Google Ads Audit

Before you write anything

Read these first. They are the source of truth and this skill does not invent around them.

1. brand-kit/brand.md: name, palette, fonts, CTAs
2. brand-kit/icp.md: the fit matrix, search terms, competitors, economics
3. brand-kit/positioning.md: the frame, pillars, traps, stats, vocabulary
4. brand-kit/voice.md: house style and banned constructions

If brand-kit/icp.md is missing or still has its template placeholders, stop. These audits score spend against your buyer definition and produce meaningless verdicts without one. Tell the user to run brand-kit-setup.

What this produces

A formatted Word document that goes past surface metrics to diagnose three things: where search spend is going to people who will never buy, where genuine buyer-intent terms are starved of budget, and how the account structure should change.

Every verdict in the document traces back to the brand kit. The Fit Matrix in brand-kit/icp.md defines who counts as a buyer. Search Terms defines what they type. Channel Economics supplies the numbers that turn "expensive" into "too expensive."

When to use

• Monthly or quarterly review of a paid search account
• After a significant spend change, structure change, or campaign launch
• Pre-planning for a new budget cycle
• When search CPA trends look wrong and the cause is not obvious
• Someone uploads Google Ads CSV exports
• Someone wants to pull the account live through a reporting connector

Optional context that improves the audit

Run this after company-recon and competitive-white-space when you can. Recon establishes what the company sells and how it is positioned, which is what makes "high intent" a specific judgment rather than a generic one. White space establishes which category terms are contested and which are open.

If neither has run, the audit still works. It reads the same facts from the brand kit, just without the market context around them.

Guiding philosophy

Apply these throughout. They are the reason the output differs from a platform-generated recommendations tab.

• Intent beats volume. A click from someone searching for a platform in your category is worth many multiples of a click from someone searching for a definition of it.
• Favor searches that signal the buyer wants to buy something. Terms carrying "platform," "software," "vendor," "solution," "pricing," or "for enterprise" are the ones that convert in long B2B cycles. Read Sales cycle length in Channel Economics to know how long "long" is for this company.
• Protect brand efficiently. Every term in Brand terms under Search Terms should hold near-total impression share at the lowest achievable CPC. The floor is Brand impression share floor in Channel Economics.
• Capture competitor demand deliberately, not reflexively. Work from Competitor and alternative terms in Search Terms and the Competitors table. Fund these only once brand and core category coverage are solid, because they are the most expensive clicks in the account.
• Structure and keyword hygiene matter more than clever ad copy. A well-built account with average creative outperforms a badly built one with excellent creative.
• Search captures demand that already exists. It cannot tell you who is behind the click. That limit is not a flaw to fix inside the account. It is the reason a paid search audit has to end with a demand-creation question rather than a bid adjustment.

Data input

Three ways in, in order of preference. All three produce the same three data sets.

Option 1: A reporting connector

If the session has a connector or data integration that reads Google Ads (a marketing data warehouse, an attribution tool, the Ads API directly), pull the data live. Ask which connector is available rather than assuming one. Confirm the account ID before pulling.

The field names below are Google Ads API vocabulary and are the same regardless of which tool fronts them. Adapt the request syntax to whatever connector is present.

Pull 1: Campaign performance overview

fields: ["campaign_name", "campaign_status", "advertising_channel_type", "spend", "clicks",
         "impressions", "conversions", "ctr", "average_cpc", "cost_per_conversion",
         "search_impression_share", "search_budget_lost_impression_share",
         "search_rank_lost_impression_share"]
date_range: last 3 months

Pull 2: Search term report, this is where the waste lives

fields: ["search_term", "campaign_name", "spend", "clicks", "impressions", "conversions", "ctr"]
date_range: last 3 months
filter: spend > 5

Pull 3: Keyword performance

fields: ["keyword_text", "campaign_name", "ad_group_name", "keyword_match_type", "spend",
         "clicks", "impressions", "conversions", "ctr", "average_cpc", "quality_score",
         "search_impression_share"]
date_range: last 3 months
filter: spend > 5

Pull 4: campaigns that have no search term report

Read advertising_channel_type on every campaign in Pull 1 first. Performance Max, Demand Gen, Display, Video, and Shopping campaigns do not produce a search term report in the sense Pull 2 assumes. Performance Max and Demand Gen in particular serve against queries and audiences the term report never shows you, so their spend must be separated out before any waste percentage is computed rather than quietly folded into the denominator.

