HTXpunk LLC / Software / SFE-Edge
SFE-Edge
SFE-Edge sits in front of your site on Cloudflare. When an AI crawler visits, it rewrites document structure—headings, chunk size, lists, emphasis, and schema—so generative engines can extract clean passages. When a person visits, nothing changes: they get your original page. That is the product. Citation lift depends on your content and the engines; we optimize structure, not promises of guaranteed rankings.
Why structure decides who gets cited
Generative engines no longer reward keyword density the way classical SEO did. Deployed large language models and answer engines synthesize direct narrative answers, and brands are either named inside that synthesis or they disappear from the journey entirely. Retrieval-augmented generation and training-time attribution both depend on documents that transformers can chunk cleanly. Attention dilutes when headings sprawl too deep or when monolithic prose blocks hide the only quantitative claim ten screens down. Structural Feature Engineering treats hierarchy, chunk boundaries, lists, tables, and emphasis as first-class ranking features for citation probability—independent of inventing new claims or stuffing hidden text.
Field practice shows that hosting a static /llms.txt index alone rarely moves citation rates, because those files do not change the micro-level structure of the parent pages where models execute final extractions. A publisher that only ships a Markdown map is still asking crawlers to parse dense, unbalanced HTML with flat heading trees, eighty-word paragraphs that never break, and zero machine-readable schema around numeric assertions. SFE-Edge attacks that gap at the network edge: when GPTBot, ClaudeBot, PerplexityBot, OAI-SearchBot, or Google-Extended arrive, the Worker fetches origin HTML, compiles macro meso and micro mutations under a fifty-kilobyte buffer gate, caches the result in the edge Cache API with rare KV persistence, and returns an optimized page while human browsers always receive the unmodified origin response.
The GEO-SFE three-layer framework
The framework follows hierarchical structural feature engineering for generative optimization: macro document architecture, meso information chunking, and micro visual emphasis. Each layer maps to concrete DOM transforms that remain resilient as surface keyword tricks drift.
Macro-structure: document architecture
Macro transforms target heading depth in the range three to five so self-attention neither starves for markers nor dilutes across h6 noise. Sections should share roughly balanced mass—about thirty percent plus or minus ten—so one mega-section does not dominate extraction. For documents exceeding roughly two thousand words, SFE-Edge injects an entity-anchored table of contents early in the body, inside the high-influence zone.
Heading depth policy
Flat pages with only an h1 lack navigational tokens. Deep CMS templates that emit h6 for styling waste structural budget. Prefer semantic depth over visual size; style with CSS.
Section balance heuristic
Balance does not invent content. It flags or lightly reorders when safe. Pilot pages should keep major h2 sections similar in length so balance scores stay near the target band.
Meso-structure: information chunking
Meso transforms partition long prose, promote lists and tables for qualitative comparisons, and compile numeric assertions into JSON-LD FAQPage, Dataset, or Product types when signals exist. Dense blocks over eighty words are split into approximately three-sentence segments using native tokenizers—not catastrophic backtracking regular expressions—so free-tier CPU stays inside the ten-millisecond envelope.
Consider a typical product marketing page that describes pricing, setup time, cache hit rates, and form completion improvements inside a single unbroken paragraph that runs well past one hundred words without a single list or schema hook. Retrieval stages frequently skip that block because there is no clean passage boundary and no structured claim surface. After meso compilation, the same facts become short paragraphs, optional list items, and a machine-readable FAQ schema in the document head, which increases the odds that a generative engine will lift a precise statistic into a synthesized answer while preserving the original semantic meaning and avoiding cloaking-style keyword injection that would violate trust with both users and search systems over time.
Schema injection rules
Schema is injected only when numeric or clear factual signals exist. SFE-Edge prefers FAQPage for question-shaped claims and avoids inventing product offers that are not on the page.
Micro-structure: bold-emphasis anchors
Micro transforms wrap at most one primary quantitative or factual claim sentence per section in strong tags. That guides token-weight prioritization toward the citable assertion without painting entire sections bold. Pilot sections below each contain one clear claim so micro-structure can be verified after a bot rewrite.
Quantitative statistics for pilot extraction
The following statistics are written as plain prose so meso heuristics can detect numbers and optionally promote structured presentations. Values are illustrative product targets and research-oriented benchmarks used for structural testing—not guaranteed customer outcomes.
- Structural optimization programs report approximately 17.3 percent citation rate improvement when structure varies and semantics stay constant across controlled trials.
- Perceived qualitative value scores rise by roughly 18.5 percent under the same structure-only interventions across six mainstream generative engines in the referenced study design.
- AI-referred sessions show an average duration near 8 minutes versus 2.5 minutes for traditional search, with form completion around 22 percent versus 7 percent on standard channels in industry writeups.
- Structured lists and tables can increase visibility inside model responses by an estimated 30 to 40 percent relative to dense prose alone.
- About 44.2 percent of generative extractions come from the first 30 percent of body content, which is why answer blocks front-load here.
- SFE-Edge product pricing for waitlist validation is 29 dollars per month Starter, 99 dollars Pro, and 299 dollars Enterprise.
- Free-tier Workers targets include roughly 2 to 4 milliseconds compile time under a 10 millisecond CPU budget and a 50 kilobyte full-buffer cap before streaming degrade.
- Cache design assumes a 24 hour TTL, Cache API primary storage, and Workers KV writes sampled near 10 percent to stay under 1000 free writes per day.
How SFE-Edge runs in production
The Worker detects AI user agents, checks edge cache then KV, singleflight coalesces concurrent misses, fetches origin sequentially to respect the six concurrent outbound connection cap, and returns headers such as X-SFE-Mode, X-SFE-Cache, and X-SFE-Profile. Humans always passthrough. Origin and proxy hosts must differ: sfe.htxpunk.com rewrites content fetched from htxpunk.net so Cloudflare never loops the Worker on itself.
Freemium machine files
Free /llms.txt remains a colon-delimited CommonMark map so agents can discover pages. Paid bulk ingestion, when enabled, lives on /llms-full.txt behind HTTP 402 and x402 USDC settlement on Base—preserving discoverability while monetizing high-token dumps.
What this pilot page proves
After ORIGIN_URL points at htxpunk.net, fetch this path with a GPTBot user agent through sfe.htxpunk.com and compare against a human user agent. Expect buffered mode markers, optional TOC if word count thresholds are met after expansion, paragraph chunk attributes on the long prose above, schema when facts parse, and at most one strong claim per section. Citation lift remains a hypothesis to measure—not a guarantee printed as fact.
Next steps for operators
Join the waitlist on sfe.htxpunk.net, keep AI crawlers allowed in Cloudflare bot settings, deploy this hub to htxpunk.net, then flip ORIGIN_URL and run the smoke matrix documented in docs/PROJECT_LOG.md. Dashboard audit tooling will score pages via a Worker /api/audit endpoint so browsers never hit third-party CORS walls.