heddle check --only helpful
Helpful Content Check
The helpful check asks a model to judge each page whole against Google's own questions for helpful, people-first content, and fails a page below the site's floor. Without it, pages that pass every rule can still be the kind Google's scaled-content policy names, written to rank rather than to help. Helpful content. Every indexable page judged whole against Google's own self-assessment ("Creating helpful, reliable, people-first content"), by the judging model.
When Heddle is set up on your site, this check runs on every build alongside 25 others, so a page that fails it is fixed before a buyer or a search engine sees it.
Why it exists
Helpful content. Every indexable page judged whole against Google's own self-assessment ("Creating helpful, reliable, people-first content"), by the judging model. Ported from the source site's check-helpful script (10 Oct 2026), where several agents publish without a person in the loop and this is the measure they are held to.
Google publishes no score for helpfulness, and its spam policy on scaled content abuse names pages generated to rank instead of to help. A site that generates pages needs a gate on exactly that.
Ten questions, each phrased so that yes is the helpful answer, so every threshold is a floor. Two kinds of page carry their own floors. An article (a post, a comparison, a definition) is written to teach or show work, and a page is written to sell or route.
The default floors are the source site's, measured halfway between the best score any deliberately unhelpful fixture reached and the 5th percentile of a good set; bookmark and accurate did not separate the sets, so they are reported and left ungated. An install overrides them in site.json helpful.floors once it has calibrated its own.
One reading of the same text moves by up to about 0.06, so a reading within 0.05 of a floor is read twice more and the mean of three counts, and a page that would fail is read three more times before the failure stands.
The baseline (heddle/helpful-baseline.json) lists pages already under a floor when the gate arrived. Such a page passes while it reads no more than its allowance (0.12, at most half the recorded score) under what was recorded; a new miss, or a baselined page made worse, fails.
The baseline only tightens: heddle check --update keeps the higher score and drops what now passes. Adding pages to it needs HEDDLE_HELPFUL_GROW=1 on that one run, and a reason in the pull request; it is how a gate is adopted on a site that is already live. New copy gets through by passing.
Known blind spot (source site, 9 Oct 2026). A draft that retells a vendor announcement passed every floor. Whether an explainer adds enough to its sources stays a person's call.
From lib/checks/helpful.mjs in the Heddle repository at ddb1b16.
Questions
What does the helpful check look for?
does each page meet Google's helpful-content self-assessment, with a baseline that only tightens?
Does the helpful check need an AI model?
Yes. It asks a model hosted on Cloudflare Workers AI and caches each answer. Without the model it reports SKIPPED, never PASS, because a gate that silently stops judging is how unreviewed copy ships.
How do I run it?
Build the site, then run node bin/heddle.mjs check <site> --only helpful from the Heddle repository.
Work with us
Let the agents bring you the leads.
Tell us about your business and who buys from you. On a short call we'll show you the searches your buyers make where you don't show up yet, and what the agents would do about them in their first month.
On the call, for your site
- The searches your buyers makeMeasured, with how many people make each one
- Where you show up, and where you don'tOn Google and in ChatGPT's answers
- The agents' first monthThe pages, fixes and links they would start with