> ## Documentation Index
> Fetch the complete documentation index at: https://docs.synscribe.com/llms.txt
> Use this file to discover all available pages before exploring further.

# SEO principles (source)

> Mined source outline of Synscribe's SEO stance — raw material for Part 1.

**Purpose:** raw material for drafting the 12 Part-1 pages. Each section gives the
*non-obvious* stance Synscribe takes, the mechanism behind it, concrete supporting
detail from sources, Raymond's verbatim phrasing worth keeping, and `❓ [needs Raymond]`
flags for anything I couldn't substantiate.

**Sources mined**

* **Workshop Notes** — "SEO/GEO Workshop, 26th May" (Notion, read in full). Primary backbone.
* **Wonderchat dossier** (Notion) — AI chatbot SaaS; BOFU-landing-page account, pillar strategy.
* **Hyperbound dossier + "Product Bible & Landing Page Strategy"** (Notion) — AI sales-roleplay
  SaaS; feature→keyword mapping, industry pillars, claims validation.
* Two referenced framework pages — `crimson-thought-866.notion.site/serp-cliff` and
  `…/2-dimension-keyword-labelling-framework` — are **JS-rendered public Notion and could not be
  fetched** (returned an empty shell). Their *concept* is captured from how the Workshop Notes
  invoke them; deeper specifics are flagged `❓`.

Two framings recur across everything and are worth stating once, up front, in Raymond's words:

> **"SEO/GEO is NOT rocket science."** — *Best match to search intent wins · Pick fights you can
> win · Fat-tailed probability distribution for queries → small fights yield meaningful results ·
> Long-tail query → higher search intent.*

> The hard part isn't SEO. *"To find \[the high-intent long-tail], ask what do you uniquely know
> about your ICP and how your product/service is positioned to solve their pain — **this is hard
> (and it's not SEO/GEO).**"* The moat is ICP knowledge, not tactics.

The house method, verbatim ("How we 'do SEO'", an **8-step, 4-week loop**):

1. Research the company → 2. Form hypothesis on how ICPs search → 3. Simple technical audit →
2. Create blog posts → 5. Create landing pages → 6. Get backlinks → 7. Track progress →
3. **Validate or invalidate hypothesis.** *"All in a 4 week loop."* — SEO is run as a
   hypothesis-testing loop, not a content quota.

***

## 1.1 — The Synscribe thesis: programmatic SEO + GEO for SaaS

**Non-obvious claim.** SEO is now a *search-intent-matching* game played across a fat-tailed
distribution of queries, and the winning move is to **pick small, winnable, high-intent fights
at scale** — programmatic landing pages + GEO — rather than chase volume/keyword-difficulty on a
few big head terms. The durable advantage is not tactical; it's *knowing your ICP better than
anyone*, which SEO merely harvests.

**Mechanism.** Queries follow a fat-tailed distribution: an enormous number of low-volume, very
specific queries collectively outweigh the few high-volume ones. Long-tail queries carry *higher
purchase intent* and are *much easier to rank* — so a portfolio of many small wins beats one hard
fight. Programmatic SEO is how you manufacture many pages that each match one specific intent;
GEO is the same logic applied to answer engines. The prerequisite (and the real work) is having
"100s of ways to describe your product" grounded in *customer-speak, not founder-speak*.

**Supporting detail.**

* Wonderchat's plan enumerates *pillars* — competitors, use-cases (by industry, tech/platform,
  data-source, objective, replacement, language, people, compliance) — i.e., the ICP knowledge
  fanned out into a page-generation program, not a keyword list.
* Hyperbound's engagement literally begins with a **Product Bible built from "our research, your
  website, call transcripts, and public sources"** sent to the client to *confirm/correct* before
  scaling page production — ICP knowledge is treated as the input to SEO, exactly as the thesis says.

**Verbatim to preserve.** "Best match to search intent win." · "Pick fights you can win." ·
"Fat-tailed probability distribution for queries → Small fights yield meaningful results." ·
"This is hard (and it's not SEO/GEO)."

