fm.

B2B SaaS demand generation case study: cost per SQO down 74% in two quarters at Pipe17

Rebuilding a B2B demand engine from the CRM up, then finding the pipeline source nobody was measuring: AI assistants.

Role Growth Marketing LeadCompany AI-native OrderOps platformChannels Google Ads, LinkedIn Ads, HubSpot, AI search
74%lower cost per SQO, $14,450 to $3,700
2xMQL to SQL conversion, 20% to 40%
$1.3Mpipeline influenced by $66.5K in paid in one quarter
$960Kclosed-won from AI-referred buyers, 32% of new-business wins

The short version

  • Farhan Munshi ran growth marketing at Pipe17, a B2B SaaS company that sells an OrderOps platform to commerce brands and 3PLs. He owned Google Ads, LinkedIn Ads, HubSpot and attribution.
  • Cost per sales-qualified opportunity (SQO) fell 74%, from $14,450 to $3,700, in two quarters. In one quarter, $66.5K in paid media influenced $1.3M in pipeline.
  • MQL-to-SQL conversion doubled, from about 20% to 40%, after I rebuilt lead scoring and routing with Sales. More than 60% of inbound had been unqualified SMB.
  • I moved Google Ads to 100% non-brand conquest and integration campaigns, fixed a conversion-goal setup that had zeroed out reporting, and proved a 4.0x ROAS with a six-month deal-level audit.
  • AI-referred buyers turned out to be $960K in closed-won revenue, 32% of new-business wins. The CRM had credited all of it to paid search.
  • A 12-agent research pipeline grew the target account list from 441 to 2,795 accounts, and the tools I built with Claude Code and Codex replaced about $24K a year in software.

01The situation

Pipe17 sells order operations software, a category it calls OrderOps, to commerce brands and third-party logistics providers (3PLs). The platform sits between sales channels like Shopify, Amazon and TikTok Shop, the warehouses that ship the orders, and back-office systems like NetSuite. Deals are large and take six months or more to close, so a form fill tells you very little about whether an ad worked.

When I joined, inbound volume looked healthy, but too little of it became qualified pipeline, and there was no CRM I could trust to explain why. Cost per sales-qualified opportunity had climbed to $14,450. More than 60% of inbound leads were small businesses that would never buy. One source report credited paid search with 0.5% of revenue, an influence report said marketing had touched 55% of it, and about three quarters of deals listed their source as "offline". Nobody could say which number to believe.

02Results at a glance

MetricBeforeAfter
Cost per sales-qualified opportunity$14,450$3,700, 74% lower, in two quarters. $66.5K in paid media influenced $1.3M in pipeline that quarter
MQL-to-SQL conversionAbout 20%40%
Marketing-sourced SQOs per quarter1217 the next quarter, then 27
Google Ads ROAS on closed dealsNot measured4.0x over six months, with one keyword at 38.2x
Google Ads spend on brand termsAbout half the budgetNone. The account ran 100% non-brand
Non-brand search clicks75% on irrelevant searchesIrrelevant clicks down about 85% after the move to exact match
Conversions column on non-brand campaignsZero, whatever the lead flowConversions recorded in three separate campaigns within about two weeks of the fix
Target accounts loaded into Google AdsNone9,818 contacts at a 92% match rate
AI-referred revenue the CRM could see7 contacts and no deals$960K closed-won, 32% of new-business wins
Net-new target account list4412,795
Companies with no type in the CRM13,000+Classified for account tiering
Marketing softwarePaid tools for form routing and enrichmentMy own tools, about $24K a year saved

03Building HubSpot from zero and migrating from Salesforce

The first job was a CRM that Sales and Marketing would both believe. I built HubSpot from zero in six weeks and migrated 3,000+ contacts from Salesforce, with lifecycle stages, scoring, routing, nurture and closed-loop reporting. Then I kept adding the parts that connect an ad click to a deal.

