LinkedIn Ads campaign review

A framework for the periodic review of a live LinkedIn Ads account. It splits the work into independent parallel workstreams, consolidates and cross-checks their findings, and turns them into ranked, human-approved changes plus manual tasks.

Defaults for BigData Boutique accounts: B2B data/search-infrastructure consulting, technical and engineering-leader audience, long sales cycle, low conversion volume. LinkedIn is the precision-targeting channel (role, seniority, company), with the highest cost per click of our paid channels, so judge it on qualified pipeline, not CTR or platform CPL.

Naming. Campaign Manager’s UI hierarchy is campaign → ad set → ad. The Marketing API and our MCP tools use the older names: campaign group → campaign → creative. In this guide, “ad set” means the API campaign (where budget, bid, targeting, Audience Network and Audience Expansion live). Always cite API ids in findings.

1. Principles

1. Measurement first. Insight Tag + Conversions API (deduplicated), lead gen forms synced to the CRM. Without them every CPL finding is noise. L1 gates the rest.
2. Judge on qualified outcomes and economics. Use lead → SQL rate, cost per SQL and pipeline against the max allowable CPL / cost per SQL from Phase 0. Platform CTR and CPL are diagnostics, not verdicts.
3. Your own trailing history is the benchmark. Compare to the account’s prior periods and to the other paid channels. Do not compare to third-party “industry averages”.
4. View-through is suspect for B2B. Report click and view conversions separately.
5. The ad set is the unit of decision. Break it down by creative, demographic (function, seniority, company, country) and serving location (on/off network) before judging it.
6. Respect learning. Batch edits per ad set. At most one spend-affecting change per ad set per cycle, in ≤20–30% steps.
7. Every finding cites data: entity ids, window, metrics and the tool call. No generic advice.
8. Humans approve every change. Reviews only propose (propose_linkedin_ads_change). LLM agents have no LinkedIn write tools.

2. Cadence: Weekly, Monthly, Quarterly

Windows: weekly = last 7 days + 28-day context. Monthly = 30 vs prior 30. Quarterly = 90 vs prior 90. Reporting is in UTC. Demographic data lags 12–24 h in the API (24–48 h in the Campaign Manager UI), so end windows 2 days ago. With long conversion windows, recent periods keep gaining conversions; compare equally “matured” windows.

3. Execution Model

Phase 4  Follow-up (runs FIRST)              ─ did last cycle's applied changes deliver?
Phase 0  Context pack (sequential)            ─ account, economics, windows, entity map, ICP
Phase 1  L1…L9 workstreams (PARALLEL)         ─ each pulls its own data → Findings[]
Phase 2  Consolidation (sequential)           ─ L1 gate, cross-check matrix, dedupe, sanity
Phase 3  Prioritize & act (sequential)        ─ score, ≤N change proposals + manual tasks + report

Workstreams read only the context pack plus their own data and return candidate actions, never proposals. Proposals and tasks are filed once, after consolidation, so duplicates and conflicts are never filed.

4. Phase 0: Context Pack

5. Phase 1: Parallel Workstreams

Severity: 🔴 critical, 🟠 material, 🟡 hygiene. Thresholds referenced here are defined once in §12.

L1 Measurement & Attribution

Data: list_conversions, get_conversion_performance, get_lead_gen_performance, CRM leads with utm_source=linkedin, Campaign Manager UI (Insight Tag and CAPI status: manual).