**❓ \[needs Raymond: positioning vs. blog-mills/agencies].** The Workshop Notes never explicitly
contrast Synscribe with agencies/blog-mills. Wonderchat "using a UK agency right now with a 6-month
contract" is a hint of the competitor, but the *sharp* positioning line ("agencies sell you N blog
posts/month; we run a hypothesis loop that manufactures winnable intent-matched pages") is my
inference — confirm the wording you want.

***

## 1.2 — Two search surfaces now: Google **and** AI answer engines (GEO)

**Non-obvious claim.** You are no longer optimizing for *one* searcher. There are now **three eras
of search running at once**, and the newest ones don't send a single query — they *fan out* into
many. You must optimize to be the source a *machine* picks from a basket of fan-out queries, which
changes what content wins.

**Mechanism (Raymond's diagram, verbatim):**

* **Past:** Human — query → Google
* **Now:** Human — query → ChatGPT — query → Google
* **Next:** Human — *task* → AI Agent — query → Google

> *"LLM turns a user query into fan-out queries which gets sent to search engines."*

Because the LLM issues many reformulated queries and then *cites* a handful of sources, the game is
to be citable across the fan-out, not just to rank #1 for the human's original phrasing. You verify
this empirically by *watching the machine's actual queries*.

**Supporting detail.**

* **"What ChatGPT Gobbles Up"** (citation-type data, promptwatch.com): *Listicles* ("Top X", "X
  Best") are the obvious top; **landing pages are "the real winner, taking both 2nd & 3rd position"**;
  *"How-to" articles are "the next secret weapon"*; comparison/review pages win *"for B2B with long
  sales cycle and buying committees."* This citation-type hierarchy is the GEO content strategy.
* Synscribe ships tools to *see the queries*: a Chrome extension to reveal ChatGPT's web-search
  queries, and **Birdseye** (a macOS app) to see Claude Code's queries. Method: *"Step in the shoes
  of your customer / a serious customer researching the space, how would they find your product?"*
* Real-world citation examples given: *best duty drawback → Zollback · hr intake automation → Jinba ·
  how to set reminder linkedin dm → Kondo · assent vs comply pro → Reglyr.*

**Verbatim to preserve.** "Landing pages are the real winner taking both 2nd & 3rd position." ·
"'How to' articles are the next secret weapon." · "Step in the shoes of your customer."

**❓ \[needs Raymond: referral tracking].** PLAN 1.2 says "why we track ChatGPT/Perplexity/Claude/
Gemini referrals," but the Workshop Notes cover *seeing queries* (extension/Birdseye), not the
*referral-tracking* rationale. Confirm the tracking philosophy (likely lives in the PostHog/Onboarding SOP).

***

## 1.3 — The atomic unit: keyword → page, and why we start **BOFU**

**Non-obvious claim.** The atomic unit of the whole system is **one keyword → one page**, and you
build **bottom-of-funnel first, then work *up*** the funnel — the opposite of the classic
"top-of-funnel blog to build awareness" playbook.

**Mechanism.** BOFU keywords are (a) *much higher relevance* — they "describe your product exactly"
— and (b) *much easier to rank*. So BOFU is where intent and winnability are both highest: you
capture people already trying to buy, on terms you can actually win, before spending effort on
broad TOFU terms that are harder to rank and convert worse.

> *"Your job is to start from the BOFU keywords first and work your way up."*

**Supporting detail.**

* **Wonderchat**, third pre-onboarding meeting: *"focus on BOFU landing pages only."* Their existing
  page *"wonderchat whitelabel ai chatbot for agency"* was generating leads — the account was
  explicitly steered to *more of that BOFU pattern*, not awareness content.
* BOFU maps onto the landing-page-first content mix (1.6): `/uses`, `/feature`, alternatives, vs,
  integration pages are all BOFU by construction.

**Verbatim to preserve.** "Start from the BOFU keywords first and work your way up." · "Much higher
relevance — describes your product exactly. Much easier to rank."

***

## 1.4 — SERP-cliff analysis: our "can we actually win this?" method

**Non-obvious claim.** Before writing anything, you run a **SERP cliff** analysis to decide *whether
the fight is winnable* — and you *drop* keywords you can't win, even high-volume ones. Volume and
keyword-difficulty scores are **not** the selection criteria; a visible "cliff" in the SERP (a
drop-off you can slot a new page above) is. *"Pick fights you can win"* is a literal gating step,
not a slogan.

