Gaps, turned into work someone can actually pick up.
A visibility number tells you there is a problem. A recommendation tells you which page, which topic, which competitor and which prompts — and shows the observations that put it on the list, so the person doing the work can disagree with it.
- Always evidence-linked
- Every item opens onto the prompts, answers and citations that generated it.
- Rated on three axes
- Impact, effort and confidence, so a backlog can be ordered without a meeting.
- Human in the loop
- Briefs are drafts to verify. AnswerRank does not publish anything to your site.
What a recommendation actually contains.
A recommendation that cannot show its working is an opinion. Every item is generated from observations in your own data and carries them on its face.
Publish an Acme vs Northwind comparison page
- 12 comparison prompts where Northwind is named and Acme is not
- 5 answers citing northwind.com/compare
- No owned page addressing the comparison topic
- Observed across 4 engines and 3 scan runs
Evidence suggests the absence of an owned comparison page is a likely contributing factor in how these prompts are currently answered.
Every item answers five questions
- What is the change, stated as a specific piece of work rather than a theme.
- Which observations produced it, listed and clickable down to the individual answers.
- How big the observed gap is, how much work it looks like, and how solid the underlying sample is.
- Which tracked prompts it is expected to touch, so the follow-up measurement is defined before you start.
- What it does not claim — every item carries its own limitation line.
Recommendations are grouped so a single root cause does not generate fourteen near-identical tickets. Closing the parent closes the set.
Impact, effort and confidence — and what each one really means.
Three ratings, each derived from something specific. The distinction that matters most: confidence describes the observation, never the outcome.
| Rating | Scale | Derived from | What it is not |
|---|---|---|---|
| Impact | Low · Medium · High | How many tracked prompts the gap touches, weighted by their buying intent, and how consistently the pattern recurs across engines. | A description of the size of the observed gap, not a prediction of the size of the gain. |
| Effort | Low · Medium · High | The shape of the work: editing an existing page, publishing a new one, changing rendering behaviour, or pursuing earned coverage over months. | An estimate for a typical team. You know your own constraints better than we do, and effort can be overridden. |
| Confidence | Low · Medium · High | Sample size behind the observation, agreement across engines, and time spread across scan runs — the same assessment used for every metric in the product. | Confidence in the observation. Never confidence that the change will work, which nobody can measure in advance. |
There is no composite priority number that hides the three inputs behind a single figure. Teams weigh effort differently, and a score that pretends otherwise just moves the judgement somewhere you cannot see it.
A recommendation has a state, and the state has consequences.
Status is not decoration. It decides what reappears, what gets merged, and where the boundary between before and after sits when you decide to measure.
Open
open
Generated from an observation and waiting on a decision. Duplicates of a still-open item are merged rather than stacked.
In progress
in_progress
Someone owns it. The evidence snapshot is frozen at this point, so you can see what the case looked like when the work started.
Implemented
implemented
Marked done with a date, which becomes the natural boundary between before and after if you run an experiment on it.
Dismissed
dismissed
Rejected with a reason that stays on the record. A recommendation you have already considered and declined should not reappear as if it were new.
A brief is a draft for a person, not a publish button.
Where a recommendation calls for content, AnswerRank assembles a brief from the observations behind it — the questions, the competitors, the sources, the structure. Then it stops.
- The question the page has to answer, taken from the tracked prompts it is meant to serve
- The competitors currently named in those answers, and the claims made about them
- The sources those answers cite, so the brief targets material an engine actually reads
- Suggested structure: the sections, the comparison points, the specifics worth stating plainly
- A list of every factual claim that needs verification before publication, flagged as unverified
Why a human has to verify
A brief is generated from observed answers and your own site. Both can contain claims that are outdated, imprecise, or simply wrong about your product — and a language model assembling them can produce a confident sentence about a feature you do not ship.
So every factual claim in a brief is flagged as unverified until a person confirms it. AnswerRank does not auto-publish to a CMS. It is a deliberate product boundary, not a missing integration.
The same applies to competitor claims. Repeating something an AI answer said about a rival, without checking it, is a legal and reputational exposure we are not going to hand you inside a one-click flow.
The sentences we allow ourselves to write.
Recommendation copy is generated from templates, and those templates are linted. The rule is simple: we report association and require a re-test, and we never promise an outcome.
| We write | We never write |
|---|---|
| AllowedEvidence suggests this gap is associated with lower visibility on comparison prompts. | BannedThis gap is why you rank below Northwind. |
| AllowedPublishing this page is a likely contributing factor in how these prompts are answered. | BannedPublishing this page will increase your visibility by 20%. |
| AllowedAcross 14 valid checks, answers citing this source named two competitors and not you. | BannedThis source is a confirmed ranking signal for AI answers. |
| AllowedRe-run the frozen prompt set after the change to see whether anything moved. | BannedYour AnswerRank Score is certain to rise once this ships. |
The forbidden column is not a style preference. Those phrases are blocked in code, and a test fails if one of them reappears in a template.
A recommendation is a hypothesis until you test it.
Because each item already names the prompts it is meant to affect, promoting it into an experiment does not require you to define anything twice.
- 01
Baseline is captured
The prompt set named by the recommendation is frozen and measured before the work starts.
- 02
You ship the change
Marking the item implemented records the date, which becomes the boundary between the two sides.
- 03
The same checks re-run
Identical prompts, engines and market. The delta is reported with its sample size, or withheld if the sample is too thin.
Find out what to fix first.
One free check, and a prioritised list with the evidence attached.
No signup required for your first visibility check.
Every figure AnswerRank reports carries its sample size and how it was collected.
Read the methodology