Findings & actions

Overview

Findings & actions (Insights › Findings & actions) is the page that answers what did the last cycle change, and what should we do about it?

A finding is a standing condition the intelligence pass measured against your own captured answers — "3 of 12 non-brand transactional questions had nobody named", "25% of 160 citations of your pages came in answers that did not name you". It stays open while it is true and closes on its own test. Findings are computed over each project's own capture cycle, so this page does not follow a date range: a finding re-sliced by a picker would show a number that was never measured.

The findings feed

The page leads with the feed. A filter row narrows it:

• Status — Open (the default: new and seen), Seen, Actioned, Dismissed, Closed.
• Type — the finding types (Nobody Named, Cited Not Named, Rival In Your Answer, Fact Stated Wrongly, …).
• Question type — one of the four question types. This is the same control as the question-type chip in the filter bar; a finding carries its question type but no reference to the individual questions behind it, so the other chips (segment, intent, journey, topic, tag, group) cannot be applied here — when one is on, the feed says so instead of showing an unfiltered list.
• Stage — where in answer assembly it fails (Stage 1 shortlist, Stage 2 filtered out, Stage 3 name dropped or described wrongly).
• Kind of work — Outside work, Content, or Not classified — tag on the edit page.
• Rank — the next-action order's groups (Fix first: answers about you · Mentions to strengthen · Unclaimed questions · Score work where the gate passes · Outside work: sites to get onto · Housekeeping). The feed stays newest first; the rank only narrows it.

Every filter is applied by the database, so the count shown is the whole population, not the page in front of you.

Above the list, one tile per question type shows how many findings are standing there, the single next action for that question type, and the rank it comes from. The same next action is the line under each question type on Home, and the Action list holds the whole order at its top.

The next action follows the method's order of work: wrong or lost answers about you first, then mentions to strengthen, then questions nobody has claimed, then work on your score, then sites to get onto. Among the mentions to strengthen, your pages used while a competitor you track was named instead come first. Within that order the more serious finding comes first, and the most recent only breaks a tie. A finding resting on fewer than 10 answers is never chosen, nor one counted in citations with fewer than 10 citations behind it, nor housekeeping (competitors you don't track, or tracked ones never named). Two kinds of finding the method does not rank (a page of yours cited less often, a competitor gaining) are placed by us, and their (i) says so. Since 26 Oct 2026 the tile can therefore name an older action than it did before, when that action ranks higher. Since 28 Oct 2026 it is chosen from every standing finding (new, seen, or already made into an action), not only from findings that have become actions: a finding not yet on your list can be the next action, and its link opens its place on the Action list, where you can add it.

A card that cannot be the next action says why, in plain words: Fewer than 10 answers, Fewer than 10 citations behind it, Housekeeping: never the next action, Not in the order yet, or, once the gates are checked, Waits until gate 1 passes (or gate 2 or 3). Its (i) gives the reason in full.

Each finding is one card:

• the strip — question type › stage › kind of work · the n it rests on, banded (low confidence under 10, below target at 10–29, at target from 30) · how it is known (measured directly, a rule of the method, a pattern not yet separable from coincidence, a forecast)
• the title and the measure sentence exactly as written when it was measured
• the test that closes it, in your own capture rhythm ("Closes when this falls to 12% or below within the next fortnight")
• the action it has already been raised as, linking to that action

Cited, not named — counted in citations from 26 Oct 2026. This finding now counts citations of your pages, not answers: an answer that cites three of your pages without naming you counts three. A figure seen before that date counted answers, so it reads differently, and the same applies to "Cited without being named" on Problem diagnosis. It is raised on Non-brand · Transactional questions when at least 2% of your citations came in answers that did not name you. On Non-brand · Informational questions, where being cited is itself the win, it is raised only when at least 2% came in answers that named a competitor you track instead. It is never raised on questions about your brand. It closes when the share halves. A finding raised under the earlier definition stays as it was until your project is next measured, after its next capture; it then closes, keeps what it said, and is replaced wherever the new definition holds.