**Mechanism.** A SERP has a "cliff" when the results below the top few are weak, off-intent, or
low-authority — leaving an exploitable gap a well-matched new page can leapfrog into. The keyword
eval also surfaces *"how other sites are getting asymmetric wins"* — i.e., where competitors ranked
without deserving it, which is a template for your own win. This is why small/long-tail fights
(1.1) are attractive: the cliff appears fastest there.

**Supporting detail.**

* Keyword Quality Gate (verbatim): *"Does this look like they want to buy? · Does my company offer
  this (or solve the same problem)? · Is the source of keyword credible? · **Can I win on this
  keyword?**"* — winnability is a first-class survival gate.
* Discovery Quality Gate closes with: *"Have I looked up those terms and verified I can win?"* with
  the footnote *"Refer to SERP gap document."*
* Operator prompt in production: *"look at ALL the keywords right now and run it through the keyword
  evals"* → the eval report returns winnability + asymmetric-win intel per keyword.

**Verbatim to preserve.** "Pick fights you can win." · "Can I win on this keyword?" · "Keyword eval
report also gives you how other sites are getting asymmetric wins." · "Refer to SERP gap document."

**❓ \[needs Raymond: the SERP-cliff mechanics].** The dedicated method page
(`crimson-thought-866.notion.site/serp-cliff`) wouldn't render, so the *specific* signals/thresholds
that define a "cliff" (DR gaps? weak-page markers? position-by-position scoring?) are not
substantiated here — pull the exact model from that doc / the Organise-Keyword SOP, which PLAN 1.4
says "has the model."

***

## 1.5 — Intent & cannibalization discipline (one keyword ≠ always one page)

**Non-obvious claim.** Every page must target **exactly one search intent with exactly one keyword**
— but the corollary is a *cannibalization discipline*: two of your own pages must not compete for
the same intent, and one keyword doesn't automatically earn its own page unless the intent is
distinct. Keywords are triaged through a **2-dimension framework (intent × relevance)** and only the
"clean accept" survive.

**Mechanism.** If two pages chase one intent, they split authority and confuse the engine about
which to rank (self-cannibalization). Conversely, near-duplicate keywords that share an intent
should *collapse to one page*. The 2-D framework separates *how commercial/high-intent* a query is
from *how relevant it is to your product*, so you keep only high-intent × high-relevance terms and
tag them for production. Relevance-first also guards against chasing volume that never converts.

**Supporting detail.**

* Landing Page Quality Gate (verbatim): *"Does it target ONE search intent with ONE keyword?"*
* Production tagging discipline: *"load all keywords into synscribe, annotate them and tag all the
  clean accept keywords as 'Core'"* — an explicit accept/reject gate before anything is built.
* Framework reference: *"see 2d keyword labeling framework"* (the BOFU-first ordering in 1.3 is the
  intent axis of this framework in action).
* Wonderchat's pillar list is deliberately *non-overlapping* by facet (industry vs tech vs
  data-source vs objective vs replacement vs compliance) — cannibalization avoided by construction.

**Verbatim to preserve.** "Target ONE search intent with ONE keyword." · "Tag all the clean accept
keywords as 'Core'." · "Does my company offer this (or solve the same problem)?"

**❓ \[needs Raymond: the 2-D framework axes + the signal hierarchy].** The framework page didn't
render. I'm asserting the axes are **intent × relevance** (consistent with the Workshop Notes'
BOFU/relevance language), but the exact labels, the quadrant→action rules, and the "survival gates /
signal hierarchy" PLAN 1.5 mentions are unconfirmed — pull from the framework doc / Organise-Keyword SOP.

***

## 1.6 — Landing-page **pillars**: horizontal (`/uses`) vs vertical (`/feature`)

**Non-obvious claim.** Landing pages come in exactly **two shapes that create two different
*feelings*** in the ICP, and you build them in a fixed order — **wide before deep**. This is a
*programmatic scale* model: a template + data (CSV/JSON) generates a whole pillar of pages, fed by
a "hub" page that funnels traffic and by internal links from the homepage.