• Insight Tag live on all pages; each key conversion recorded recently. A primary conversion with zero events while clicks continue → 🔴.
• Conversions API sending key events server-side. Missing → 🟠. When the same event is sent by Tag and CAPI, both must use separate conversion rules sharing an eventId, or it is counted twice. LinkedIn keeps the Tag event and discards the CAPI duplicate, so the Tag rule normally shows the higher count. Similar counts on both rules with no eventId → 🔴 double counting.
• Association: every Lead Gen / Website Conversion ad set has the right conversions associated. An unassociated or disabled rule on a conversion ad set → 🔴 (proposable fix).
• Windows: options are 1/7/30/90 days click and view (180/365 for some CAPI/CSV lead and purchase categories). The default is 90-day click / 90-day view. Attribution is last touch. Check each rule is deliberate. In efficiency findings, use click-through conversions.
• Lead gen forms → CRM: form leads arrive in the CRM (native sync or integration) with campaign and ad set ids. Missing or lagging → 🔴. Form leads with no CRM record cannot be judged on quality.
• Reconciliation: platform conversions vs CRM leads for the same window. Gap above threshold → investigate (dedup, spam, stripped UTMs, consent).
• UTMs on every destination URL (source/medium/campaign = ad set id/content = creative id). 🟠
• Revenue Attribution Report (Business Manager, needs a connected Salesforce, Dynamics 365 or HubSpot CRM) used for pipeline influence if available. Absent → 🟡 note (manual).
• Self-reported attribution field on demo/contact forms. Missing → 🟡.

Verdict (required): measurement_confidence: high | medium | low.

L2 Account Structure & Governance

Data: get_campaign_groups, list_campaigns (all statuses), list_creatives, UI for access.

• Structure follows funnel role: campaign groups by goal/stage. Ad sets separated by audience type (prospecting vs retargeting vs account list) and by format, so results are attributable. 🟡
• Fragmentation: many small ad sets splitting a low budget, each below the learning floor → consolidate. 🟠
• Rejected or paused-by-review creatives on top spenders → 🔴.
• Stale objects: ended ad sets still “active”, ad sets with no active creative, empty groups. 🟡
• Naming encodes funnel stage, audience, format, offer. 🟡
• Access: least privilege, former vendors removed (manual). 🟡

L3 Budget, Pacing & Delivery

Data: get_campaign_performance (DAILY), get_daily_trends, budgets from get_campaign_details / get_campaign_groups.

• ACTIVE with no or near-zero spend for 3+ days → delivery problem: bid/cap below market, audience too small, creative rejected or missing, schedule. 🔴 on key ad sets.
• Under-delivery (spend below the threshold share of daily budget for a week) → budget is not the constraint. Look at bid, audience size and creative (L4/L5/L6). 🟠
• Daily overspend is normal: LinkedIn may spend up to 100% above the daily budget on a given day while managing the total. Flag only sustained over-plan pacing.
• Pacing vs monthly envelope outside the threshold band. 🟡/🟠
• Concentration: one ad set takes most of the spend without matching results. 🟠
• Group budgets / Dynamic Group Budget: if a campaign-group budget is on, ad-set budgets are not the lever. Check which level controls spend before proposing a budget change.
• Schedule: custom day/hour schedules exist only with a lifetime budget, Sponsored Content and Audience Network off. Check that schedules and end dates are deliberate.

L4 Objective & Bidding

Data: get_campaign_details (objective, costType, optimizationTargetType, unitCost), results.

• Objective and format fit the role. Both are fixed at launch (§12), and an ad set must use its group’s objective, so a misfit is fixed with a new ad set (X15). Formats per objective follow LinkedIn’s eligibility table (B2B objectives only):

  Role	Objective (objectiveType)	Eligible formats	Prerequisites	Judge on · red flag
  Awareness	Brand awareness (BRAND_AWARENESS)	single image, carousel, video, document, event, text, spotlight, follower, article/newsletter, Thought Leader, conversation	—	CPM, reach, frequency · judged on CPL
  Consideration	Engagement (ENGAGEMENT); Video views (VIDEO_VIEW, video only)	single image, carousel, video, document, event, follower, article/newsletter, Thought Leader, conversation	Engagement and video-view retargeting audiences to harvest	engagement/view rate, cost per engaged member · expected to produce SQLs 🟠
  Consideration	Website visits (WEBSITE_VISIT)	single image, carousel, video, document, event, text, spotlight, message, conversation	UTMs, LP	CPC, engaged sessions · funding a demo goal 🟠
  Conversion	Lead generation (LEAD_GENERATION)	single image, carousel, video, document, event, article/newsletter, message, conversation	Lead gen form (or CAPI), CRM sync (L1/L7)	CPL → SQL · pre-filled-only form, leads missing in CRM
  Conversion	Website conversions (WEBSITE_CONVERSION)	single image, carousel, video, document, event, text, spotlight, message, conversation	Insight Tag/CAPI rule associated, enough events (last bullet)	click-through CPA → SQL · unassociated rule, too few events