  1. One definition of an SQO.A deal becomes an SQO the moment it enters the first sales stage, Technical Evaluation. I rebuilt the quarterly count from stage-entry dates because the default report undercounted. One quarter read 18 when the real number was 21.
  2. Routing with no gaps.Every new contact without an owner went to the account's owner if there was one, and round-robin across the sales team if there wasn't.
  3. Forms that keep the click.Every website form posted server-side to the HubSpot Forms API with the visitor's tracking cookie, first-touch UTMs, landing page and Google click ID. The click ID is what lets a closed deal find its way back to a keyword.
  4. Reporting people see.A weekly funnel report used one written definition each for MQL, SQL, SQO and pipeline. A Slack alert checked every five minutes and posted each new inbound SQO with its channel, and a summary went out twice a week.
  5. Imports that don't overwrite.Every list was deduplicated and checked against existing records first. One enrichment run added 516 net-new contacts without overwriting an existing record.
  6. Intent on the account record.I connected Propensity intent scores to HubSpot as 12 synced contact and company properties, so reps could see account intent without leaving the CRM.

04Fixing lead scoring so Sales trusted MQLs

More than 60% of inbound was unqualified SMB, and the score explained why. An MQL was anyone who crossed 100 points, and a page view was worth the same 20 points as a demo request. A busy visitor who would never buy crossed the line before a good-fit buyer who had visited once. About 70% of MQLs got there on engagement alone and had never asked to talk to anyone.

  1. Hand-raisers only.A hot MQL became a demo request or a contact form and nothing else. Everyone else went to nurture.
  2. Three tiers, agreed with Sales.Tier 1 is hand-raisers and accounts showing intent, worked by account executives. Tier 2 gets SDR outreach. Tier 3 gets marketing nurture. Four outbound segments (3PLs, brand aggregators, Shopify Plus brands and NetSuite users) each got messaging that matched the ads they were seeing.
  3. Qualify in the ad.The top reason leads were disqualified was company size. So every search ad said who the product is for: high-volume brands.
  4. Audit the rejected pile.I joined 820 recorded sales calls to 999 disqualified leads. 94 leads had a call on record, and 35 records across 30 companies turned out to be worth another look.
  5. Check response before blaming targeting.31 of 95 disqualified paid leads were marked no-response. That's a speed-to-lead problem as much as a media problem.

MQL-to-SQL conversion went from about 20% to 40%. I wrote up the general method in my note on MQL-to-SQL conversion rate.

05How I cut cost per SQO 74% in two quarters

Cost per SQO was $9,875, then $14,450 the quarter after. It came down to $3,700. In that quarter $66.5K in paid media influenced $1.3M in pipeline and produced 23 qualified demo leads, a record. Marketing-sourced SQOs went from 12 to 17 to 27 across three quarters.

  • A different scoreboard. I stopped reporting form fills and cost per form fill. The numbers that mattered were cost per SQO and pipeline per dollar, both read from HubSpot.
  • Spend pointed at accounts that buy. I uploaded 9,818 contacts from Tier 1 and Tier 2 target accounts to Google Ads as a Customer Match audience, got a 92% match rate, and bid 30% higher on them. Customers and past converters were excluded from every non-brand campaign.
  • Less waste. Non-brand keywords moved from phrase match to exact match, and the negative keyword list grew past 1,500 terms.
  • Each channel got one job. LinkedIn ran against the 3PL target account list and brought in 17 qualified 3PL leads that quarter. Google got integration-specific campaigns and influenced 10 SQOs.

One caveat a peer should know. The $3,700 counts opportunities that paid media influenced. Strict first-touch gives a higher cost, which is why I reported both views side by side (see the attribution section below).

What the search terms showed

About half the Google budget went to our own brand name. On the generic campaigns, 75% of clicks came from searches that had nothing to do with the product, because phrase match on "order management software" matched all kinds of unrelated software. GA4 agreed: non-brand paid search bounced 99% of the time with a 1.5 second average session. Brand traffic bounced 33% and stayed 226 seconds.

Match types, negatives and audiences

  1. Phrase to exact.I paused 19 phrase keywords and created 19 exact versions. Irrelevant clicks dropped about 85%.
  2. Negatives by hand.I had installed a script that added negative keywords to a shared list every morning. I removed it when it became a risk to the new integration keywords, and went back to manual search-term reviews.
  3. Audiences in observation mode.One new campaign sat at zero impressions for about 11 days. The target account list had been attached without an explicit observation setting, so Google restricted the campaign to only those people. Once fixed, it served within 24 hours.
  4. Ad schedules back to all hours.I had put the non-brand campaigns on business hours. Then the data showed brand converting at midnight, 3am and 10pm, so everything went back to running around the clock.

