How to Refine an Audience (Fast) Without Losing the Plot
This guide shows a practical, repeatable loop to tune an audience after you “ship it,” using real examples and clear guardrails. It covers both directions—narrowing when you’re too broad and broadening when you’re starved for volume—plus how to use preview, keywords, and historical win rates to make smart cuts (or adds) without guesswork.
TL;DR — The Refinement Loop
Check basics: size → projected match rate → early performance (CPM, CTR, CVR, CPA).
Decide the direction: Are you too big (low quality, low penetration) or too small (low volume, learning never exits)?
Run the right play:
Narrowing play (too big): cut by industry, titles, and negative keywords; use win rates to keep what converts.
Broadening play (too small): add adjacent industries/titles/departments; relax firmographics; add lookalike/expansion where useful.
Preview → spot-check → revise: inspect companies/titles in preview, remove obvious misfits, re-run projections.
Measure with intent: compare before/after on CPM/CTR/CVR/CPA; explore Primer's conversion lift in Overview to see impact compared to holdouts.
Repeat until you hit your penetration + efficiency target.
Key Concepts (Quick Definitions)
Audience size: number of reachable people/companies after filters + match.
Projected match rate: estimated % of your list/filters the ad platform can actually match.
Penetration (coverage): reflection of how much of your audience your campaign has reached within your time window; heavily influence by budget, CPMs, and audience size.
Historical win rates: opportunity closed won patterns from your CRM (by industry, size, region, etc.) that indicate where you close deals most often.
Preview: a spot-check of companies, titles, and traits to validate fit before spending.
Rule of thumb: For initial tests, aim for an audience that you can reach meaningfully within the test window (e.g., 10–30% penetration over 2–4 weeks), without forcing CPMs to the moon.
Scenario 1: You Started Broad → Now You Need to Narrow
Typical symptoms
Audience is massive.
Penetration is low given your budget/timeframe.
CTR mediocre, downstream quality uneven.
Example (real):
We launched a brad ICP audience. Penetration was low. First pass: added keyword filters to keep only software companies, regardless of their industry categorization. Better, but Hulu (clearly not a fit because it's B2C and we're a B2B company) still showed up and the audience was still large. Second pass: used historical win rates to select only industries where our close rates are high. That cut size and improved expected quality.
Step-by-step “Narrowing” play
Apply win-rate filters
Review Historical Win Rates by industry, company size, location, etc.
Keep only top-converting buckets (e.g., “Software; 200–2,000 employees”) and exclude low performers.
Use preview to create your “Not-a-Fit” list
Skim companies → idenfity obvious mismatches that may show up even if the industry win-rate is strong. (e.g., media/streaming like “Hulu” if you sell B2B SaaS infra)
Add negative keywords tied to those patterns (e.g., “streaming,” “broadcast,” you repeatedly see but don’t serve)
Title/department shaping (precision pass)
Review Historical Win Rates by department and job title.
(optional) Still keep buying committees that may not have a high win rate but you believe influence your deals (e.g. C-Suite, etc.).
Add negative titles (e.g., “Intern,” “Student,” “Freelance”) if noise creeps in.
If still too broad, tighten firmographics and technographics even further
Company size bands (e.g., remove 1–10 if they almost never buy).
Required technologies (if your solution aligns well with them).
Region/geo if your sales coverage doesn’t support certain areas.
Re-check projections
Confirm audience size vs. current budget and current CPMs → your estimated penetration should improve.
If still broad, iterate steps 1–4 again (small, logical cuts—don’t over-prune in one go).
What to watch in performance
Penetration should improve.
CTR should rise (relevance ↑).
CVR should rise.
If volume collapses, you cut too far—undo the last cut or go to the Broadening play.
Scenario 2: You Started Too Narrow → You Need to Broaden
Typical symptoms
Audiences on Meta/LinkedIn are tiny; learning never exits.
CPM volatile; CTR okay but not enough conversions to judge.
Example (real):
For BLVD, we asked them to broaden because they were running lead gen on Meta with very small audiences.
Step-by-step “Broadening” play
Add adjacent industries
Remove overly constricting industry filters.
Use win rates to find secondary but decent converters (e.g., “IT Services” next to “Software”).
Relax firmographics
Widen company size bands (add 50–200 or 2,000–5,000 if they’re not terrible).
Expand regions you can sell to.
Grow the buying committee
Add adjacent departments/titles (e.g., beyond “Marketing Ops” to “Sales Ops,” “Finance,” “Procurement,” “Legal” for enterprise motions).
If primarily targeting by job title, use an LLM to suggest additional job titles similar to the existing ones. Paste them in as a comma sepratead string.
Re-check projections
Audience size increases; penetration at current budget should look healthier → better chance to collect statistically useful data.
Ensure Conversion APIs are set up
Ad platforms need signal to exit learning mode. Ensure that you're sending offline conversions to ad platforms.
Keep in mind that B2B audiences in B2C ad platforms may never exit "learning mode" given audience size and time to conversion constraints. That's ok. But you'll have to do more of the work of campaign optimization and measurement as a result.
Enable platform expansion tools (where sensible)
Carefully test lookalikes/optimized targeting on top of a quality seed.
Combine with anti-ICP exclusion audiences to box in the algorithm
What to watch in performance
Conversion volume should rise.
CPA might wobble at first—judge after sufficient data.
If quality falls sharply, revert the last broadening step and try a different expansion (e.g., a different adjacent industry rather than looser size).
Penetration: A Simple Planning Check
Impressions ≈ (Spend / CPM) × 1,000
Rough reach ≈ Impressions / frequency (if you assume ~1.5–2.5 for prospecting)
Penetration ≈ Reach / audience size
Goal: Your budget over the test window should produce meaningful penetration (e.g., enough unique reach to generate statistically useful results). If penetration looks thin, either narrow (to lift penetration) or boost budget / broaden (to get enough data).
Guardrails & Pitfalls (Read This)
Small-but-precise ≠ always better. Ultra-tiny audiences can trap you in learning with spiky CPMs, but sometimes this may be necessary if your ICP is narrow.
Over-segmentation drives cost. Each extra constraint raises auction pressure; make changes in small steps, measure, then proceed.
Platform behavior matters. LinkedIn handles small audiences differently than Meta/Google; give each platform what it needs to learn.
Match rate lies by omission. A great match rate on a tiny base won’t move the needle; a moderate match rate on a bigger, qualified base often wins.
Preview is truthy. If preview looks wrong, fix your filters before spending another dollar.
Putting It Together—Two Real Patterns
A) “ICP Audience Clean-Up” (Narrowing)
Start broad → low penetration.
Add software-only keyword filters.
Remove obvious misfits from preview (e.g., Hulu) via negative keywords/industry exclusions.
Keep only high win-rate industries; exclude the rest.
Re-check size & penetration; ship.
B) “Meta Lead Gen Can’t Learn” (Broadening)
Audience too small; volume poor.
Add adjacent industries (still relevant).
Expand titles/departments (more of the buying unit).
Relax size bands slightly.
Consider lookalikes/optimized targeting layered on a clean seed.
Re-check size & penetration; ship.
Measurement That Actually Helps You Improve
Timebox your changes: evaluate at 7 and 14 days (or after N impressions) before deciding.
Track CPM, CTR, CVR, CPA—and holdouts to see incremental lift.
When you find a winning slice (e.g., “Software, 200–2,000, RevOps + Finance”), clone it and test one additional variable at a time.
Last updated