Unclaimed questions — confirmed on each engine's last 7 runs from 28 Oct 2026. A question is now read on each engine as it runs now, on that engine's 7 most recent answers to it, whatever their dates (up to a year back). It is unclaimed there when none of the 7 named a company (transactional questions) or cited a source (informational questions); it is confirmed with 7, and with fewer it is observed, not yet confirmed. An engine is one engine configuration as it is set now (two models from one provider are two engines) or one search surface, and an answer counts toward it when it was requested from the model the engine runs now: answers from a model the project no longer runs are not counted. An engine set to a moving alias (a model id ending in "-latest") is one engine on that id, even when the provider changed the model behind it. The finding counts the questions confirmed unclaimed on at least one engine whose answers cited sources (open ground); questions unclaimed with nothing cited (answered from memory, outside the contest) and questions only observed are told in its detail and not counted, so on informational questions it is not raised. It closes when a company is named in one of these questions. On Performance, each question under Nobody is named on these questions carries the same reading per engine, with the date of the nightly findings run that read it; it does not follow the period picked. A finding raised under the earlier definition stays as it was until your project is next measured, after its next capture; it then closes, keeps what it said, and is replaced wherever the new definition holds.

Kind of work — read from the context map from 30 Oct 2026. For unclaimed questions and a competitor gaining, the kind of work now follows the kind of context the questions are about on your context map. A question tagged with an offering (what you sell, from the industry down to the item) is Outside work: credible independent sites have to carry it. A question tagged with a context (who it is for, what for, where, at what price, under which specification, against which certification) is Content. A question tagged with both counts as outside work. A question with neither is Not classified — tag on the edit page: the kind of work cannot be read from it, so its card asks you to tag its questions on their edit pages (What this question is about) and does not become an action. The card's link opens the Query list on the active questions of that type with no offering or context tagged, which can be more questions than the finding names. Unclaimed questions are filed as one finding per kind. Before this date the kind was read from whether a page of yours was cited on that question type. When you tag questions, the next nightly findings run reads the tags, even if the project has not captured since; a finding whose questions moved to another kind closes and records where they went. Every other finding keeps the kind its type means: getting onto the sites that cover a topic is outside work, and correcting what is said about you, attaching claims and improving the pages you own are content.

When nothing is standing, the page says so: "Nothing is standing. New findings are raised after each capture cycle."

Lost questions about you — from 30 Oct 2026. A question about your brand (one that does not name a competitor) is lost on an engine when that engine's answers to it carry three or more accuracy verdicts and most of them got you wrong: not accurate, or carrying an issue against your brand facts. A check that found nothing it could verify is no verdict and is not counted. The finding, Lost Question About You, sits under Fix first: answers about you. It counts these questions out of the questions about you with three or more checked answers on an engine, names up to three of them with the engine and its count ("on Gemini, 4 of the 5 answers checked since 12 Oct 2026 got you wrong"), and closes when none is left. It is read per engine, so it can differ from Problem diagnosis, which reads each question over all of its checked answers together. The method's fuller close, right on four or more engines for two cycles in a row, is told in the finding and not applied. With no brand facts on file nothing is checked, so the finding is not raised.

Problem diagnosis — the four stages

Linked to the category heads the page. Someone who can edit the project chooses five of the non-brand questions you already track that ask about your category in general: no brand, no price, no location beyond your market. It adds no questions and costs nothing, the page says who chose them and when, and you can choose again. It then shows runs per question across the engines, for example "Named in 2 of 12 answers, about 3 an engine", and your tracked rivals on the same answers. That is an observation, not a reading: "3 of 5 questions" is the Gate check's count, at 20 runs a question, and it isn't printed before one has run. Until five are chosen, or if one of them is later paused or deleted, it can't be checked.

On the shortlist comes next. It counts how often each engine named you on your most-answered non-brand transactional questions, up to 20, over the page's period, for example "Named in 3 of 60 answers, about 3 a question". That is an observation, not a reading: it comes from your regular captures, at whatever depth they run, so it shows answers and never a percentage or a pass. The questions are suggested until someone who can edit the project confirms them: choose 10 to 20 from the ones you already track. It adds no questions and costs nothing, the page says who confirmed them and when, and you can choose again. With fewer than 10 that have answers it can't be checked. A reading comes from a Gate check at 100 runs an engine, which isn't run yet.