**Mechanism.**

* **Horizontal** (`/uses`, `/use-cases`, `/industry`): go *wide* across use-cases/features. ICP
  reaction: *"this does everything I need."*
* **Vertical** (`/feature`, `/product`, `/integration`): go *deep* into one feature. ICP reaction:
  *"oh wow, it's exactly what I want (that other products don't have)."*
* **Wide first because:** *higher volume, easier to build.* Depth comes after the wide net is cast.
* Scale is mechanical: *template + CMS, split static vs dynamic components, bring .csv/.json to
  generate many pages.* A page can be *dynamically composed* — Hyperbound tags each of 12 features
  with a `feature id` and a "when to use it" rule so the *right feature image/section is selected by
  the page's keyword intent*.

**Supporting detail.**

* **Wonderchat**: all near-term capacity *"allocated to uses (120 uses page)"* — a single horizontal
  pillar of \~120 pages — with vertical/other pillars (competitors, integrations, compliance,
  industry) planned behind it; competitor sitemaps (chatfuel, botpenguin, tars, zendesk) mined to
  populate the pillar. Example horizontal pillar in the wild: `reglyr.com/uses`.
* **Hyperbound**: industry pillar (Manufacturing, Medical/Pharma, Financial Services, Insurance,
  Oil & Gas), each page mapping a *feature set* (e.g., Multiparty Roleplays + Enterprise Security
  for regulated verticals) to the vertical's pain. Feature→keyword rule example: *"AI Real Call
  Scoring — highlight for keywords related to call scoring, QA… strong for enterprise/management
  keywords."*
* Other page types named: **Alternatives** ("strapi alternative"), **2-way VS** ("contentful vs
  tina"), **Directories** (esp. B2C: "plumbers in sf"). *Comparison/VS pages exploit the LLM's baked-in
  bias* — "use it to your advantage with aka 3-way or 4-way vs pages."
* Organization: a **hub page** ("Use Cases for AI Chatbot") lists the pillar, is itself SEO-ed,
  linked from header/footer — *"Use a 'hub' page to filter traffic into pSEO landing pages."*

**Verbatim to preserve.** "Go wide before going deep." · "this does everything I need" /
"oh wow, it's exactly what I want (that other products don't have)." · "Use a 'hub' page to filter
traffic into pSEO landing pages." · "Landing pages account to >30% of what gets cited by ChatGPT."

***

## 1.7 — Speed-to-rank: the new-domain playbook

**Non-obvious claim.** A brand-new domain can put **5–10 pages into positions 1–5 within \~2 weeks**
— ranking is *fast* when you stack the deck: BOFU + long-tail + a verified SERP cliff + exact-match
titles, on the winnable small fights. Speed is a *design choice*, not luck.

**Mechanism (from the parts, assembled).** Each ingredient compounds to make ranking quick:
long-tail/BOFU terms are low-competition (1.3); the SERP-cliff gate means you only enter fights with
a visible gap (1.4); exact-keyword-in-title matches the intent the engine (and LLM) reads first
(1.11); programmatic pillars ship *many* attempts at once so the fat tail pays off quickly (1.1/1.6);
IndexNow + manual GSC indexing get pages *seen* fast (1.12/indexing). Synscribe runs its own domains
this way — the *"zero-to-ranked"* launches and a *"One Startup A Day launch"* SOP are the proof the
cadence is real.

**Supporting detail.**

* Reference repo is the "template nextjs website we use for **our own launches**" (`synscribe.com/
  zero-to-ranked`) — Synscribe dogfoods the fast-launch playbook.
* Indexing reality that bounds the speed: *"Your manual crawl budget starts out small \~3–5 to
  \~10–20 per day"* — so you *manually* index the priority pages first; Bing via IndexNow is
  automated on publish.

**Verbatim to preserve.** "zero-to-ranked." · "Manual crawl budget starts out small \~3–5 to \~10–20
per day."

**❓ \[needs Raymond: the "why it's achievable" theory — HIGH PRIORITY].** The specific claim
"5–10 pages, pos 1–5, in 2 weeks" is from **PLAN.md, not the Workshop Notes**, and PLAN explicitly
tags 1.7 as "I draft the quick-win theory (Raymond reviews)." I've assembled a *plausible* mechanism
above from the other principles, but the real theory — *why* a fresh, low-authority domain ranks
that fast (is it purely cliff+long-tail? does GEO/citation rank differently from classic SEO? what
role do backlinks/press play in the 2 weeks?) — is **not substantiated by any source** and needs
your authoritative version.