  Job applicants and Talent leads (needs a Recruiter contract) are recruiting objectives. In a marketing account → 🟡. Proposable: a new DRAFT ad set via create_campaign (type and format: STANDARD_UPDATE, CAROUSEL, SINGLE_VIDEO, SPONSORED_UPDATE_NATIVE_DOCUMENT, SPONSORED_UPDATE_EVENT, SPONSORED_INMAIL = message, SPONSORED_MESSAGE = conversation, TEXT, DYNAMIC + SPOTLIGHT/FOLLOW_COMPANY). Its ads go through create_inline_ad (an existing image/video URN) or create_creative (an existing post, an adInMailContent or a clone). Manual: Thought Leader Ads (member permission), new video/document/carousel assets, message/conversation copy, event pages.

• Bid strategy (LinkedIn offers three: Maximum delivery, Cost cap, Manual bidding; there is no “Enhanced CPC”):

  Situation	Fitting strategy
  New ad set, no cost history, want full budget spent	Maximum delivery (CPM-charged, no bid input)
  Known cost per result, need control	Cost cap at the real max allowable cost. It’s an average target, and actual cost can differ
  Small exact audience (account lists, retargeting), strict CPC/CPM control	Manual bidding

• Cap or manual bid far below actuals → no delivery (cross-check L3). Maximum delivery with cost per result above the max allowable → test cost cap.

• Learning resets: changes to audience, budget, bid strategy or creatives restart learning. LinkedIn recommends running cost-cap ad sets ≥15 days. Repeated edits inside that window → 🟠.

• Optimization goal too deep for volume (e.g. optimizing for a conversion with a handful of events a month) → step up the funnel or consolidate. 🟠

L5 Targeting & Audiences

Data: get_campaign_details (targeting), get_audience_demographics (function, seniority, title, industry, company size, company, country/region), get_ad_segmentation (SERVING_LOCATION, IMPRESSION_DEVICE_TYPE), get_audience_reach, estimate_audience_size, list_saved_audiences, search_targeting_facets.

• Who actually saw/clicked vs the ICP map: share of spend on off-ICP functions, junior/unpaid seniority, wrong company sizes or companies (resolve_organization_names after MEMBER_COMPANY). High off-ICP share → narrow or exclude. 🟠
• Demographic pivots (get_audience_demographics demographicType=): MEMBER_JOB_FUNCTION, MEMBER_SENIORITY, MEMBER_JOB_TITLE, MEMBER_INDUSTRY, MEMBER_COMPANY_SIZE, MEMBER_COMPANY, MEMBER_COUNTRY_V2, MEMBER_REGION_V2, MEMBER_COUNTY (prefer the V2 geo values). Filter with campaignIds; sort by impressions/clicks/costInUsd. Its conversions are website conversions only. Lead gen form leads by demographic come from the Campaign Manager UI (manual).
• Reading demographics: rows with < 3 events are dropped, and so are ads with too few engaged members. Only the top 100 values per creative per day are returned, so long-tail titles undercount. Values are approximate (within ~3 units) and kept 2 years. Reach and carousel metrics aren’t available on these pivots. Compute off-ICP share over the returned rows and state their coverage of the ad set total. MEMBER_COMPANY legitimately includes non-targeted companies (members with several employers or who changed jobs, Audience Expansion, parent companies), so that alone isn’t a targeting error.
• Selection logic: values within a facet are OR, while facets and “Narrow audience further” are AND. Exclude removes members. LinkedIn recommends job function + seniority over a short list of job titles, skills where titles miss the expertise, and no more than 2–3 facets. It suggests testing titles vs skills, and industry vs function, in separate ad sets rather than layering them. Internal rule of thumb: the ICP base is function × seniority × company size (+ geo); titles only for account-list or test ad sets; ≤ 3 AND layers before exclusions.
• Audience size: LinkedIn’s minimum is 300 members. It recommends ~50,000+ in general and 300,000+ for Sponsored Content and Sponsored Messaging. Account lists and retargeting are legitimately smaller; flag small audiences only when delivery is constrained (L3).
• Audience Expansion is on by default where available. Expansion metrics are blended into the ad set’s reporting, so it can’t be split out. On ICP-precise or account-list ad sets with weak quality → turn off (proposable). 🟠
• LinkedIn Audience Network is on by default for new single image, carousel, document and video ads. Compare on vs off network (SERVING_LOCATION). Off-network spend with poor engagement or conversion quality → disable, or apply a block list/IAB exclusions (UI, manual). 🟠
• Predictive Audiences replaced Lookalikes (Lookalikes discontinued Feb 29, 2024; existing ones are static). They need a source of 300+ members, and Audience Expansion is off on ad sets that use them. Static lookalikes still in use → 🟡 replace with a Predictive Audience test.
• Exclusions: customers, open opportunities, recent converters, employees and competitors excluded from prospecting. Missing → 🟠.
• Matched audiences (contact/company lists, website/video/form retargeting): status ACTIVE, sizes growing, lists refreshed (stale > 90 days → 🟡). Retargeting should beat prospecting on CVR. If not, look at frequency/creative (L6).
• Overlap: the same audience in several active ad sets → self-competition. 🟡
• Geo & language (folded-in): location options are permanent, recent, or “recent or permanent” (IP-based recent visits widen reach to travellers). Use permanent for standard B2B. Geo targeting is by member location, not company HQ. Multi-country ad sets let expensive markets absorb the budget (no per-geo bid adjustments), so split by country when cost per SQL differs materially. Profile language can’t be changed after launch (new ad set needed). Creative and LP must match it.

