Strategy
Part of A social listening strategy that begins with a decision and ends with a review gate
Social listening ninety day plan: stop rules for every stage
A ninety day social listening plan for England: a stop rule for each stage, the sources and tools to test, and a proceed, revise or close record.
A ninety day social listening plan ends in one of three records: proceed, revise or close. The call is only as good as the stop rules written into the brief and the evidence pack assembled before the decision. Both are set early and read late.
The plan tests one bounded method. It does not target sentiment, reach or return on investment. For England-only analysis, include a record only when its author explicitly self-identifies as being in England, and mark UK-wide, ambiguous and unknown locations unassigned.
What to take away
- Write a stop rule for every stage into the brief before any data arrives.
- Stop if no finding could affect the choice or if qualified review rejects the use.
- Set sample size, minimum precision and the maximum tolerable error rate before testing, not after.
- Build an evidence pack that gives source, dates, query version, numerator, denominator, missing data and inference limit for every count.
- Record proceed, revise or close with evidence, rejected options and a next review date.
Fix the stop rules before any data arrives
Every stage carries a stop rule, and the brief approved by day 15 holds them. Stop if no finding could affect the choice or if qualified review rejects the use. A rule written after the data arrives only justifies a conclusion already reached.
Days 1-15: Approve the decision
- Decision owner supplies choice, deadline, population, risk tolerance
- Research lead turns assumptions into one question
- Approve briefobservation unit, England criterion, period, exclusions
- Retain brief and decision log
- Metriccompleted mandatory fields divided by all fields due
- Stop if no finding could affect the choice
The decision question itself comes from a decision-first strategy that names the owner, the deadline and the alternatives. The table gathers every stop or pause trigger in one place.
| Stage | Stop or pause when | Record that proves the gate |
|---|---|---|
| Days 1 to 15 | No finding could affect the choice, or qualified review rejects the use | Decision brief and decision log |
| Days 16 to 30 | Unauthorised access, uncontrolled sensitive data, or a route without acceptable exit provisions | Dated source register and route approvals |
| Days 31 to 45 | Measured precision falls below the agreed target, or unresolved disagreement passes its limit | Query log, item IDs, labels and adjudication notes |
| Days 46 to 60 | High-risk errors exceed the agreed limit, or a listed consequence occurs | Classification record and escalation route |
| Days 61 to 75 | Wording presents the stream as representative of England | Evidence pack |
| Days 76 to 90 | Evidence cannot inform the choice | Proceed, revise or close record |
Name who can call each stop
A stop rule needs a person who can invoke it. Write the named individual and job function beside every owner, because a role label alone is not an assignment. The stage owners supply different things to the gate.
- Days 1 to 15the decision owner supplies the choice, deadline, population and risk tolerance, and the research lead turns assumptions into one question.
- Days 16 to 30the platform lead supplies first-party terms, the privacy lead maps fields, users, retention, sharing and rights, and security records supplier evidence.
- Days 31 to 45the query editor supplies versioned terms and exclusions, and two reviewers label a sample from the approved period.
- Days 46 to 60the analysis lead supplies label definitions and the model version, while privacy, AI and security reviewers assess errors, access and incidents.
- Days 61 to 75the analyst uses approved returns and the decision owner checks relevance.
- Days 76 to 90the authorised owner considers eligible recommendations.
Treat source and query failures as stop events
Source authorisation ends in a dated source register and an approval for each route. Stop unauthorised access, uncontrolled sensitive data or a route without acceptable exit provisions. Recheck after any change to terms, fields or subcontractors.
Days 16-30: Authorise sources and data flow
- Platform lead supplies first-party terms
- Privacy lead maps fields, users, retention, sharing, rights
- Security records supplier evidence
- Produce dated source register and approval for each route
- Metric: approved routes divided by routes proposed
- Stop unauthorised access or uncontrolled sensitive data
Assess each route under UK GDPR and the Data Protection Act 2018. The ICO's Guide to data protection explains that public-source data can still require privacy information, subject to exceptions. Screen whether a data protection impact assessment (DPIA) is required, complete one where required, and keep a record of processing activities (ROPA) where applicable.