***

## 1.8 — Authority & internal linking (hub/spoke, footer links)

**Non-obvious claim.** Internal linking is not an afterthought — it's the *authority-distribution
mechanism* that makes a programmatic pillar actually rank. The rule is concrete and easy to get
wrong: pillar pages must be **reachable by clicking from the homepage**, funneled through a **hub**,
and the hub's links must be **present in the DOM even when visually collapsed**.

**Mechanism.** Crawlers and LLMs discover and weight pages through links they can actually parse. A
pillar of 120 pages with no path from the homepage is invisible; a hub page concentrates internal
link equity and gives the crawler one high-value page that fans out to the spokes. The subtle
failure mode: designers hide long link lists behind accordions/JS for readability, which removes
them from the DOM — so the links stop counting. *"Nothing beats having a navigable site."*

**Supporting detail.**

* Landing-page technical rules (verbatim): *"Must be added to sitemap · Must be discoverable by
  clicking from home page · Use a 'hub' page to filter traffic into pSEO landing pages."* Link the
  hub from header or footer.
* The DOM-visibility trap, called out twice: *"If you collapse the list to make it more readable,
  make sure links are still 'visible' in the DOM even if they are collapsed."* There's even a
  ready-made Lovable prompt to fix it: *"The links in my header nav bar are hidden in the DOM, can
  you make it visible in DOM even though it's hidden from sight…"*
* Wonderchat focus item: *"Ship header and footer"* and route *"'Enterprise' link to uses"* — header/
  footer wiring is treated as a shippable SEO task, not cosmetics.

**Verbatim to preserve.** "Nothing beats having a navigable site." · "make sure links are still
'visible' in the DOM even if they are collapsed." · "Use a 'hub' page to filter traffic into pSEO
landing pages."

**❓ \[needs Raymond: hub/spoke authority theory depth].** PLAN 1.8 references a "How to build strong
internal links" SOP (🔬 mine from prod). The Workshop Notes give the *rules* but not the *authority-flow
model* (how much equity a hub should pass, spoke→spoke linking, cross-pillar linking). Pull from that SOP.

***

## 1.9 — Off-page: link building & press for the bump

**Non-obvious claim.** Off-page work is run for a **specific, timed "bump"** — a press release or
link-building push to *nudge* rankings at the right moment — and outreach is treated as an
*always-on setting you turn on day one*, not a one-off campaign. Backlinks are step 6 of the 4-week
loop, deliberately *after* the pages exist.

**Mechanism.** Programmatic pages get you into contention (1.4–1.7); off-page authority is the extra
push that converts "ranking on page 1" into "top few / cited." Because it's sequenced after content,
the links point at pages already matched to intent, so the authority lands where it can convert. A
press release is used specifically for a *quick bump* to accelerate an otherwise slow climb.

**Supporting detail.**

* Setup instruction, verbatim: *"Turn on Link Building Outreach **NOW**"* — it's the first toggle on
  the Features page, alongside IndexNow/GSC/PostHog.
* The 4-week loop lists *"6. Get backlinks"* between "Create landing pages" and "Track progress."
* **Hyperbound** is the worked press example (SOP "Press Release for a Quick Bump"; PLAN 2.5.6/3.10).
  **Wonderchat** has a dedicated *Link Building* workstream in its dossier.
* Workshop-hours menu offers *"link building"* as a deep-dive topic — a standing capability.

**Verbatim to preserve.** "Turn on Link Building Outreach NOW." · "Press release for a quick bump."

**❓ \[needs Raymond: the "bump" mechanism].** *Why* a press release produces a ranking bump (fresh
backlinks? referral traffic signal? timing relative to indexing?) and *how big/durable* the bump is
aren't spelled out in the sources — confirm the causal story and the Hyperbound outcome if there's a
number to cite.

