A narrow keyword inventory can miss useful buying questions. Unreviewed negatives can exclude relevant traffic. Neither list size tells you how much of a market you cover.
Connect a supported search intent to an ad and destination, then evaluate the result against the account's economics. The seven research lenses, RSA rubric and decision gates below organize that work; they do not prescribe one architecture for every account.
Why Most Search Campaigns Hit a Ceiling
Inspect whether an ad group combines different buying questions without giving each a useful message or destination.
Consider "running shoes," "trail running sneakers," and "marathon training shoes." If one RSA and page must answer all three, check whether their assets address those needs. Do not assume every query receives the same headline: responsive search ads can assemble different assets for different searches.
A price question may need offer conditions; a comparison question may need substantiated differences. Group a coherent intent with suitable messages and destinations where that improves control. Product categories can still work when their messages, economics and reporting needs align.
Evaluate actual query/ad/destination evidence and conversion quality. The grouping alone establishes neither a Quality Score change nor a CPC discount.
Use Quality Score as a Diagnostic
Quality Score summarizes expected CTR, ad relevance and landing-page experience at keyword level. It is not the auction's pricing formula.
Google's Quality Score guidance advises against treating it as a KPI or aggregating it into one. Inspect the keyword components and underlying queries, ads and pages instead.
An inherited report may contain this calculation:
Campaign weighted diagnostic =
SUM(keyword Quality Score × keyword impressions) / SUM(keyword impressions)
It is a legacy internal summary, not a recommended target or official account score. Unavailable scores remain unknown, not zero.
Prioritize investigations from actual spend, business value and the affected experience. A better headline, clearer grouping or more useful page is a change to evaluate, not a performance result established by a score threshold.
The 21-Query Keyword Protocol
Use these seven research lenses where they fit the offer. Some may yield no relevant demand.
Layer 1: Core Product. Product-defined searches. Check actual terms and competitive conditions; their CPC is not automatically highest.
Layer 2: Specificity. Ingredient, feature, material or use-case searches. Added detail must represent a supported buying question; it does not guarantee better traffic.
Layer 3: Conquest. Competitor comparisons and "alternative to" queries. Review product fit, permitted claims and applicable brand/policy constraints.
Layer 4: Seasonal. Questions whose relevance changes with a season or event. Inspect the appropriate market and period.
Layer 5: Question/PAA. Questions from search results and related research. Their presence alone establishes neither your audience nor acquisition cost.
Layer 6: Discovery. Adjacent problems or categories. Require a defensible connection to what the product actually does.
Layer 7: Shopping Micro-Intent. Purchase modifiers combined with product types. Generated phrases are candidates, not evidence of search demand.
The 21-query method organizes research across five categories. Deduplicate candidates, retain sources and keep the intent the offer and destination can support. There is no required output of 1,300–2,500 keywords.
An earlier skincare-account narrative reports CPA moving from $34 to $26.50 while spend increased 40% after this research work. Account/comparison details are not retained here and other changes are not isolated. Treat it as an unverified retrospective, not the protocol's expected effect.
Seven layers of search demand.
A coverage map of the article’s research framework. These are research categories, not a performance ranking.
Source: Published article · this section
Angle-Based Ad Group Architecture
An intent map helps decide what the ad and page must answer:
- "Best Running Shoes": relevant comparison evidence, where available.
- "Running Shoes Sale": the actual offer and purchase conditions.
- "Running Shoes for Flat Feet": product fit without unsupported medical benefits.
- "Nike vs Adidas Running": substantiated differences and reviewed claims.
It does not promise consistent Quality Scores of 8–10. Keep useful signal together; split control or reporting when the decision warrants it.
A possible account map separates four roles:
- Branded Search: defined own-brand demand.
- Prospecting Search: supported non-brand questions and destinations.
- Branded Shopping: defined products and brand-demand reporting.
- Prospecting PMax: a documented acquisition role and applicable brand controls.
These roles do not require four campaigns, manual CPC or a universal phrase/exact split. Choose structure and bidding from objectives, supported controls, economics and measurement needs. Explicit brand/non-brand definitions keep a blended ratio from concealing the question.
The RSA System: Overgenerate and Score
Draft candidates that answer different buyer questions. Twenty-five headline ideas can be an editorial timebox, not a required output or test result.
One internal 100-point rubric is:
- 3-Rule Filter (30 points): visualizable, falsifiable and product-specific.
- Keyword Relevance (20 points): appropriate to the intended query theme.
- Clarity (15 points): understandable without extra explanation.
- Proofability (15 points): supported by retained product evidence.
- Uniqueness (10 points): a different proposition within the reviewed set.
- Reason to Click (10 points): a useful next step for the buyer.
These are editorial weights, not Google's score or a CTR prediction. Reject unsupported claims regardless of the total. The three-rule review asks:
- Visualizable: can the product demonstrate the named feature or situation?
- Falsifiable: is there relevant evidence for any rating, test, performance or certification claim?
- Product-specific: is the difference defensible, beyond a brand name inserted into generic praise?
Authority, a buyer's desired change and value can supply different briefs where supported. They do not require three RSAs, before/after promises or ROI language. Compare the shortlisted messages and proof; a similarity cutoff cannot establish distinct ideas.
Review assets and possible pairings for coherence. Pin where a reviewed requirement justifies a fixed position; do not invent a universal performance penalty.