The rebuild

Then I rebuilt the account around non-brand demand. Six new campaigns: one conquest campaign against legacy order management vendors, one against integration-platform tools, integration campaigns for NetSuite, Shopify and TikTok Shop searches, and one for category terms like order orchestration and retailer EDI. With a small protected category campaign and a remarketing campaign, that made eight campaigns and 29 non-brand ad groups.

  • Exact and phrase match only. An earlier conquest attempt had failed on broad match drifting into unrelated searches and our own brand name.
  • Brand-name negatives on every non-brand campaign.
  • Keyword matched to page. NetSuite searches land on the NetSuite integration page. Conquest traffic lands on a comparison or category page, never straight on the demo form.
  • United States and Canada, people present in the country, no search partners and no display.
  • Target accounts and site visitors attached as observation audiences. Customers, past converters and employees excluded.
  • Fulfillment-flavored searches stay out. People typing those want a 3PL to ship their orders. They aren't shopping for software.
  • Judged on SQO dollars in HubSpot, never on lead counts.

Two things caught me in the first week. The shared negative list was silently blocking 51 of the 146 new keywords, and two ad groups were completely dead, because old competitor-name negatives were still buried in it. I removed 35 entries, then ran a 619-keyword serving audit that cleared 16 more. And ads that name a partner platform run with limited approval, so each one needs a clean fallback ad next to it.

The conversion-goal fix: why Google Ads showed zero conversions

HubSpot showed 12 new contacts with a Google click ID in one month, 11 of them legitimate. Google Ads showed zero conversions. The tags were fine and so was the offline upload. The problem was configuration: the demo and contact form goals were set as secondary at the account level, so form fills only ever landed under All conversions. On top of that, six of eight campaigns were bidding on clicks or impression share, so nothing was optimizing toward a form fill.

I made both goals primary and moved all eight campaigns to conversion-based bidding the same day, seven on Maximize Conversions and one on Maximize Conversion Value. Over the next three weekly checks the conversion count on enabled campaigns went from 0 to 1 to 3, spread across three different non-brand campaigns.

Turning brand off to see what it was worth

I paused the brand campaign and compared the 22 days after with the 22 days before. Leads carrying a Google click ID went from 9 to 7. Non-brand leads rose from about 1 to 4. Organic contacts went from 13 to 15 and direct from 16 to 19. Total lead flow barely moved, which is what you'd expect if most brand searchers were going to convert anyway. It's a small sample, and it's the cheapest incrementality test a B2B account can run.

The deal-level ROAS audit

With a six-month sales cycle you can't judge paid search in 90 days. So I joined six months of Google Ads spend to HubSpot deals through the click ID, keyword by keyword. The window held 90 paid search contacts. The account returned 4.0x on closed deals. One keyword returned 38.2x. Five keywords had spent real money and produced nothing, and I cut them.

The same join ran in the other direction too. HubSpot held 731+ contacts with a Google click ID, Enhanced Conversions for Leads was on, and opportunity and closed-won stages went back to Google Ads as offline conversions with a value on each, so bidding could see revenue and not just form fills.

07The LinkedIn Ads audit: cutting a channel on deal-level evidence

LinkedIn was the bigger line and nobody had checked it against deals. I traced a full year of LinkedIn contacts into HubSpot. The channel had spent 88% more than Google, created one deal, and that deal closed lost. Google created ten deals in the same period. A six-figure LinkedIn line had produced no closed revenue.

The reasons were more useful than the verdict:

  • Retargeting an audience that was too small. An estimated half of the spend went to retargeting a site with 5,000 to 10,000 monthly visitors. Retargeting needs something closer to 50,000.
  • Leads that fell through routing. Of 65 LinkedIn contacts, 27 (42%) never had an owner assigned and another 30 (46%) had no follow-up logged. A channel can't produce pipeline from leads nobody is assigned to call.
  • A broken handoff from Lead Gen Forms. Leads arrived without the form name or company type that downstream workflows keyed on, so the ebook they asked for never sent and the 3PL nurture skipped them. I fixed it with one five-step workflow and backfilled 108 historical contacts.
  • Impressions on the wrong company size. 44% of impressions went to companies with 10,000+ employees, while 59% of opportunities came from companies with 11 to 500.