***

## 1.10 — Attribution that matters: AI traffic + pillar-level conversion, not vanity

**Non-obvious claim.** We attribute to **business outcomes at the pillar level and to AI-answer-engine
traffic** — not to rankings/impressions/traffic as ends in themselves. The performance question is
"did this *pillar* produce enquiries/sign-ups/paid conversions?", reconciled across GSC (what ranks)
and PostHog (what converts), with explicit tracking of traffic *from AI answer engines*.

**Mechanism.** Rankings are a leading indicator, not the goal; the loop's step 8 is *"validate or
invalidate hypothesis,"* so attribution exists to *decide which hypotheses to double down on*. GSC
tells you what got seen/clicked; PostHog tells you what those visitors *did*; the join tells you
which pillar/intent actually drives revenue. Tracking AI-referred traffic separately matters because
that surface is the growth edge (1.2) and is invisible to classic rank tracking.

**Supporting detail.**

* **Wonderchat's declared key metrics** (verbatim): *"Enterprise enquiries (form submissions) · Free
  trial sign ups (landing page → account creation) · Subsequent conversion to paid services"* —
  conversion, staged down the funnel, *per landing surface*. Not "traffic."
* The reference-repo template ships attribution primitives: *"User journey trace in contact form"*
  and a *"Dual contact form (contact & get free consultation)"* — attribution designed into the page,
  not bolted on.
* Workshop-hours deep-dive topic: *"attribution with GSC & Posthog."*

**Verbatim to preserve.** "Enterprise enquiries (form submissions) / Free trial sign ups (landing
page to account creation) / Subsequent conversion to paid services." · "Validate or invalidate
hypothesis." · "User journey trace in contact form."

**❓ \[needs Raymond: the PostHog philosophy].** PLAN 1.10 says the attribution philosophy comes from
the Onboarding SOP (PostHog section) — not in the Workshop Notes. The *specific* AI-traffic-insight
method and the SQL/funnel setup (PLAN 2.6.1 references "SQL snippet in hand") should be pulled from
that SOP to make this page concrete.

***

## 1.11 — Quality as a gate: our Definition of Done

**Non-obvious claim.** Quality is enforced as a **hard gate at every phase**, phrased as
**answerable yes/no questions the operator must pass** — not a post-hoc checklist. Each stage
(Discovery, Product Bible, Keywords, Blog Post, Landing Page) has its *own* gate, and a piece of
work is not "done" until it passes. Critically, the gates are written from the *reader's/ICP's* point
of view ("can I *see* it?", "would ChatGPT cite this?"), and include a **claims-safety gate** — flag
anything you can't say publicly *before* publishing.

**Mechanism.** Gates make "done" objective and repeatable across operators/agents (this is the
discipline delta between self-serve and agency-run accounts). Phrasing them as ICP-perspective
questions forces the operator to simulate the searcher/engine rather than grade their own effort.
The claims gate exists because programmatic pages scale *fast*, so an unvalidated claim would
replicate across a whole pillar.

**Supporting detail — the actual gates (verbatim):**

* **Discovery Gate:** *"Can I see my ICP when I close my eyes? · Do I have >20 ways to describe my
  product? (if 20 is hard, try starting with 200) · Have I looked up those terms and verified I can
  win?"*
* **Product Bible Gate:** *"Do I know which feature is most important for which ICP? · Do I know what
  my ICPs are going to name those features? · Do I know what are my ICPs buying triggers?"*
* **Keywords Gate:** *"Does this look like they want to buy? · Does my company offer this? · Is the
  source credible? · Can I win on this keyword?"*
* **Blog Post Gate:** *"Does this blog post look like what ChatGPT would search for and cite? · Does
  it have the potential to sell your product? · Can you win with this blog post? · Are the
  keyword/query exist in the blog post title?"*
* **Landing Page Gate:** *"Can I 'see' my landing page? · Does this make my ICP go 'THIS IS EXACTLY
  IT!'? · Does it target ONE search intent with ONE keyword? · Do I have >20 pages?"*
* **Technical DoD (the "if you can only get 3 things right"):** Title, Description, H1 — *"surprisingly
  important because both human & LLMs are looking at just these in the search results and ask 'which
  of these is more relevant or interesting for me?'"* Then: first paragraph, heading flow, robots.txt,
  sitemap.xml, JSON-LD, llms.txt. And: *"Most technical SEO is just about improving experience for a
  human… \[tools give you] 100s of things to fix, they don't even improve the experience, ignore those."*
* **Claims-safety gate (Hyperbound):** the Product Bible is sent to the client to *"Flag any claims
  we should NOT make (metrics, customer names, positioning)"* and to confirm metrics like *"50%
  faster ramp," "150% increase in DM→demo conversion," "2x faster time to first won deal"* before any
  of them ships to a landing page.