For those campaigns, pull what does exist:

fields: ["campaign_name", "advertising_channel_type", "asset_group_name", "spend", "clicks",
         "impressions", "conversions", "cost_per_conversion"]
date_range: last 3 months

And ask the connector for the search terms insight report, which returns grouped search categories rather than individual terms for these campaign types:

fields: ["campaign_name", "search_category_label", "impressions", "clicks", "conversions"]
date_range: last 3 months

If that report is unavailable through the connector, say so and treat the spend as uninspected rather than clean.

Also pull monthly totals across the trailing six months for the trend section.

Option 2: CSV export

The user exports the campaign report, the search terms report, and the keywords report from the Google Ads UI and uploads them. Parse with pandas or the csv module. Column headers vary by locale and by which columns the user selected, so map headers to the field names above before analyzing rather than assuming positions.

Option 3: Pasted data

The user pastes numbers into the conversation. Structure them into the same three data sets first. If a data set is missing, say which sections of the audit you cannot produce instead of estimating around the gap.

Analysis workflow

Step 1: Establish the buyer context

Before touching a number, write down four things from the brand kit:

1. What the company sells, from Hallway Pitch and What We Do.
2. Who the buyer is, from The Fit Matrix: the ICP column across seniority, company size, job function, and industry.
3. What high intent looks like, from Category terms and Competitor and alternative terms in Search Terms. High intent means the searcher is trying to evaluate, pilot, or buy something in this category.
4. What a non-buyer looks like, from Negative terms in Search Terms plus the Non-ICP column of The Fit Matrix.

This context drives every verdict in the document. Without it the audit is arithmetic.

Step 2: Campaign performance diagnosis

For each campaign, assess four things:

• Spend efficiency. Compare CPA against Acceptable CPA band and CPA red line in Channel Economics. Those numbers are for a sales-qualified lead, not an MQL, so check what the account is actually counting as a conversion before you compare.
• Impression share gaps. Separate budget-lost from rank-lost. Budget-lost is a funding decision. Rank-lost is a relevance or bid problem. They have different fixes and conflating them produces a wrong recommendation.
• Intent alignment. Does the campaign name and structure map to one clear intent tier: brand, competitor, category, or solution?
• Status. Active versus paused, and whether anything paused should be revived or killed outright.

Step 3: Search term waste analysis

The highest-value section. Categorize every search term with meaningful spend, after two scope checks that decide what "every" covers.

Scope check one: which campaigns are in. Only campaigns that produce a search term report. Everything from Pull 4 sits outside this analysis and is handled in Section 7 under Campaigns with no search term report.

Scope check two: how much of the in-scope spend has a term. Run the reconciliation in Section 4 before classifying anything. If a third of search spend has no term attached, every conclusion in this step describes two thirds of the account, and the document has to say so in that many words.

Terms taking money that should not be:

• Informational and educational intent: definitions, "what is," "how does it work"
• Research-stage curiosity with no buying signal
• Job-seeker intent: anyone searching to be the role rather than to buy for it
• DIY intent where a platform purchase is unlikely: open-source, build-your-own, tutorials
• Navigational queries for other companies' products, including logins and support pages
• Misspellings, foreign language, and gibberish
• Anything matching Negative terms in Search Terms
• Tangential terms outside the category in Category we compete in

For each waste category: explain why it is misaligned, quantify the spend, and recommend a negative keyword theme rather than a list of one-off exclusions.

High-quality terms that budget is starving:

• Terms carrying intent modifiers: platform, software, enterprise, solution, pricing, vendors, demo, comparison, "vs," "alternatives to"
• Terms with high CTR and low impression share
• Terms with a good conversion rate and a "limited by budget" signal
• Any term appearing in Category terms or Competitor and alternative terms that the account is either underfunding or missing entirely

Step 4: Account structure assessment

• What is working. Clean intent separation, logical ad group grouping, brand isolated in its own campaign.
• What is broken. Mixed intent inside one campaign, core and exploratory keywords competing for the same budget, broad match doing too much work, brand cannibalizing non-brand, competitor terms buried inside generic campaigns.
• Match type hygiene. Whether exact, phrase, and broad are being used for the jobs they are good at.
• Campaign types that are not search. Whether Performance Max or Demand Gen is running alongside search, what share of budget sits there, and whether it is cannibalizing brand search or covering genuinely different demand. Check whether brand terms are excluded from Performance Max, since without an exclusion it will absorb the cheapest brand clicks in the account and report them as its own conversions.
• Budget allocation. Money should flow in this order: protect brand, fund solution-intent, then competitor, then category exploration. Check it against Monthly paid budget in Channel Economics.