Google's responsive search ad guidance describes adaptive combinations from up to 15 headlines and four descriptions. Not every asset or position is guaranteed to show.
For illustration, 15 × 14 × 13 × 4 × 3 = 32,760 ordered arrangements could fill three headline and two description positions without repeats. This counts one full-slot model; it does not mean Google generates, serves or tests every arrangement. Pinning, eligibility, format and available space affect what appears.
Asset labels and combination reports describe their actual exposure. They are not controlled experiments; inspect conversion quality and business results before calling a revision a winner.
Negative Keyword Architecture
Exclude demonstrated irrelevant demand at the appropriate scope. More negatives can also block useful queries. Retain the term, reason, match type and affected campaign for each proposal.
Organize candidates by their routing question:
- Irrelevant-demand candidates: jobs, free, DIY, how-to and similar terms can be relevant to some offers. Inspect the full query, product and campaign role before excluding them.
- Competitor-name candidates: decide where comparison demand belongs, then review scope and match behavior.
- Own-brand candidates: define which eligible campaigns should cover the demand. An exclusion neither forces the lowest CPC nor proves nonincrementality.
Check current supported settings and actual query evidence. Same-domain keyword overlap is not an internal bidding war.
Retain the date range, source, conversion delay and reviewer too. A relevant query with no conversions in a short window may have immature outcomes. Exclude demonstrated irrelevance; do not equate an unmatured zero with waste.
Competitive Intelligence: The Messaging Matrix
Capture competitor ads with source, market, date and placement. Twelve categories can organize observations: price/value, quality/premium, urgency, social proof, authority, convenience, guarantee, newness, personalization, comparison, free offer and exclusivity.
State prevalence within that captured set and its denominator. It is not complete market saturation or a profitability forecast. A crowded message can remain useful when your product has a substantiated reason to use it.
Turn comparison into proof questions:
- "Premium quality": what relevant material, test or specification can you demonstrate?
- "Fast shipping": which delivery conditions and markets can you actually state?
- "Lowest price": what genuine specification or ownership-cost difference can you explain?
Check the answers before writing ads. Do not invent a rating, review count, certification or "lasts 3× longer" example.
The original supplement-brand narrative reports ROAS moving from 2.1× to 4.7× over 60 days after messaging work. It retains no account/report source, attribution basis, comparable periods, spend or concurrent-change record here. This is an unverified retrospective, not a verified win, causal effect of the matrix or expected result.
Staged Deployment
Propose a manageable change set with a reviewer, budget/risk boundary, measurement question and rollback path. Let economics and conversion delay determine the review period.
Stage 1: Establish the baseline. Check tracking, value or qualified-lead definitions, destination readiness and selected ready-to-buy themes. Distinguish usable evidence from missing or immature data.
Stage 2: Extend supported coverage. Test a bounded adjacent intent or message while keeping costs and comparisons interpretable. Require useful mature evidence, not a universal 14-day, 20- or 50-conversion rule.
Stage 3: Evaluate discovery demand. Define less direct traffic's role and measurement limits. Platform-attributed ROAS alone cannot establish incremental contribution; expand only within the agreed economics and risk controls.
Stage 4: Review broader matching and scale. Inspect conversion quality, tracking, budget and relevance controls. There is no universal 200-conversion permission slip for broad match or mandatory budget split.
Hold the proposed advance when evidence is incomplete or inconclusive. Calendar time and conversion counts do not approve a change by themselves.
The Weekly Optimization System
Use this illustrative 30-minute triage timebox, assigning deeper work outside it. Select data windows that cover volume, conversion delay and returns rather than automatically judging seven days.
Minutes 1–5: Budget pacing. Compare spend with the agreed budget and period. Check account/billing status, disapprovals and delivery constraints before blaming one setting.
Minutes 5–10: Quality and destination checks. Prioritize query/keyword rows by spend and business context. Inspect relevant Quality Score components, the message and destination, not a weighted account-score target.
Minutes 10–20: Search-term review. Separate supported demand, demonstrated irrelevance and immature outcomes. Record the reason and scope before proposing a negative or keyword change; conversion alone does not require a new exact-match keyword.
Minutes 20–25: RSA and claim review. Inspect assets, combinations and destinations against the buying question. Check each claim's evidence, note limited exposure, and retain the creative revision and review decision.
Minutes 25–30: Competitive check and next actions. Review captured competitors and relevant auction observations. Record what changed, which question warrants investigation and who owns the action.
Produce a short decision log: accepted changes, evidence gaps and assigned follow-through. The timebox guarantees neither a completed optimization cycle nor better performance.
The Complete System in Practice
The account retrospective reports monthly Google Ads spend growing from $21K to $336K over eight months. It describes an earlier 4.2× blended dashboard ratio and 12× brand ratio, plus a later $127K non-brand Search snapshot at 3.8× attributed ROAS.
These are reported snapshots, not a matched measurement tuple or complete contribution ledger. Attribution, comparison windows and concurrent changes are not fully retained here. Spend growth and attributed ROAS do not establish incremental profit.
The narrative describes separating brand/non-brand reporting, researching queries, rebuilding ads/pages and reviewing exclusions. It cannot establish that this architecture caused the change or that copying asset counts reproduces it.
Choose one supported buying question, inspect the ad and destination that answer it, and define the evidence for an accepted change. Let the client's goal and observed constraints determine the work.
Put the framework to work
Map one intent to an ad and page