L6 Creative & Fatigue

Data: get_creative_performance, list_creatives (age, status), get_audience_reach (penetration), get_ad_segmentation (CARD_INDEX, CONVERSATION_NODE).

• Variants: active creatives per ad set within the threshold band. < 2 → no learning. 🟠
• Fatigue: CTR down beyond the threshold vs the creative’s first two weeks while frequency or audience penetration rises → refresh. 🟠
• Winners/losers: rank within an ad set on CTR and CVR (click-through). Pause the bottom performer once it has the minimum data (§10).
• Format fit: judge video on view metrics, carousels per card (CARD_INDEX), Conversation Ads per node (CONVERSATION_NODE). Don’t judge awareness formats on CPL.
• Copy truncation: single-image intro text truncates after ~150 characters and headlines after ~70. Key message and CTA must sit before the cut. 🟡
• Thought Leader Ads (sponsored member posts): with the member’s permission, from employees or other members. Objectives: brand awareness or engagement (and video views for video). Documents, polls, multi-image, celebrations and reposts are ineligible. If not tested for awareness/engagement, suggest a test (manual: needs the member’s permission).
• Sponsored Messaging (Message / Conversation Ads): charged per send. LinkedIn strictly limits how often a member receives them, so judge on open/click/lead per send, not frequency. EU/EEA/CH members receive them only if they opted in.

L7 Lead Gen Forms & Landing Pages

Data: get_lead_gen_performance, list_lead_forms, destination URLs from list_creatives, HTTP checks, GA4 for utm_source=linkedin.

• Form completion rate (opens → submits) per form vs its own history. Drop → form change or audience shift. 🟠
• Qualifying questions vs volume: forms with only pre-filled fields produce easy, low-intent leads (cross-check L8). Too many custom questions cut completion.
• Shared form edits: updating a PUBLISHED form changes every ad set that uses it. Prefer a new form when testing.
• LPs: broken/redirecting URLs, stripped UTMs → 🔴/🟠. Message match ad ↔ LP. Mobile speed.
• Offer fit by role: cold audiences → content/benchmark/guide. Retargeting and account lists → consultation/assessment. Cold traffic straight to “book a call” with poor CVR → 🟠.

L8 Business Outcomes & Lead Quality

Data: CRM (HubSpot) leads by ad set/creative ids or UTMs, lifecycle stages, pipeline; Revenue Attribution Report if connected; self-reported attribution.