Step 5: Strategic recommendations

• Budget reallocation, stated as a from-and-to with a monthly dollar amount
• Negative keyword themes to add immediately
• Missing keyword classes, checked against Category terms, Competitor and alternative terms, and the trap names in Named Traps in brand-kit/positioning.md
• Structure changes: what to split, merge, or rebuild
• The demand gap, which is Section 9 below

Document structure

Eleven sections. Do not force all eleven out of a thin account. If the account is small or new, say so and compress.

Document design

Read the Palette, Typography, and Document Footer sections of brand-kit/brand.md and use them. Never hardcode a color or a typeface into the document.

Page size US Letter, one-inch margins.

If Palette still has blank placeholders, stop and tell the user to fill it in. Do not substitute defaults.

Section 1: Executive dashboard

Page one, and it sets the tone for everything after it. Build a two-by-three grid of KPI tiles as a table with colored cell backgrounds. Each tile carries one large metric value (20pt bold), a label (9pt bold), and a one-line subtitle (8pt italic).

Below the tiles, three to four sentences summarizing the finding, not the method. One of those sentences states the coverage of the audit: what share of account spend sits in campaigns with a search term report, and what share of that spend had a term attached to it. A reader who does not know the denominator cannot judge the numerator.

Section 2: Performance trends, six months

| Month | Spend | Clicks | Conv | CPA | Trend | Notes |

The Trend column uses directional words: Baseline, Improving, Declining, Stable, Best Month, Scaled Up. Bold the significant ones. Notes carries brief context: a structure change, a match type shift, a seasonal effect.

Below the table, two or three sentences naming the single most important trend. Tie the change to a cause when the data supports one, and say the cause is unclear when it does not.

Section 3: Account overview

| Campaign | Type | Status | Spend | Clicks | Conv | CPA | CTR | Imp Share |

Type carries the channel type from Pull 1. Impression share is blank for campaign types that do not report it, and blank is the honest cell there. Two or three sentences of interpretation below it, including what share of account spend sits in campaign types this audit cannot inspect at the term level.

The benchmark column is conditional, and it is never yours to fill.

Before building this table, ask the user one question: do you have a benchmark source you want this account measured against, and can you give me the source name and its publication date?

• They supply one. Include their column, use their figures unchanged, and put the source name and date in a footnote line under the table so a reader can check it. Add the Verdict column beside it, a short bold call: "On track," "Above benchmark," "Critical gap." If they offer figures without a source or without a date, ask for both. If both do not arrive, treat it as no benchmark.
• They do not. Omit the benchmark column and the Verdict column entirely, and write under the table: no external benchmark source was available, so these figures are reported against this account's own trailing distribution rather than an outside reference. Say it plainly. An account measured against its own six months is more defensible than an account measured against a number nobody can source.

You may not supply a benchmark figure or a citation from your own knowledge. Not a CTR, not a CPC, not a CPM, not a conversion rate, not a "typical B2B software account runs around," and not a source name with a year attached to it. This document tells the reader on its own writing-rules page that it never invents a number. A remembered benchmark with a remembered citation breaks that promise in the one place an executive is most likely to check, and it is the single most likely way this audit embarrasses the person presenting it.

Metric table, always:

| Metric | This account | Trailing six-month median | Direction |

Rows: brand CTR, non-brand CTR, brand CPC, non-brand CPC, brand impression share, conversion rate. Brand impression share is the one row with a real target regardless: compare it to Brand impression share floor in Channel Economics, which is the company's own number and not a benchmark.

Section 4: Spend waste analysis

The money section. Before any waste layer, reconcile the search term report against campaign spend, because the term report is not complete and treating it as complete overstates how much of the account you have actually inspected.

Google withholds search terms that too few people searched, so a share of spend in every search campaign arrives with no term attached to it. Your own spend > 5 filter on Pull 2 removes more rows on top of that. Both gaps have to be measured and shown, and they are different things: one is a privacy floor you cannot cross, the other is a choice you made.

Run the reconciliation in this order and show it as the opening rows of the waterfall:

total account spend                      from Pull 1, all campaigns
  minus spend in campaigns with no
        search term report               PMax, Demand Gen, Display, Video, Shopping
= search campaign spend                  the only spend this section can judge
  minus spend on reported terms          the sum of Pull 2 before filtering
= unattributed search spend              withheld terms plus anything below the pull filter

Split that last line into two rows where you can: spend on terms your own filter dropped, which you can recover by rerunning Pull 2 without the filter, and spend Google never attributed to a term, which you cannot recover at all.

