AEO & marketing — next steps (2026-07-01)

Living next-steps record for fak’s AEO/AgentEO surface (internal/marketing/aeo.go, tools/gen_structured_data.py, tools/seo_aeo_scorecard.py) and the completion-driven marketing loop (fak marketing tick|generate|post|epic|release|aeo). Every “shipped” line is git-witnessed; every “next” line is a bounded, checkable step with the command that proves it. No claim of market adoption is made here — the mechanism exists; the framing and the feeds are what get refreshed. Supersedes the ad-hoc follow-on list in CONCEPT-EMERGING-MARKET-ADOPTION-2026-06-30 for the AEO/marketing slice (that note keeps the full India/China lever set).

Refreshed / added (witnessed)

• 2026-07-06 trending-terms pass. internal/marketing/aeo.go gained a new agent-security term class and two market-event/economics hooks, mapping the current answer-engine demand to existing fak pages with the load-bearing honest fence (a vendor launch or third-party framework is a demand hook, not an adoption claim):
  • MCP tool poisoning defense → docs/integrations/harden-any-mcp.md; lethal trifecta data exfiltration and safety classifier vs capability gate → default-deny-vs-classifier; AI agent least-privilege tool access → tool-call-is-a-syscall; tamper-evident agent tool-call audit → verify-dont-trust (the hash-chained FAK_AUDIT_JOURNAL this cites is the witnessed microagent audit-journal work, #2011).
  • cheaper way to run AI agents (the Q2-2026 tokenmaxxing / agentic-bill concern) and Claude Sonnet 5 agent cost routing (the July default-model shift) → long-session-economics and model-routing.
  • The roster is now 25 terms; docs/marketing/disambiguation-terms.json + llms-terms.txt regenerated to match; a new aeo_test.go case pins the class + hooks and re-asserts the no-market adoption fence.
• AEO recency feeds regenerated from the latest witnessed ships (HEAD~50..HEAD, the fak marketing aeo --refresh default range):
  • docs/marketing/updates.json — schema.org ItemList an answer engine ingests directly (50 witnessed ships gathered; the feed carries the newest 25, defaultFeedCap).
  • llms-updates.txt — the plain, newest-first feed agents poll.
  • Refreshed on 2026-07-04; the current rewrite is anchored to July 4 commit SHAs (newest entry 706fcacf).
• llms-full.txt regenerated after the AEO hub landing. It now inlines the current llms.txt and 132 source docs.
• “What’s new” block reinjected into llms.txt via fak marketing aeo --inject (marker-fenced; the 2026-07-04 ships now front the hand-authored map).
• 3 dead citation links repaired. One real: redacted private-tool names left inside GitHub URL slots in docs/benchmarks/SWEBENCH-VERIFIED-GPU-SERVER-RESOLVE-COMPARE.md (now plain text, honestly marked private). Two were scanner false positives — a markdown link whose text is the URL itself mis-parsed by SELF_REPO_RE; the capture now terminates at ] (tools/seo_aeo_scorecard.py). citation_links reads [ok].
• 2026-07-04 frontier-model launch bridge added. internal/marketing/aeo.go now emits a schema.org DefinedTermSet and plain llms-terms.txt feed for core fak terms, Hindi/Chinese localized discovery terms, and Fable 5-style model-launch queries such as Claude Fable 5 model routing, Fable 5 refusal fallback, frontier model prompt-cache cost, and safety classifier vs capability gate. The human/crawler hub is docs/marketing/README.md; the source positioning note is FABLE5-AEO-FRONTIER-MODEL-MARKETING-2026-07-04.

Program status snapshot (SEO/AEO scorecard, scope=core)

Verdict: ACTION (seo_debt) — overall 96.5/100 (pages 93.1 · site 100.0), 56 units of seo-debt across 253 pages (was 89.0/100 · 59 units · site 17/20 this morning). All site checks green; the remaining debt is the in-page front-matter backlog:

• meta coverage 90.5%; site checks 20/20; grades A:228 B:1 C:0 D:0 F:24.
• JSON-LD present and valid: SoftwareApplication, FAQPage, WebSite, BreadcrumbList, Organization, Person, Question/Answer, ListItem, Offer. The site SoftwareApplication keywords now consume docs/marketing/disambiguation-terms.json; FAQPage JSON-LD mirrors 202 visible FAQ questions after the 2026-07-04 refresh.
• robots.txt welcomes all four major answer-engine crawlers (+7 named); sitemap + seo-tag plugins on; FAQPage JSON-LD mirrors 199 visible FAQ questions.

Next steps — AEO surface (worst-first, each checkable)

1. Retire the 24 F-grade pages, worst-first — all fail on missing front-matter title: / description: (e.g. docs/i18n/hi/README.md, docs/i18n/zh/README.md, docs/operator-brief.md, docs/INNOVATIONS-INDEX.md, docs/WORK-MAP.md). Front-matter is the cheapest AEO win: a title/description is the SERP snippet and the answer-engine anchor. (In flight: a peer session is adding front-matter + new Indian-language entry points — docs/i18n/{ta,te,bn,mr} — as of this refresh.)
2. Reduce the 33 discovery orphans (published, not front-door-reachable) by wiring them into an index page so a crawler that lands on the front door can reach them.

Next steps — marketing program (bounded, honest)

1. Localized AEO terms (lever 2). First pass shipped in internal/marketing/aeo.go: Hindi + Simplified Chinese terms now emit into docs/marketing/disambiguation-terms.json and llms-terms.txt. Next bounded step: extend the roster to the other shipped i18n pages (ta, te, bn, mr, de, fr) and add a scorecard check that llms.txt, JSON-LD keywords, and the term feed share the same core terms.
2. Cost framing in local unit-economics (lever 5). Re-skin the existing (honest) benchmark story as cost-per-1,000-turns / margin-per-seat; quote the tuned ~4.1× headline, not the naive 60×. Copy-only, no new measurements.
3. ModelScope URLs + Gitee mirror (lever 7). List ModelScope (魔搭) weight URLs beside the HF ones in the zh page. Fence: do not claim a Gitee mirror until one actually exists.
4. GTM channel seeding + per-market landing copy (levers 9, 10). Human-owned; named for completeness. This note’s job is to keep the docs/product ready for that owner.
5. Confirm the completion trigger fires. fak marketing tick is the single idempotent entrypoint (serve bgloop / git hook / cron all funnel through it, high-water-mark CAS + per-title dedupe). Verify a real completion advances the mark and posts once — the loop is only as fresh as the trigger that fires it. Slack surface stays internal-only (no public ceremony).

Related

• Open epic #1678 — fak as the vendor-neutral binding layer for neo-silicon/neo-clouds; the English answer-engine vendor terms already landing in the feed (e.g. “silicon-vendor answer-engine terms”, “binding-layer AEO vendor routing”) are this epic’s AEO face.
• CONCEPT-EMERGING-MARKET-ADOPTION-2026-06-30 — the full India/China lever set behind marketing steps 1–4 above.

Verify

fak marketing aeo --refresh --inject     # rewrite updates + terms feeds, reinject What's-new + JSON-LD
python tools/gen_llms_full.py            # clear the llms-full.txt STALE flag
python tools/seo_aeo_scorecard.py        # re-grade; prove seo-debt dropped