Backbrief Site Check

Site Check sample: backbrief.ai

Dated 2026-09-29. Read from the live site between about 23:00 and 23:20 UTC. Made by hand, on the owner's own site.

This is a sample of what Backbrief Site Check produces. It was made by hand with the discipline the base uses: every finding carries its URL, the raw evidence and a way to reproduce it, every claim on the page is tagged verified, unverified or vendor claim, and a check that could not run says NOT CHECKED instead of passing. The site is the owner's own, backbrief.ai. The checker never makes the fix, so this report names problems and leaves the changes to the person who builds the site. AI helped produce this report.

Verdicts

Verdicts use three words: pass, fix, not checked. There is no score.

How to read the evidence

• Every finding gives the URL, the raw text returned, and a command to reproduce it. Commands use curl, which the site answers.
• Sources for claims are the base product repository at main, commit c1728307ed43, read 2026-09-29, which was public on that date and is private as of 2026-10-05, and the site's own legal pages. Tags in the ledger mean:
  • verified: the sentence on the page matches a named source, read on the date above. This shows the page agrees with the owner's own files. It does not show the claim is true in the world.
  • unverified: plausible, and no source was checked. Unchecked, not false.
  • vendor claim: the source is the seller of the thing described. Every source here is the owner's own, so the owner's product statements are self-description and the ledger says so where it matters.
• Legal statements on the site (governing law, statutory rights) are unverified local-law assumptions. This check did not research them.

1. Slop

Verdict: fix. The prose check applies the stop-slop rules to the live text of the pages. The visual check (layout, type, imagery) is NOT CHECKED: it needs a rendered page and none was rendered.

Passes, with evidence:

• No em dash or en dash on any page read, on /llms.txt or /pricing.md. Reproduce: curl -s https://backbrief.ai/backbrief | grep -c -P '\x{2014}|\x{2013}' returns 0.
• No -ly adverb in the /backbrief body text (scanned by pattern, then read).
• No "here's what" openers, no "the truth is", no jargon from the banned list.

Lines that fail, each on https://backbrief.ai/backbrief unless noted:

Reproduce: curl -s https://backbrief.ai/backbrief, then read the visible text.

2. Search eligibility

Verdict: fix. Read on free data only: direct fetches of the pages and machine files. No paid tool was used.

3. Structured data

Verdict: pass, structure only. There is no validator in this run. "Structure only" means each block was pulled from the page and parsed as JSON, and its fields were compared with the visible text. Validation in a rich results tool is NOT CHECKED and belongs after every deploy.

Reproduce: curl -s https://backbrief.ai/backbrief | grep -o '<script type="application/ld+json">[^<]*'.

Observations, none blocking:

• The logo is og.png, a 1200 by 630 social card, not a square mark. Validation may warn.
• The Product block has no image and no url, and the Offer has no availability. None is a parse error; a rich results tool decides what it accepts.
• sameAs lists an X profile that no visible page links to. The visible-content rule asks the owner to confirm the profile belongs to the entity.
• No block describes content the page does not show. No review, rating or FAQ markup exists, and none is displayed.

4. AI-answer baseline

Verdict: not checked. This section is a single-run record. It supports no conclusion about whether AI assistants mention or cite backbrief.ai, because a verdict needs repeats across runs and a fixed prompt panel. These queries were written by hand for this sample, not drawn from query research, so they are not a panel.

• Engine: the web search tool available to Claude Code. The underlying engine and model are not disclosed by the tool.
• Date: 2026-09-29. Market: US only. Each query ran once.
• Volatility note: results from a single run on a system that changes by model version, location and sampling. Do not compare this run with a later one unless the same queries are run again under the same label.

What the record shows: in these responses on this date, backbrief.ai appeared as neither a name nor a source. What it does not show: how often that would repeat, what any other engine returns, or why. The repeat runs, the fixed competitor set and the panel from query research are NOT CHECKED here because they were not run.