The unattributed share is a reported number, not a rounding difference. It goes in the waterfall as its own row, tinted Background soft rather than Problem, because it is unexamined rather than wasted. It goes in the callout under the waterfall. And it goes in the sentence that names the waste percentage.

Then the waste waterfall itself, which starts at the spend you can actually see terms for and strips each waste category away layer by layer, ending at effective working spend.

| Layer | Amount | % of search spend | % of total spend | Running total |

Fill waste rows with the Problem tint, the final working-spend row with the Positive tint, the opening total row and the unattributed row with Background soft.

Two percentage columns, and both are required. A waste figure quoted against attributed spend and read as a share of the account is the mistake this table exists to prevent.

Below the waterfall, a problem-colored callout with the waste percentage, the share of spend the analysis could not see, and the annualized figure. Annualizing is what makes a monthly number land with an executive.

Rules on projecting the waste rate. Waste percentages are computed on attributed spend only. Do not scale the measured waste rate up onto unattributed or non-search spend to reach a bigger headline number: that assumes the invisible spend behaves like the visible spend, and low-volume queries are exactly where it is least safe to assume that. When unattributed spend exceeds 30% of search spend, say so directly and give the recoverable figure only as a floor, since a third of the search account was never in the sample.

Then list each waste category as a short block: the category name, the dollar amount, three or four real search term examples from the data, and one line on why that intent does not convert for this buyer.

Section 5: Top 25 search terms

Sort by spend descending, take the top 25. State above the table which campaigns are in scope and what share of account spend that represents, so nobody reads the top 25 as the top 25 of the whole account.

| # | Search term | Spend | Clicks | Conv | CPA | Conv % | Verdict |

Color the verdict cell: Positive tint for KEEP, Caution tint for WATCH, Problem tint for CUT.

Classification is intent-first, not conversion-first.

The primary question for every term is this: is the person behind this search plausibly someone in the ICP column of The Fit Matrix, looking to evaluate, pilot, or buy in this category?

If yes, the term is KEEP regardless of whether it has converted in ninety days. Read Sales cycle length in Channel Economics. When the cycle is measured in months, a ninety-day window of zero conversions on a strong-intent term tells you nothing about the term. It tells you to investigate the landing page, the offer, or the ad copy.

KEEP, buyer intent is present:

• Anything in Brand terms under Search Terms, including founder and executive names. Always KEEP, regardless of performance.
• Anything in Category terms carrying an enterprise or scale modifier.
• Terms combining "platform," "software," "vendor," "solution," "enterprise," or "for business" with the core category from Category we compete in.
• Terms combining "pricing," "demo," "alternative," "vs," "alternatives to," or "comparison" with a name from the Competitors table. This is comparison shopping and it is late-stage.
• Competitor navigational terms only when paired with a qualifier that signals switching intent, matching the patterns in Competitor and alternative terms.
• Any term with conversions at any CPA. Conversion validates intent.

WATCH, ambiguous intent, monitor for six months:

• Broad category terms with no platform or solution modifier. Could be buyer research, could be idle curiosity.
• Generic single-concept queries naming the technology without naming a need.
• How-to and troubleshooting terms. Frustrated practitioners sometimes become buyers and more often stay DIY.
• Anything where intent genuinely splits close to even between buyer and non-buyer.

CUT, clearly not a buyer:

• Navigational searches for other companies' products: logins, support portals, training sites, documentation.
• Foreign language, misspellings, gibberish.
• Job and hiring searches where the searcher is the role rather than buying for it.
• UTM strings and other technical artifacts leaking into the term report.
• DIY and build-it-yourself tooling searches.
• Purely academic intent: research papers, coursework, thesis topics.
• Anything matching Negative terms in Search Terms.

The rule that matters most: a term with heavy spend and zero conversions is not automatically a CUT. If the intent behind it is "I am looking for a platform in this category," that is a landing page problem or an offer problem, not a targeting problem. Cutting it hides the real failure and removes the demand at the same time. Only cut when the intent itself is wrong.

Verdict summary table below the term table:

| Category | Terms | Spend | Notes |

KEEP, WATCH, and CUT totals with the reasoning for each. Close with a paragraph on the intent quality distribution and which WATCH terms to revisit in ninety days.