**Verbatim to preserve.** "Can I see my ICP when I close my eyes?" · "THIS IS EXACTLY IT!" · "Does
this blog post look like what ChatGPT would search for and cite?" · "ignore those" (on junk
technical fixes) · "Flag any claims we should NOT make."

***

## 1.12 — Agent-operated SEO: why an AI assistant changes the operating model

**Non-obvious claim.** SEO here is **operated *through* an AI agent (Pi), not by a human clicking
tools** — the operator's job shifts from *doing the steps* to *directing an agent that does them in
parallel with subagents*. This is the "Next" era from 1.2 turned inward: the *human gives a task,
the agent fans out the queries and work*. The leverage isn't speed on one task; it's running the
whole 4-week loop as delegated agent work with durable memory.

**Mechanism.** Pi is *"the AI agent that's central to entire operations… what you'll mainly deal with
90% of the time."* Work is issued as a *task*, and Pi decomposes it — *"do in sequence, passing data
from prev to the next, use subagents"* — so research → product bible → keywords → pillar planning
run as a chained agent pipeline. State persists in a virtual filesystem (`/memory/[company]/
research.md`, `product-bible.md`) so the agent's knowledge compounds across the loop rather than
living in an operator's head. The same agent-perspective is used to *audit* the market (Birdseye
shows how Claude Code searches), closing the loop between "how agents search" and "how we build for
agents."

**Supporting detail.**

* The canonical operator task (verbatim, the workshop's live demo prompt):
  > *"I'm working on pixo.video, help me: 1. do a website research 2. build a product bible 3. update
  > my org info 4. do deeper research on what bofu keywords I need, see 2d keyword labeling framework
  > 5\. expand on the keyword and think about programmatic seo templates 6. (in parallel to 5) using
  > the product bible help me plan a horizontal landing page pillar (create it). do in sequence,
  > passing data from prev to the next, use subagents."*
* Other real one-line operator commands: *"Do a website research" · "Build a product bible" · "look
  at ALL the keywords right now and run it through the keyword evals" · "load all keywords into
  synscribe… tag all the clean accept keywords as 'Core'" · "Do a technical SEO audit on \[site]" ·
  "setup my image generators / setup my default refiners."*
* Files land in Pi's working directory (`/memory/[company]/research.md`, `product-bible.md`) — the
  operator *reads the agent's work product*, they don't produce it by hand.

**Verbatim to preserve.** "The AI agent that's central to entire operations." · "What you'll mainly
deal with 90% of the time." · "do in sequence, passing data from prev to the next, use subagents." ·
"Next: Human — task → AI Agent — query → Google."

**❓ \[needs Raymond: the leverage thesis + VFS conventions — flagged in PLAN as your draft].** PLAN
1.12 tags this "I draft the Pi thesis / where the leverage is (Raymond reviews)." The *economic*
argument (what an agent-operated model lets one AM do vs. a manual operator, and where the real
leverage sits — parallelism? consistency? the compounding memory?) is my framing above and needs
your authoritative version. Also, exact VFS folder conventions (`/reports/`, memory, soul) are to be
confirmed from prod/code per PLAN 2.0.4.

***

## Cross-cutting phrases worth a "voice" callout somewhere in Part 1

* "SEO/GEO is NOT rocket science."
* "Pick fights you can win." / "Fat-tailed probability distribution… small fights yield meaningful results."
* "Customer-speak, not founder-speak." / "100s ways to describe your product."
* "'Free' is a strong modifier, use it when relevant."
* "Do not 'argue' with customers on the technicalities." (meet the query as phrased, don't correct it)
* "Using LLM's Bias to Win" — comparison/VS pages exploit the model's baked-in preferences.
* "Nothing beats having a navigable site."