The budget moved to what did produce revenue. What stayed on LinkedIn was narrow: 3PL lead generation, and a co-published total cost of ownership report shown to a Shopify Plus audience of about 700, which brought in leads from a COO and a supply chain director.

08Marketing attribution for a six-month sales cycle

The CRM held two answers to "what did marketing do?" and they were more than 100x apart. First-touch source said paid search was 0.5% of revenue. A 180-day influence field said marketing had touched 55% of revenue, and 79% in the most recent quarter. Both were true. Leadership dashboards showed the first one.

Three views, never merged

I stopped trying to pick one number. Leadership reporting moved to three views side by side, each with the claim it can support.

ViewQuestion it answersOne quarter of new business
First-touch paid searchWhere did the contact first come from?6 deals
Influenced, 180-day lookbackDid marketing touch the buying group before the deal opened?29 deals
CRM proofIs there a Google click ID on a contact tied to the deal?6 of 6 paid deals

When deal-level source fields broke, I worked from contacts. A contact's first-touch UTM is reliable and so is its association to a deal, even when the deal's own source field is blank.

Replacing the form pipeline

Website forms had been routed to the CRM through Segment, a customer data platform, with Store Leads adding enrichment. I replaced that with a small server-side snippet that posts every form straight to HubSpot, plus a first-party cookie that stores the visitor's first-touch UTMs, referrer, landing page and Google click ID. Before I trusted it I compared a week on each side of the switch: 14 form submissions with 3 paid search and 3 click IDs before, 12 submissions with 5 paid search and 5 click IDs after. Nothing was lost, and all 5 paid search submissions kept their click ID.

A UTM standard people follow

An audit of about 8,000 contacts found that 5% of utm_source values, 11% of mediums and 10% of campaign names followed any convention. I wrote one standard: lowercase, a closed list of 11 mediums, source is always the real platform, and campaign names follow period, initiative, asset. Then I shipped it as a command-line tool with a linter and a hosted link builder, because a rule people have to remember doesn't survive a busy week.

This is the part that changed how I think about attribution. Paid search looked like the best revenue channel in the company on paper. When I read the click paths, 14 of 15 paid-search SQOs had first arrived by searching for the company's own name. So where had they heard the name?

Our forms asked buyers how they heard about us. I read the answers against the CRM source. The share of respondents naming an AI assistant had gone from 6.6% to 16.5% in a year. Four opportunities carried a written answer naming one, and all four were logged as paid search. The ones that closed came to $960K in closed-won revenue, 32% of new-business wins and 80% of the closed-won revenue paid search had been given credit for.

The path is simple once you see it. An assistant recommends the product. The buyer searches the brand name, clicks the brand ad and fills in the demo form. Paid search gets the credit. HubSpot's own AI referral source had captured 7 contacts and no deals. And because that answer was filled in on only about 21% of SQO deals, the real figure is higher than what I could prove.

How I track AI referrals in GA4

  1. One saved report.Filter hostname to the marketing site only, then session source containing chatgpt, openai, perplexity, claude, copilot, gemini and a few others. Report sessions, engaged sessions and conversions by source, then landing pages by source.
  2. A baseline.The first 90 days showed 343 sessions from AI assistants: 203 from ChatGPT, 48 from Gemini, 44 from Perplexity, 35 from Claude and 11 from Copilot, with 8 conversions. That was about 2.7% of Google organic sessions.
  3. A monthly email.A scheduled job reran the report on the first of each month and emailed it to marketing leadership with the change since the last run.
  4. Know the limits.GA4 can't show the prompt. The landing page is the best proxy for what an assistant cited you for. And GA4 later started filing this traffic under its own medium, so a report keyed on referral will undercount.

By my last monthly report, integration pages were the strongest group of AI-referral landing pages after the homepage. I also designed a 20-page freshness test, ten refreshed pages against ten matched controls over six weeks, built on real edits instead of changed dates. I don't have a result from it to report. More on the measurement in how to measure AI visibility.

10AI agents for marketing: the 12-agent account research pipeline

LinkedIn matches 50% to 70% of an uploaded company list. Our first target list had 441 brands, built by one agent working through sources in order. That would match about 250 to 300 companies, too few to run a campaign on. The next day I rebuilt it with 12 agents running in parallel.

