mirror of
https://github.com/superdesigndev/treg.git
synced 2026-10-02 03:24:35 +08:00
Jason ZhouandClaude Opus 5
6d92d9341d
treg.to had six thin marketing pages and 41 organic clicks a month. This adds
the first slice of a programmatic set built on the one thing nobody else can
publish: live per-call prices and measured reliability for every provider that
does a job.
What ships:
/agents/<client> ChatGPT, Claude, Claude Code, Cursor. Install steps for
that client, then the whole catalog as a menu of jobs by
buyer category, priced from the catalog.
/use-cases the crawlable hub the job pages hang from.
/use-cases/<cat>/<job> seven job pages: the prompt that works, why it works,
then what the agent sees before it calls (cheapest,
most reliable, how the providers compare).
<path>.md a plain-Markdown mirror of every page, declared with
rel=alternate, for agents and answer engines.
The page is a template, not prose. Everything job-specific lives in
`agent_pages.USE_CASE_PAGES`; the route picks one of three FORMS from the
catalog itself:
short one provider, nothing to compare (19 of the 66 jobs)
platforms the job spans platforms, which are not alternatives (19 of 66)
compare several providers, one platform
Two correctness rules earned their own tests. Cheapest is claimed PER BILLING
UNIT, never overall: 38 of the 66 jobs mix per-call, per-result and per-success
endpoints, and ranking those together names the wrong winner, because one call
returning a thousand rows is not dearer than one row. And providers are keyed by
(provider, platform), not provider: ScrapeCreators serves Instagram and YouTube
for the same job, and collapsing on provider alone silently dropped a whole
platform from a multi-platform page.
Honesty constraints, all tested: no page states a number the catalog did not
produce; every capability id named in a page must exist; per-vendor reliability
renders only when there is traffic and always carries the "live traffic, not a
controlled benchmark" caveat; a boxed methodology says how each figure is made;
the pages 404 off the reference hosts, because "search treg in ChatGPT's Plugins
directory" is false on a self-hosted registry.
Also here: `.agents/skills/treg-page`, which makes the research that precedes a
page repeatable, including how to spot the vendor astroturf that dominates these
searches; and `marketing/` records the measurement, the plan, an independent
review and a risk audit against the SEO playbook.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Gfwm9hzVd8JBQ73YycJpah
4.3 KiB
4.3 KiB
Risk audit of the programmatic pages — /seo-playbook, 2026-08-21
Run against the kill list and extraction rules in .claude/skills/seo-playbook (Lily Ray on what
Google punishes, Kevin Indig on how LLMs select, Fishkin/Natividad on the zero-click response).
Subject: 4 agent pages, 7 use-case pages, 1 hub, plus the 66-job menu they are the first slice of.
Kill list: no hits, and two near-misses worth naming
| # | Tactic | Verdict |
|---|---|---|
| 1 | Self-promotional "best X" ranking our own brand first | Clear. The comparison ranks providers; treg.to is the access layer and is not one of the ranked items. This also sidesteps Ray's law 5 (69% of self-promotional listicle citations ended with the citing brand excluded from the recommendation) |
| 5 | Comparison / "X alternatives" at scale | Near-miss, structurally avoided. The banned shape is one page per competitor pairing. Ours is one page per job with every provider on it, so nine providers produce one page, not 36 pairings. Keep it that way |
| 6 | FAQ farms | Clear. Four questions inside a page, never one question per URL |
| 3 / 7 | Templated scaling with minimal uniqueness | Near-miss, and the real risk for waves 2 and 3. 12 pages with hand-written notes and quoted research is fine; 66 generated from the same shell would be the fingerprint Ray describes. The treg-page skill exists to keep the per-page research mandatory |
| 4 | Artificial freshness | Clear. Dates come from the catalog's own verification stamp, so nothing bumps without a real re-check |
| 10 | Schema misuse | Was a hit, now fixed. SoftwareApplication carried Offer price: "0", which reads as "the product is free" while the page says installing is free and calls are metered. The offer now says exactly that |
| 8, 9, 11 | Hidden instructions, bought consensus, one page per query | Clear |
Extraction rules (Indig)
- Answer in the first 30% — passes. The provider count, the from-price, the $0.000 markup and the named cheapest provider all appear before the 30% mark; the economics block sits above the fold on desktop.
- 15+ unique data points — passes comfortably (per-provider price, billing unit, accepted inputs, verified date, success rate and latency where measured). The bar Indig cites is that the top ten average only about four.
- Methodology boxed — was missing, now added. Every comparison page carries a four-box block: how prices are derived, what the success rate counts (2xx vs 5xx, 4xx excluded), what it is not (not a controlled benchmark), and what "verified" means.
- Visible date — was missing, now added to the hero subline ("last checked 2026-08-20").
- Named author — still missing. Indig's data says named authors outperform brand bylines. This needs a person's name and is Jason's call, not something to invent.
- Focused beats ultimate guide, but cover the intents inside the page — passes: one job per page, with the prompt, the comparison, the caveats and the FAQ all inside it.
Where the pages are weakest, in order
- No off-site presence. Nothing in this repo fixes it. Indig and Ray converge here: authority is
per-topic, three placements in sources the models already cite beat a dozen scattered mentions, and
nofollow counts for AI mentions. The directory submissions in
pseo-ship-plan.md(mcpservers.org, PulseMCP, Glama, Smithery, Docker MCP, Anthropic's connector directory) are the highest-leverage remaining work, and they are Jason's to send. - No AI-visibility measurement. Indig's method is polling, not rank tracking: ~40 seed prompts, five runs per platform per week, per-platform scores with confidence intervals, never one blended score. Nothing like this exists yet. Ray's caveat applies: treat it as directional.
- Scaled-content risk lives in waves 2 and 3, not in what shipped. The guard is the research step, and it is only a guard while someone actually runs it.
Kept deliberately, against a rule
- Traffic is not the KPI. Fishkin's decoupling argument and the wiki agree: the verdict metric is
repeated successful calls per landing page, which still needs
first_call_atattribution. - Comparison pages at all. Ray's caution is about scale and self-promotion. These are neither, and the first-party price and reliability data is the thing no competitor can regenerate.