• Lead → SQL rate and cost per SQL per ad set vs max allowable and vs Google/Reddit for the same window and definitions. 🟠/🔴
• Junk patterns: students, job seekers, vendors, wrong geos, personal emails concentrated in an ad set, form or expansion/LAN traffic → exclusions.
• Account-level influence: target accounts engaged (MEMBER_COMPANY), opportunities opened at engaged accounts. LinkedIn often influences without last-click credit; record that evidence before cutting.
• No CRM access → say so. Efficiency findings are then platform-based and lower confidence.

L9 Competitive & Market (quarterly)

Data: LinkedIn Ad Library (public, manual), search_targeting_facets (“similar” companies), own cost trends.

• Competitor messaging and offers in the Ad Library. Do our angles differentiate?
• New ICP segments or account-list additions (similar companies, new titles/skills).
• Cost trend per ad set over 12 months (own data) to separate seasonality from performance drift.

6. Finding Format

- id: L5-03
  workstream: L5
  title: "Audience Network takes 41% of 'Eng Leaders - Prospecting' spend, 0 SQLs"
  entities:
    campaign_group_ids: ["…"]
    campaign_ids: ["…"]        # API campaign = UI ad set
    creative_ids: []
  evidence:
    window: "2026-09-04..2026-10-01"
    metrics: {spend: 1240, offsite_share: 0.41, offsite_ctr: 0.0011, sqls_offsite: 0}
    source: "linkedin_ads_get_ad_segmentation segmentationType=SERVING_LOCATION"
  severity: material           # critical | material | hygiene
  confidence: high             # inherits L1 measurement_confidence
  est_monthly_impact: 1300     # account currency
  root_cause_hint: "LAN enabled by default on a precision ad set"
  candidate_actions:
    - kind: proposable         # proposable (propose_linkedin_ads_change) | manual
      tool: update_campaign
      args_hint: {campaignId: "…", offsiteDeliveryEnabled: false}
      summary: "Disable Audience Network on ad set …"
  depends_on: []
  conflicts_with: []

No finding without metrics and a window. Click-through conversions unless stated.

7. Phase 2: Consolidation & Cross-Checks

7.1 Measurement gate (L1)

7.2 Cross-check matrix

7.3 Dedupe & sanity

• Merge same-entity findings across workstreams (L3 under-delivery + L4 low cap → one finding).
• Collapse leaf findings into one cause (12 junior job titles → “exclude Entry/Unpaid seniority”).
• Totals reconcile with get_campaign_performance for the account (±2%).
• Ids exist and are in the stated status. Savings ≤ actual spend. No proposal contradicts a recorded human decision, an open change task, or an open manual task.

8. Phase 3: Prioritization, Change Proposals & Manual Tasks

Score: impact × confidence ÷ effort (confidence 1.0 / 0.6 / 0.3; effort 1 = one API call, 2 = new creative/form/LP, 3 = restructure/tracking). Critical items jump the queue: broken tracking, double counting, forms not reaching CRM, rejected top creatives, runaway spend.

Buckets: Fix now → Quick wins (exclusions, AE/LAN off, pauses, cap/bid fixes, one budget step) → Optimizations (creative refresh, form questions, ad-set splits by geo/audience) → Strategic (CAPI, CRM loop, Revenue Attribution, Predictive/TLA tests, channel role).

8.1 (a) Proposable changes → linkedin-ads-change tasks

Call propose_linkedin_ads_change(tool_name, arguments, summary, expected_impact, account_id, rationale), one call per change, highest priority first, capped (weekly ≤10). Each call creates a task on the Paid Ads board (todo, labels linkedin-ads-change + LinkedIn Ads, requires_human=true) assigned to the Ad Change Applier. After a human clears Requires human, the applier runs the call exactly once and records applied_at / apply_result / applied_resources (urn:li:…). A duplicate open proposal returns status: duplicate. Don’t re-file it. Use the tool’s own argument names (accountId, campaignId, …). An error lists them.

upload_image is allowlisted but takes a file path on the MCP server, so it isn’t practical from a review. New image assets are manual.