Section 6: High-intent terms that are underserved

Terms that convert, or should, and cannot because of budget or absence:

• Terms below 50% impression share carrying a strong intent signal
• Terms converting at a high CPA because waste terms are bidding up the same auctions
• Missing keyword classes: solution-level and competitor-alternative terms absent from the account entirely
• A coverage check against Named Traps in brand-kit/positioning.md. A trap you have named is a phrase buyers eventually search. If the traps are not in the account as keywords, say which ones to test and at what match type.

Section 7: Account structure assessment

• What is working, and should be left alone
• What is broken, with the specific campaign named
• Recommended structure changes
• Match type recommendations
• Budget reallocation:

| From | To | Amount / mo | Rationale |

Campaigns with no search term report. Give Performance Max, Demand Gen, Display, Video, and Shopping their own block here, sized by spend, and say up front that the term-level verdicts in Sections 4 through 6 do not cover them.

| Campaign | Type | Spend | % of account | Conv | CPA | What the data does and does not show |

For each one, work with what exists rather than pretending the search term analysis reached it:

• Performance Max. Report spend, conversions, and CPA against Acceptable CPA band. Where the search terms insight report is available, report the search categories by impressions and conversions, and say clearly that these are grouped categories, not search terms, so they support a direction and not a KEEP, WATCH, or CUT verdict on any one query. Check that brand terms are excluded, and check whether conversions here overlap with what the brand search campaign was already converting. Where asset group detail exists, name the weakest asset group by CPA.
• Demand Gen. Same treatment. It buys audiences, not queries, so judging it on search intent is a category error. Judge it against its own role in the funnel, and against Sales cycle length in Channel Economics before calling a high CPA a failure.
• Everything else. Report spend and conversions, and name it as uninspected.

Recommendations for these campaign types stay at the level the data supports: exclusions, budget, conversion action selection, and asset quality. Do not recommend a negative keyword list as if it worked the way it does in a search campaign, and confirm what the account actually supports before recommending term-level exclusions at all.

Section 8: Quick wins with projected impact

A table, not a numbered list. Seven to ten specific actions.

| Action | Est. savings / qtr | Est. add'l conv | Effort | Confidence |

Effort is a real estimate: "5 min," "Low," "Medium," "1 hour." Confidence is High, Medium, or Low in bold. A quick win with Low confidence is still worth listing; hiding the uncertainty is what makes an audit lose credibility on the second read.

Below the table, a Positive-tinted callout with the combined projected impact: total savings range, total additional conversions range, and the note that none of it requires additional spend.

Section 9: The demand gap

Paid search captures demand that already exists. It cannot create it, and it cannot tell you who is behind the click. This section quantifies both limits and says what follows from them.

It is also the section most likely to be built on numbers nobody can source. Sizing category demand requires search volume for terms the account does not bid on, and that volume is in none of the pulls above. Decide which version of this section you are writing before you write a line of it.

The inputs, and where each one comes from

Ask for all of these in one message, by name. Do not substitute a realistic, typical, or industry-standard rate for the account's observed rate. A rate with no source is an invented benchmark, and in this section it drives the headline conclusion. The account's own non-brand CTR and conversion rate are the only rates allowed here, and each one is labeled in the document with the pull it came from.

Version A, when the inputs arrive

Write the chain out in the document, one line per step, each line labeled with its source:

monthly searches across Category terms      Keyword Planner export, date of pull
x observed non-brand CTR                    this account, last 3 months
= reachable clicks per month
x observed non-brand conversion rate        this account, last 3 months
= reachable conversions per month
x stated lead to closed-won rate            supplied by the user
x Average deal size                         Channel Economics
= reachable revenue per month at full coverage

Three rules on this arithmetic:

1. Run it twice, once at the observed conversion rate and once at half of it, and report the result as a range. A single number here reads as a forecast and it is not one.
2. Full coverage is a ceiling, not a plan. No account holds all available impressions on a category. Present the top of the range as the most this channel could ever return, then say what the account holds today from its own impression share figures.
3. Compare the ceiling to the stated target, not to ambition. When the ceiling sits below the target, the gap is not a bidding problem. No budget increase closes it, because the searches do not exist yet. State that in one sentence with both figures next to each other.