  1. Pull the exclusion list.All 26,965 company records came out of HubSpot first, so nothing we already had could come back as new.
  2. Give each agent its own source.Twelve source categories, one per agent: software vendors' customer galleries, investor portfolios, wholesale and marketplace seller lists, trade show exhibitor lists, growth rankings, subscription brands, made-in-USA registries, accelerators, trade publications and recent platform migrations among them. Every agent returned the same columns.
  3. Merge with a script.5,873 raw rows became 4,326 unique domains. When several agents found the same brand, the row kept every source, and a brand found by three or more agents ranked highest.
  4. Exclude and flag.1,324 domains were already in HubSpot. 98 fuzzy name matches went to a file for a person to review.
  5. Tier it.2,795 net-new brands remained, 6.3x the first list: 968 in Tier 1, 857 in Tier 2 and 970 borderline. No paid database was used.

The same approach handled the CRM. 13,316 companies, half the database, had no type, which blocked intent scoring and tiering, so I wrote a HubSpot API script that classified them by industry. A full pull also showed 1,974 good-fit brands sitting in the lowest tier, only 13 of them with any intent score. That was the fastest win available.

The tools underneath

  • Three command-line tools for Google Ads, HubSpot and Clay enrichment. Every write shows a dry run first and reads the live system back after. Batch lookups took HubSpot deal queries from 90 seconds to 3.
  • Approval rules. Agents can read and draft freely. They stop before changing budgets, bids, campaign status, conversion goals, CRM imports, workflows or a live web page.
  • A second agent that doubts the first. For a full business review, 16 agents ran the analysis and a separate set re-derived 30 key numbers from fresh pulls. 24 reproduced exactly, 6 were adjusted and none were refuted.
  • Sales calls as research. A pipeline over 258 external calls, 154 hours in 30 days, pulled out objections and competitor mentions for messaging.

Together with the form pipeline above, this tooling replaced about $24K a year in software. I describe the general setup in AI agents for marketing.

11Trade show demand generation: Manifest and Shoptalk

Enterprise deals in this market often start with a conversation at a conference, so events got the same treatment as paid: a named account list before, clean data after.

  • Manifest. I ran an integrated campaign for about 200 named accounts. A gated commerce operations research report anchored LinkedIn Sponsored Content, paid search, retargeting and SDR outreach. The booth logged about 160 distinct scans. Contacts were enriched in Clay, booth conversations and report downloads got separate HubSpot follow-up, and I tracked outcomes for 90 days.
  • Shoptalk. Before the show we marked 146 target companies, logged 159 meeting interest requests and used AI to filter 555 pages of attendee contacts. I supported the booth design and ran the contact imports and follow-up segmentation.
  • Executive dinners. For a three-city dinner series I built the marketing contact list, 147 contacts across 24 companies, and a sync that reconciled registrations back to HubSpot.

Two rules kept the data honest. A scanned badge is not an MQL, and an invited account is not an attendee. Registration, attendance, booth scans and meetings each stay in their own field.

12Landing pages, an ROI calculator and programmatic SEO

  • Website refresh. I managed the refresh that moved the site's positioning from SMB to enterprise.
  • Landing pages through the WordPress REST API. A paid landing page for order management software searches, a 3PL software page, and gated ebook pages with automated delivery. The generic paid page had been sending visitors to a separate demo page to find a form, one extra click before anyone could convert.
  • ROI calculator. A live calculator with three visible inputs (orders, headcount and software spend), a soft lead gate, HubSpot capture, round-robin follow-up and a Slack notification to Sales.
  • Integration pages as data. I exposed the site's integration page type, 48 fields per page, through the REST API so those pages could be managed in bulk.
  • Programmatic comparison pages. A Next.js and TypeScript system that builds "us vs them" comparison pages from JSON content files, with structured data, a sitemap and source links for every claim. Six comparison routes were built.

13What didn't work

  • Generic category keywords. In one full month the two generic campaigns bought 147 clicks and produced zero conversions. The 3PL software campaign went 90 days without one before I paused it. Exact match fixed the junk clicks. It didn't create buyers.
  • The non-brand rebuild was unproven when I left. The plumbing worked and three campaigns were converting, but volume was small. I can't tell you non-brand search pays at this deal size, only that the account can now measure it.
  • What stayed on LinkedIn. Over the following two quarters the remaining LinkedIn spend produced 32 contacts and no SQOs.
  • Comparison pages never shipped. The six-route system stayed a prototype, and the conquest campaigns needed exactly those pages to land on.
  • Shoptalk. About 100 booth scans, below what we expected from a show that size.

