Recommendations
A work queue derived from measured findings — each card says what was found, where the fix happens, what to do, and what it was counted in.
Recommendations is the prioritized work queue. Each card is one move, argued from measured evidence and broken into steps you can assign.
Nothing here was thought up. The list is derived in SQL from findings that were counted in your collected answers; a model writes the wording and is not allowed to add items. Every card shows the count it rests on — seen in N of M answers — so you always know how much evidence a card carries.
The four signals
| Signal | What was measured | Where the fix happens |
|---|---|---|
| Wrong information | An engine stated something false about you, with the sentence it wrote as evidence. | Your site + the sources engines read. State the correct fact on pages you control first, then fix external pages that actually carry the stale version. |
| Repeated criticism | A criticism consistent across two or more answers. | Your site and your product — published evidence where it is unfair, a real fix where it is fair. |
| Coverage gap | A publication cited for your rivals and never for you. | Off-site: the publication itself. Nothing on your own pages closes it. |
| Lost question | A prompt where competitors get named and you do not. | Your site — the page that makes you a defensible answer to that question. |
The where column answers the question every suggested fix gets asked first: is this work on my site, or somewhere out in the world? The card's detail view states it above the steps. If a step names an external property you do not have — a Wikipedia article, a directory profile — that step does not apply; correct the fact where you can.
Anatomy of a card
Title — the move as an instruction: "Correct the founder count on the About page."
What we found — the observation with the engine's verbatim sentence where there is one, and the correction (what is actually true) for misconceptions.
Where + Suggested fix — where the work happens, then 2–4 concrete steps.
Expected impact — the mechanism the work acts on, never a promised number, plus the measured count: seen in N of M answers.
Impact / Effort — impact is derived from reach against the corpus (misconceptions are always high: a factual error is the most fixable thing on the tab); effort is a property of the class of work.
Working the queue
Cards move through Inbox → Planned → Done, with an Archive for dismissals and cards whose signal stopped being measured.
- Plan what you commit to; planned cards keep their wording even as the measured numbers refresh.
- Done is permanent — a completed card is never resurrected.
- Dismiss with a reason (not relevant, cannot do now, disagree). A dismissal survives every future generation run; regenerating never re-proposes it.
- When a signal stops being measured — the misconception stops being repeated, the publication starts citing you — the pipeline archives the card as no longer measured, and reopens it with its history if the signal returns.
Prioritizing
Filter by impact, then order impact against effort:
- High impact, Quick — do these immediately.
- High impact, Long haul — plan these. One per cycle, resourced properly.
- Medium impact, Quick — good filler when the team has capacity.
- Low impact, Long haul — dismiss without guilt.
Closing the loop
After shipping a fix:
- Re-analyze the affected page in Pages.
- Give the engines a cycle or two to re-crawl and re-answer.
- Check whether the affected answers moved — misconception cards are re-checked against the next runs automatically.
- Watch the card's seen in N answers fall, and let the pipeline archive it when the signal dies.
A recommendation you shipped but never measured taught you nothing.