Run targets like a pipeline, one row each: source · page · answers fed · mentions us today? · ask · contact · status · date. Work the top first; small placements land fastest.

- After a placement goes live, re-ask the prompts it was feeding in the next cycle. If nothing moves, re-rank.

- Never present a paid inclusion as organic.

Research & build the tracked prompt set (loop step 1, Foundation; reviewed quarterly, owner: the brand's GEO lead). Define WHAT we track before measuring anything, using the GEO best-practice prompt framework.

1 · Research the prompt set (best practice), cover all three axes:

- Journey stages: *awareness* (problem-aware, pre-category, concerns/objections mined from sales calls, Reddit, 'what worries people about [category]'), *consideration* ('best [category]', segment-specific variants), *purchase* (where commercial intent exists).

- Customer context: ICP segments, use-cases, constraints, reverse-engineer from the brand's nav, pricing tiers and positioning if not documented.

- Geography: track core prompts per target region.

- Seed from the brand's top non-branded Google Search Console queries, converted 1:1 into natural questions.

- Keep brand-evaluation prompts ('grex vs X', 'is grex worth it') in a SEPARATE tag/project so guaranteed mentions don't skew visibility.

- Target ≈50–100 tagged prompts (stage · intent · segment · region) + a separate sentiment-theme set (safety, support, pricing, reliability, …).

2 · Build in Peec (primary), or another LLM-analysis tool. Create/organise the prompts in Peec.ai with those tags; add competitors (with name aliases); select tracked models (ChatGPT, Google AI Mode, Perplexity, Gemini). Cross-check coverage in LLMrefs if connected.

3 · Handoff: once the set is live and collecting, weekly GEO Sense begins. Re-run quarterly to expand clusters and retire dead prompts. Human approval the owner to approve the prompt set before it goes live.

Weekly re-measure & report (loop step 5, Verify; owner: the brand's GEO lead, weekly + monthly deep-dive). Close the loop honestly.

- Re-measure the same prompt set via Peec.ai and compare to the pre-action baseline.

- Report trend and correlation, NOT causation, AI answers are probabilistic and attribution is noisy, so track which actions *preceded* movement and build the picture over weeks; never claim one post/mention 'caused' a change.

- Promote tactics that correlate with sustained lift; drop those that don't. Feed self-reported 'how did you find us?' data into the AI-attributed-leads metric (cross-check traffic via Google Analytics, directional only; LLM visits are undercounted).

- Monthly: NotebookLM qualitative synthesis over exported answers to surface recurring narratives/complaints.

- Post the internal weekly trend report automatically; escalate before any external or client-facing reporting.

Loops back to GEO Sense.

Reading movement.

- Count a clear flip in vote counts (e.g. absent in every sample → present in every sample) as movement; a wobble at the margin is noise.

- If the prompt set changes, mark the break in the series, never splice before and after into one trend.

On-page content & optimisation (loop step 4a, Act / On-page). Execute the on-page items from GEO Decide so the brand's own pages get retrieved *and* cited. Produce long-tail, ICP-question content, and hold every page to the extractability bar:

Write for extraction

- Lead with the answer, the first 40–60 words of each section answer the query directly and declaratively (no scene-setting; that's the #1 citation-killer).

- Question-based H2s that mirror real prompts; one H1 per page; paragraphs <120 words; a summary at the top of each section; tables, bullets and numbered steps.

- Be specific & declarative, concrete numbers, dates, named processes; put the brand first in numbered lists.

- Gut check: screenshot any one H2 section, does it answer the question on its own?

Own the high-yield formats

- Build and defend listicles / comparisons, 'best [category]', '[competitor] alternatives', 'X vs Y' (the most-cited format; keep tables genuinely balanced). Consolidate thin single-topic posts into comprehensive listicles with a pillar hub.

Technical & trust

- Schema (Organization, Article w/ author + dateModified, FAQPage, Product/BreadcrumbList), must reflect visible content; validate before publish.

- Entity consistency, identical brand name / NAP / descriptors; real author bios + credentials.

- Retrievability, server-rendered, fast, crawlable; AI bots allowed in robots.txt.

Freshness & quality floor

- Quarterly refresh of top citation-earning pages; bump dateModified honestly.

- No raw-AI slop, every page passes 'would I read this myself? does it add value others miss (original data, first-hand insight)?' Human approval editorial/Tom approval before publish.

Feeds GEO Verify.

Weekly AI-visibility pull (loop step 2, Measure; owner: the brand's GEO lead, weekly). With the tracked prompt set live (see GEO Prompt Research & Build), pull the week's data via Peec.ai:

- Visibility % (answers mentioning the brand), position (rank when mentioned), share of voice vs competitors, sentiment by theme, citation rate vs platform benchmark (ChatGPT ≥2.0 · Google AI Mode 1.1–1.5 · Perplexity 1.5–2.0), and the top cited sources.

- Segment every metric by tag (stage · intent · segment · region), never read a single blended number.

- AI answers are probabilistic: trust the weekly aggregate, not day-to-day swings (≈30% of brands hold visibility answer-to-answer).

- Record the values on this board's metrics, compute week-over-week deltas, post a short summary to Slack/Claude.

- Optional daily lightweight check (visibility + sentiment on the top prompts) that escalates only on a material move.

Output feeds GEO Decide.

Sampling standard.

- Never score a prompt from a single answer: ask each tracked prompt several times per engine per cycle (follow the playbook's sample size) and record a vote count per prompt × engine, not just a verdict.

- Keep the full dated answer text for every sample, it's the evidence behind any claim.

- Separate three outcomes: recommended (offered as an answer), mentioned (named without endorsement), cited (linked as a source).

Continuous, background, run by the GEO expert; visibility is low by design.

1. Find Reddit/forum threads where our ICP asks questions AI search later draws on.
2. Post genuinely helpful, non-spammy comments that surface the brand's POV.
3. Vary timing and communities to stay natural and within subreddit norms, real, agreed accounts only, never sock-puppets (see the Off-page guardrails).
4. Note seeded threads so we can later check whether they show up in AI answers.

Before any post or link (every time, in order).

1. Read the community's rules in full, and its wiki/pinned posts, real posting norms (flair, self-promo days, karma minimums) often live there.
2. Restricted or private community → stop.
3. Check the link hasn't already been posted there.
4. Verdict: go with the constraints that apply, or no-go with the rule that blocks it and an allowed alternative (e.g. the weekly promo thread).

Before any brand-adjacent comment, six checks. Complies with the rules · respects the self-promo policy · answers what the post actually says · reads like a person (varied sentences, specifics from the thread, honest hedging) · matches the community's tone · nothing that looks like a spam pattern.

- Where self-promotion is restricted: no links, answer the question fully, the brand at most in passing. Disclose affiliation where the community expects it.

- A comment that fails gets one rewrite against the specific failure. Fails again → skip the thread. A skipped comment costs nothing; a removed one costs the account.

Account standing. A new or thin account contributes with zero brand mentions for at least its first week, and only raises its promotional share as karma grows (default ladder, tune on your own data: none below ~10 karma, ≤10% to ~50, ≤20% to ~100, ≤⅓ to ~500, never above half). Spread activity across communities and times, bursts in one subreddit read as automation.

escalate any removal, flag or shadowban, stop posting in that community and raise it rather than retrying.

Inputs: target query/topic list.

Outputs: seeded high-signal threads feeding AI-search answer sets.