14What I'd tell you

  • Before you trust a zero in Google Ads, check which goals are primary. Compare click-ID contacts in the CRM with the Conversions column for the same month. If they disagree, look at goal settings before you touch a tag.
  • Match the lookback to the sales cycle. With six-month deals, join spend to closed deals at the keyword level over six months. It took a six-month join to find the keyword that returned 38.2x and the five that returned nothing.
  • Ask buyers how they found you, in their own words. One question on a form moved $960K from "paid search" to "an AI assistant told me". No tracking pixel would have found that.
  • Fix routing before you judge a channel. 42% of one channel's leads never had an owner. You can't call the media bad until someone has called the lead.
  • Pause brand for three weeks and count total leads. If the total holds, you were paying for people who were already on their way.

If your pipeline looks like this one did, this is the work I do now: B2B SaaS paid ads, B2B demand generation, Google Ads audits, AEO and AI agent setups for marketing teams.

FAQQuestions people ask

How do you reduce cost per SQO in B2B SaaS?
At Pipe17, Farhan Munshi cut cost per sales-qualified opportunity 74%, from $14,450 to $3,700, in two quarters. He stopped optimizing to form fills, pointed Google and LinkedIn spend at named target accounts, moved non-brand keywords to exact match, built a negative list of 1,500+ terms, excluded customers, and moved budget away from a channel that had produced no closed revenue.
How do you improve MQL to SQL conversion rate?
Fix the definition before the volume. At Pipe17 a page view scored the same 20 points as a demo request, so most MQLs had never asked to talk to anyone. Farhan Munshi made hand-raisers the only hot MQLs, sent everyone else to nurture, agreed three follow-up tiers with Sales and fixed routing. MQL-to-SQL conversion went from about 20% to 40%.
Why does Google Ads show zero conversions when leads are coming in?
Check which conversion goals are set as primary. At Pipe17, HubSpot showed a dozen contacts with Google click IDs in a month while Google Ads reported zero conversions. The demo and contact form goals were set as secondary, so form fills only counted under All conversions. Farhan Munshi made them primary and moved all eight campaigns to conversion-based bidding.
Should a B2B SaaS company bid on its own brand name in Google Ads?
Test it. At Pipe17, Farhan Munshi paused the brand campaign and compared the 22 days after with the 22 days before. Leads carrying a Google click ID went from 9 to 7, non-brand leads rose from about 1 to 4, and organic and direct contacts both ticked up. Total lead flow barely moved, early evidence that most of the brand spend wasn't adding leads.
How do you track revenue from ChatGPT and other AI assistants?
Use two sources together. In GA4, filter session source for ChatGPT, Perplexity, Gemini, Claude and Copilot and report it monthly. In the CRM, ask buyers how they heard about you and read the answers against the recorded source. At Pipe17 that field showed AI-referred buyers were $960K in closed-won revenue, 32% of new-business wins, all credited to paid search.
Is LinkedIn Ads worth it for B2B SaaS?
Only if leads get worked and the audience is big enough. Farhan Munshi's audit at Pipe17 found a six-figure LinkedIn line with no closed revenue. An estimated half of the spend went to retargeting a site with 5,000 to 10,000 monthly visitors, and 42% of LinkedIn leads never had an owner assigned. Fix routing before you judge the channel.
Can AI agents build a target account list?
Yes, if each agent has its own source and the output is deduplicated against your CRM. At Pipe17, Farhan Munshi ran 12 parallel agents across 12 source categories. They returned 5,873 rows, 1,324 were already in HubSpot, and 2,795 net-new brands remained, up from a 441-account list built by a single agent the day before.
What is a good Google Ads ROAS for B2B SaaS?
It depends on deal size and sales cycle, so measure it on closed deals. At Pipe17, Farhan Munshi joined six months of Google Ads spend to HubSpot closed-won deals through the Google click ID. The account returned 4.0x, one keyword returned 38.2x, and five keywords that had lost money were cut.

Notes on how I did it

More work