5. Page claims

Verdict: fix. This is the claim ledger for the live page https://backbrief.ai/backbrief. Quotes are from the page as served on 2026-09-29. Repo paths are relative to the base product repository. Abbreviations: DOCS = docs/README.md, LIC = LICENSE.md, GR = payload/.claude/rules/grading.md, GRADE = payload/.claude/commands/grade.md, GV = payload/.claude/commands/grade-video.md, SKILL = payload/.claude/skills/video-claim-grading/SKILL.md, TN = examples/trailnotes-rerecord-2026-09-29.md, WE = payload/scaffold/examples/worked-example-business.md, REC = examples/grade-video-run-record-2026-09-29.md, IDX = payload/.claude/CLAUDE.md.

The promise in the headline, "Get the second opinion before the work costs money.", is positioning and not a checkable claim, so it has no row.

What the product is and does

How the grade works

The proof

Setup, scope and terms

Claims on the other pages that the ledger touched

Fix list

Ordered by what each fix prevents. Whoever builds the site makes the changes; the owner decides the ones marked owner decision, because they change what is public.

1. Prevents the site contradicting its own promise that paid files are not free. Owner decision. The public repository https://github.com/charlesashe/backbrief-base is "private": false and carries the product's agents, commands, rules and skills under payload/.claude/, while LICENSE.md lines 19 to 22 bar redistribution and /terms says "Nothing in the line is free." The repo's own description still reads "Private until release." Decide whether the payload should be public, and if not, change visibility. Validate by re-reading the repo page and the API field private. RESOLVED 2026-10-05: the repository was made private on the owner's instruction. Re-validated the same day by the step this row names, the API field private reads true and an anonymous read of the repo page returns 404.
2. Prevents a retired free download from undercutting the price. Owner decision. https://backbrief.ai/agent-os-kit.zip, /getting-started-with-claude.pdf and /start-simple-with-ai-checklist.pdf return 200. Remove the files, or change the site copy to say the retired kit stays downloadable. Robots rules do not remove a file. Validate: curl -sSI on each returns 404. Removing files is destructive: keep a rollback copy first.
3. Prevents a stated withholding from being false. Owner decision. The site says no checker price is published (/llms.txt, /pricing.md, /site-check). PRODUCT.md line 130 in the public repo quotes a checker price. Remove the number from the public file, or change the site line. Validate by reading PRODUCT.md after the change.
4. Prevents a buyer misreading what the product does. On /backbrief, P38: "the doing stays yours" contradicts approve.md line 10, where approved units go to the builder, researcher and runner agents. Rewrite so the owner keeps outward action (publishing, sending, spending), which is what escalation.md says. P10: add that on regrade passes an advisor also gets its own earlier scores.
5. Prevents an unsourced claim about a third party. P30, the footer on every page. Replace with a sentence about Backbrief that the owner can verify, for example that Backbrief earns no commission from Claude Code.
6. Prevents the page overclaiming its own video result. P16 and S3. Remove "the grade that comes back when the claims do not hold up". The page's own next sentence says an unverified claim is unchecked, not false. Consider saying the record is a small set of videos, without a count, and check that the grader gaps listed in examples/grade-video-run-record-2026-09-29.md lines 29 to 40 are fixed before "Every one graded C" stays.
7. Prevents stale public documents. Owner decision. The repo root README.md line 10 says "Status: payload assembled, unreleased. Version 0.1.0 is reserved for the first release." The file VERSION reads 0.1.0 and the site calls it released. The same README also names a local folder path on the owner's machine (line 8). Update both.
8. Prevents weak link previews and a stale freshness signal. Give each page its own og:title, og:description and twitter:title, and add og:url. Set sitemap <lastmod> per page from the page's own change date, not the build time. Validate by re-fetching /sitemap.xml and comparing entries.
9. Prevents prose that reads as generated. S1 to S6. Rewrite S1 with a named actor; state S3 without the pull-quote; remove the repeated footer line on /backbrief. For S2 and S6, record the exception in the project's context notes if the phrasing is the owner's voice.
10. Prevents unrun checks from passing by silence. Run the checks marked NOT CHECKED, then re-read this report: PageSpeed Insights on /backbrief (after the daily quota resets), Search Console indexation and the generative-AI features setting, the rich results test on /backbrief after each deploy, the checkout amount by the owner opening the buy link, a rendered visual pass, and repeated AI-answer runs under a fixed panel.

This is a stress-test, not a certification. A pass is not a guarantee. An unverified claim is unchecked, not false.

Get Backbrief, the base

Write to contact@ashecorp.com with the subject "Site Check" to be told when it is available.