Each proposal carries summary (becomes the title), expected_impact (quantified from the finding) and rationale (evidence + cross-checks passed).

8.2 (b) Manual changes → human tasks

Anything not on the allowlist, or UI-only, becomes a human task. If the agent has the tasks output, emit one task per action: board_id: paid-ads, labels LinkedIn Ads + linkedin-ads-review, requires_human: true, fields.source_task_id = the review task (a pointer only, no copied prose). Use a short title and description holding the action, entity ids and evidence. Otherwise, list them in the report’s “Manual actions” section. Typical manual items: Insight Tag / CAPI / eventId dedup, conversion window changes, CRM form sync, Revenue Attribution setup, LAN block lists, Thought Leader Ad permissions, new image/video creative, LP work, A/B tests, access cleanup.

8.3 Experiments over guesses

For uncertain changes (Maximum delivery vs cost cap, AE or LAN on/off on a converting ad set, Predictive Audience vs ICP targeting, form vs LP), recommend a Campaign Manager A/B test (ad, audience or placement variable; runs 2 weeks to 90 days) or a duplicated ad set with one variable changed, with cost per SQL as the agreed metric (manual task).

9. Phase 4: Follow-up of Last Cycle’s Changes

Run first, before new findings. For each linkedin-ads-change task since the last review:

Also check that manual tasks from the last review were done (e.g. CAPI live, exclusion list uploaded). Unfinished critical manual items stay at the top of the report.

10. Decision Frameworks

Minimum data before acting

Pause vs optimize (ad set)

Cost per SQL ≤ target (or CPL ≤ max allowable with OK SQL rate)? ── yes → keep; scale if budget-limited
        │ no
Enough data (above)? ── no → watch; fix only obvious delivery/creative gaps
        │ yes
Tracking healthy (L1)? ── no → fix tracking; don't pause on CPL
        │ yes
Off-ICP delivery / AE / LAN (L5)? ── yes → narrow/exclude, AE/LAN off
        │ no
Creative fatigued (L6)? ── yes → refresh, re-evaluate next cycle
        │ no
Form/LP/offer mismatch (L7)? ── yes → fix offer/form
        │ no
→ lower bid/cap, or pause

Scale when all hold

• Cost per SQL ≤ target, SQL rate OK, L1 ≥ medium
• Spend ≈ budget (budget-limited) and audience penetration not yet high
• Not in learning; last change ≥ 72 h ago
• Step ≤ 20–30%; re-check after 7 days. Or clone into an adjacent ICP segment, one variable at a time

Refresh creative when

• CTR down beyond threshold vs the creative’s first two weeks with frequency/penetration rising
• Creative live past the age threshold at meaningful spend
• Comments or sales feedback raise objections the ad doesn’t address

11. Symptom → Root-Cause Diagnosis

12. Platform Facts & Operating Thresholds

Verified platform facts (LinkedIn help center):

Operating thresholds (internal heuristics, not LinkedIn rules; tune per account):

13. Report Format

One markdown asset per review: LinkedIn Ads Review: <account_id> <period>.

1. Verdict (healthy / needs attention / critical) + measurement confidence (L1).
2. Scorecard: spend, impressions, CPM, CTR, CPC, conversions (click / view), CPL, leads → SQLs, cost per SQL, vs previous period and target. Windows stated.
3. Last cycle follow-up (Phase 4).
4. Critical issues.
5. Ranked actions: id, action, entity, expected impact, confidence, proposable/manual, task id.
6. Proposed changes (change task ids) and Manual actions (task ids or list).
7. Workstream notes, including “checked and fine”.
8. Deferred / blocked (cross-check id, revisit date).
9. Open questions (targets, ICP, CRM access, tests to run).

Appendix: Data Pulls (LinkedIn Ads MCP)

Tools are prefixed linkedin_ads_. LLM agents can bind only read tools (get_ / list_ / search_ / compare_ / estimate_ / resolve_). Writes go only through propose_linkedin_ads_change. List tools default to ACTIVE, so pass all statuses for audits.