Version B, when the inputs do not arrive

Say in the document that category-level demand could not be sized, and name the one export that would size it. Then report what account data alone supports, which is real and useful:

• Headroom inside the current keyword set. For each campaign and each significant keyword, eligible impressions are impressions divided by search impression share. The gap between eligible and served is demand the account is already eligible for and not getting.
• Why that headroom is unserved, split into budget-lost and rank-lost impression share. One is a funding decision, the other is a relevance or bid problem.
• Coverage of Category terms, as a count: how many of those terms have no keyword in the account at all. A count, never a volume estimate.

Then state the limits plainly, in the document, not in a footnote: this analysis covers only terms the account already bids on. It cannot say how many people search the category, whether category demand is large enough to hit the target, or whether the ceiling sits above or below it. Those questions need the Keyword Planner export, and until it arrives the honest answer is that the account has measurable headroom of a stated size and an unmeasured category around it.

The three questions that follow, in either version

1. How much of the traffic is out of ICP. Estimate from the search term mix, since the platform will not tell you the searcher's title or company size. Say plainly that this is an estimate from intent signals rather than a measured figure, and that it covers only the spend with a term attached to it.
2. What a demand gap implies, where Version A found one. A demand gap is a demand-creation problem. The buyers exist, they have the problem described in Named Traps, and they are not yet searching for the category because they have not named their problem the way you name it. Reaching them requires targeting by who they are rather than by what they typed. Where Version B is what ran, this conclusion is available as a hypothesis and must be written as one.
3. Which awareness and consideration gaps search structurally cannot fill. Name them against the pillars and traps in brand-kit/positioning.md. This one holds in both versions, because it does not depend on volume data.

Write this as analysis that follows from the data. It is a strategic conclusion, not a pitch for another channel, and it should read as one.

Section 10: Cross-channel integration summary

Include this section only if a LinkedIn ads audit has been run in the same conversation. Otherwise skip it and renumber. Do not invent LinkedIn findings.

| Dimension | Google Ads | LinkedIn Ads |

Rows: primary role, biggest strength, biggest gap, retargeting, total waste, priority fix.

Below the table, a unified recommendations paragraph with the combined recoverable figure and total projected additional conversions across both channels.

Section 11: Key takeaways

Five to seven executive bullets: account health, total recoverable waste with the share of spend the term analysis could not see, the KEEP/WATCH/CUT split across the top 25, the share of budget sitting in campaign types with no search term report, the top structural change, the demand gap finding and which version of Section 9 produced it, and the next three moves in order.

Writing rules

Follow brand-kit/voice.md in full. It governs. These are the audit-specific additions:

• Confident and consultative. This is written for a CMO or a VP of Demand Gen.
• Quantify everything. Dollars, percentages, impression share points.
• No hedging on findings. If something is broken, say it is broken.
• Short paragraphs, two or three sentences, or bullets.
• Tables for anything data-heavy.
• Never invent a number. Every figure traces to the account data or to the brand kit. If a figure is an estimate, label it as one in the same sentence.
• Never supply a benchmark from memory. No CTR, CPC, CPM, CPA, conversion rate, or impression share presented as an industry, category, or platform norm may come from your own knowledge, and no citation may be attached to one. A benchmark enters this document only when the user hands you the figure, the source, and the date. Where none is supplied, the comparison is against this account's own history and the document says so in a sentence. This rule outranks the wish to fill a column.
• No source links in the document body. A user-supplied benchmark gets a footnote line under its table.
• Bold sparingly. Everything bold is nothing bold.

Word document output

Generate the document with the docx skill available in the session. Reference it by name and let the host resolve it. Follow its rules in full, and specifically:

• US Letter page size
• Real bullet list formatting, never unicode bullet characters typed into a paragraph
• Tables need both column widths and per-cell widths set, in DXA
• Cell margins: top 80, bottom 80, left 120, right 120
• Override the built-in Heading1 and Heading2 styles rather than styling headings inline
• Never put a newline character inside a paragraph. Use separate paragraphs.
• Validate the file after generating it, using whatever validation step the docx skill provides

If no docx skill is available in the session, output the full audit as clean markdown, preserving every table, and tell the user plainly that a Word document could not be generated and why. Do not silently downgrade the deliverable. If brand-visuals is installed, offer to turn the markdown into a branded PDF instead. Nothing in the analysis changes; only the format does.

Output notes

• After generating the document, write three or four sentences in chat: total waste found, the single biggest issue, and the top recommendation. Then present the file.
• Do not reproduce the document in the chat.
• If the account is small or newly launched, say so and shorten the audit. Eleven sections forced out of a thin account reads as padding and undermines the sections that matter.