Name the approved source set and the listening platform before testing. Examples include public posts on X or Reddit, and tools such as Brandwatch, Talkwalker or Meltwater, though access and query syntax depend on the route and product. Two reviewers then label a sample from the approved period.
For example, on a platform that supports this syntax, log ("brand name" OR #brandname) NOT "job vacancy" and test the exclusion against reviewed items. Adapt operators to the selected platform and record the exact query version used. Keep the item IDs, so the test can be repeated after any vocabulary or syntax change.
Query testing stops on numbers agreed before the test. Set and record the sample size, sampling method and minimum acceptable precision according to decision risk. The gate fails if no sample size or target was agreed, or if measured precision falls below the target.
Report precision by source and query version, and do not claim recall without a complete relevant set. Record a maximum unresolved-disagreement rate, meaning unresolved items divided by items reviewed, and pause if either limit is breached. If no defensible limits are agreed, fail the gate.
Set automation limits before the model runs
Before any automation, record the maximum tolerable error rate, meaning incorrect classifications divided by adjudicated items. Define which consequences count as high risk. If no defensible limit is agreed, do not automate.
Stop high-risk automation if observed errors exceed the agreed limit or a listed consequence occurs. The ICO's Guidance on AI and data protection says AI error measures should reflect purpose and risk. Reassess after model updates, drift or material errors.
Walk the owners through the failures worth rehearsing before the pilot starts, such as unauthorised access or an error rate past its limit, so each stop rule has a named person who will act on it.
Assemble the evidence pack
The evidence pack gives source, dates, query version, numerator, denominator, missing data and inference limit for every count. Keep observation, finding, recommendation and decision records separate. The analyst uses approved returns, and the decision owner checks relevance.
Measure completeness as findings with complete fields divided by findings proposed in days 61 to 75. Exclude retractions but preserve their audit trail. Each count needs a complete definition, with population, numerator, denominator, period, source and owner, as set out in the metric definitions.
Reopen the pack for missing context or conflicting evidence. Stop publication if the wording presents the stream as representative of England.
Record proceed, revise or close
The authorised owner records proceed, revise or close, with the evidence, the rejected options and the next review date. Only eligible recommendations are considered. Items outside the owner's authority are excluded and named, so the record shows what was left out and why.
Recommendations accepted, divided by eligible recommendations considered in days 76 to 90, measure workflow uptake. That figure is not impact or return on investment. Use HM Treasury's Magenta Book to plan evaluation from the outset.
Public claims need review by legal and advertising specialists against the advertising rules, which for UK advertising means the CAP Code applied by the Advertising Standards Authority. Reopen the record after a new purpose, a source-policy change, a privacy complaint, an incident or an unsupported claim. Close if evidence cannot inform the choice. Named specialists must approve their fields before use, and nothing here grants compliance assurance.
Before you close
- Check that every stage has a written stop rule in the approved brief.
- Confirm the sample size, minimum precision and error limit were recorded before testing.
- Assemble an evidence pack with the seven fields for every count.
- Record proceed, revise or close, with rejected options and a next review date.
- Name any items excluded because they fall outside the owner's authority.
- List the triggers that reopen the record.
Common questions
What goes into the proceed, revise or close record?
It holds the decision, the evidence behind it, the rejected options and the next review date. The authorised owner makes the call on eligible recommendations only, and items outside that owner's authority are named and excluded.
What reopens a closed record?
A new purpose, a change to source policy, a privacy complaint, an incident or an unsupported claim reopens it. The pack itself reopens for missing context or conflicting evidence. If evidence cannot inform the choice, close.
Is recommendation uptake the same as return on investment?
No. Recommendations accepted divided by eligible recommendations considered is workflow uptake. It is not impact or return on investment, and the plan does not target sentiment, reach or return on investment.