The Problem diagnosis page (Insights › Problem diagnosis) reads every view down to one named problem per question type. An answer is built in three stages — the engine assembles a shortlist, filters it, then writes the answer — and each symptom points at one stage and one kind of work:

Where each question type is losing — read per question from 30 Oct 2026. Each contested question is sorted into exactly one pattern by what most of its answers did, so a question type reads on one count of questions:

• Non-brand · Transactional — not named, your site not cited (Stage 1); not named, your site cited; named, not in the first three; named, described wrongly (all Stage 3); named, position not readable; named in the first three.
• Non-brand · Informational — not cited, not named (Stage 1); named, your site not cited (Stage 3); your site cited.
• Questions about you — not named (Stage 1); described wrongly (Stage 3); described correctly. A question named in its answers with no accuracy verdict (never checked, or checked with nothing the check could verify) is shown as "Named, no accuracy verdict" and never counted as correct. With no facts marked approved, nothing can check it, and the row says "No facts marked approved to check against", with a link to your facts; facts added between captures are read from the next capture.
• Questions naming a competitor — on their own, thinner bar ("Questions about a competitor"): not named, named but not in the first three, named, position not readable, named in the first three. They never enter the count of questions about you.

A named answer with no position to read. Being named says nothing about position when the answer gives none, so that answer is left out of the first-three test rather than counted against it. An answer that doesn't name you stays in, because you aren't in its first three. A question where every answer naming you has no position reads "Named, position not readable", which isn't a losing pattern. The Top-3 presence card says how many answers its figure rests on whenever any were left out.

A step in November 2026. Before then, a position on Google AI Overview or AI Mode was the line your brand first appeared on, which isn't a rank. Since then it comes from the same reading as a chat answer: the order companies are named in the AI answer's own text. First-three figures on search answers may step at a project's first capture after the change, and an answer captured before it keeps what it stored. The Visibility score may step with them, in either direction and by up to about three points where search answers are a large share: a position read from prose weighs less than an unmarked one did, and an answer that now has a position weighs more than one without.

The reading is the losing pattern that holds the most questions, called only when its own share is past the six-point noise floor. Two patterns within the floor of each other both apply. Losing patterns that pass the floor only together say "no single pattern is past the floor", never "clear". Fewer than 5 contested questions are too few to read as a rate. Under the reading, a line says when one engine alone differs past the floor from the others, or that the engines differ; an engine with fewer than 10 contested questions of its own is listed, not compared.

The verdict strip at the top, this section and the line under each question type on Home read the same reading: the nightly findings run writes it over the window your pages open on, with its date. It does not follow the date picker; the strip's next action still comes from the findings standing. Before this date the section compared two shares — not named against named but not in the first three — over the picked window. Figures counted in citations or runs, such as the citations of your pages in answers that did not name you, are printed beside the reading with their own denominators and never enter it. Stage 2 is reported in words as not yet distinguishable from Stage 1 until your page's search rank for the same question (Search Console) can separate them.

Every count is printed with its population ("N of M contested questions"), question types are never added together, and the page composes figures the Performance review, Citation sources, Content coverage gaps, Website readiness check and Fact check already compute — it computes nothing new. The outside-work-or-content card at the bottom sorts the open gaps by the budget that closes them: outside work is binary, and two or three credible independent sites carrying it is enough; content is continuous and routes to the content plan.

Actions

The actions themselves live on the Action list. This page shows one line — how many are open, in progress and completed — and the door to the list. That line counts the whole project and carries an "All prompts" badge when a filter chip is active.

Verified impact and what works best

Two cards under the actions line:

• Verified impact — completed actions with a verification verdict. Today's verification is a project-wide before/after on the retired blended score; the change in points on the action's own question type arrives when the per-pool verification ships.
• What works best — success rate per kind of action, from the actions that were verified. It fills as you verify more; treat it as directional until each kind has a handful of verified rows.

Model changes

Shown only when there are any: brand-mention shifts around the dates a new model was first seen on an engine. The delta printed beside each date is an observation, not a claim that the model change moved the number.

Actions lifecycle

1. Detect — the intelligence pass raises a finding with its measure and its stop condition.
2. Raise — a finding becomes an action (automatically for the finding types that map to one; by hand for the rest), and lands on the Action list.
3. Assign — open the action from the list (one hop) and set its owner and due date on the action's record page.
4. Execute — work the checklist; mark it In progress, then Completed.
5. Verify — after completion the platform compares a project-wide figure before and after — today a check on the retired blended score, kept until the measurement on the action's own question type ships. Verified and Failed verification are set by that check, never by changing a status.
6. Learn — verified outcomes feed the "What works best" table.

How findings are generated

Findings are written by the intelligence pass that runs after each capture cycle, over the project's own captured answers. There is no setup: the more cycles a project has, the more its findings can say. Each finding closes when its own test is met, and a closed finding is kept under the Closed filter so you can see what stopped being true and when.