The Signal-First Google Ads Playbook, 2026

Research cutoff: September 22, 2026
Account window examined: August 23 through September 21, 2026, against the preceding equal 30 days

Evidence legend

Every claim below carries one of six marks; the legend is the key, and the hover text on each mark repeats its label.

• ● Verified fact. A directly documented product mechanic, dated event, or other checkable fact.
• ■ Company claim. A performance or benefit statement from Google or a named advertiser case distributed by the platform.
• ◆ Independent measurement. A third party's analysis of observed data. Independence does not mean randomized or bias-free.
• ○ Anecdote. A practitioner report without enough disclosed data to evaluate causally.
• △ Inference. An interpretation that follows from cited facts but is not itself stated by the source.
• ◇ First-party data. Pulled from the 19 client accounts we managed on September 22, 2026, and from our own lead records, aggregated. No client is named.

Research and access note

● Mechanics were checked against Google's help pages, cited by answer id where they appear. The queries in section 3 were run read-only against our own Google Ads account through API v25 on September 23, 2026; the date is printed beside them because field names move between versions.

◇ Account figures come from one dated pull of the 19 client accounts under our management on September 22, 2026, covering August 23 through September 21 against the preceding equal 30 days, with lead records from our own reporting database. Amounts are in each account's own currency, Canadian dollars except for two US accounts. They are aggregated and no client is named. They describe that day. An account's rung changes the week its pipeline does, so the counts are a snapshot of one fleet on one day.

1. The ladder in one screen

A Google Ads account run for leads sits on one of five rungs, and the rung is set by what the bid strategy was allowed to see in the last 30 days. Budget doesn't set it. Industry doesn't set it. One of the fleet's largest accounts and one of its smallest sat on the same rung in September, because neither had 15 matched graded outcomes a month for bidding to learn from. The rung decides which campaign types the account has earned. Search is available on every rung because it's the one campaign type whose queries can be read. Performance Max, AI Max and Demand Gen are earned, one rung at a time, by evidence that the leads they add are worth having.

T0

Signal untrustworthy

3 of 19 accounts, September 22

T1

Raw lead reliable

12 of 19 accounts

T2

Graded, not yet steering

4 of 19 accounts

T3

Graded and steering

0 of 19 accounts

T4

Valued

0 of 19 accounts

rungs with accounts on themempty on a 19-account fleet

Place your account

Six questions place an account. Answer them from the account's own settings and the lead record, not from memory; question one alone put three of our 19 accounts on the bottom rung in September. The rung you land on is the section of this playbook to read next, and the gate you failed is the work.

2. Why the ladder comes before the buying journey

Two urgent, local, phone-first businesses sat in the same buying-journey branch in September and needed entirely different playbooks. One had logged 245 tracked calls in the month with no click attached to any of them; the other had a single clean form action and a call path that had stopped reporting. A cabinetry reseller and an auto service shop, in different branches by every journey model, needed the same playbook, because they were two of the four accounts among the 19 whose leads were being graded by somebody. What predicted the work was whether anyone was grading the leads, and the journey had nothing to say about that.

None of the 19 had earned value-based bidding. The floor Google documents is 15 conversions in 30 days on the action the strategy bids on, and no graded action among the 19 reached it; the best case matched 12.98 graded contacts in the window. One account came closer by a different route: it had 16.94 matched qualified leads sitting in the all-conversions column and zero in the column bidding reads, because the upload action had never been promoted to primary, so bidding had never been given permission to use them. The top two rungs of the ladder, where value bidding and the expansion campaign types live, were empty on a fleet of 19.

The lag reports can't rescue the journey model either. Every account booked 90 to 100 percent of counted conversions in the under-one-day bucket, and the figure says nothing about sales cycles: every counted conversion was a first touch, a form or a call, and a first touch has no lag by definition. The long-cycle branch of the journey model can't be observed until an uploaded outcome becomes the primary action somewhere, and on this fleet that was true of exactly one account, at 12.98 matched.

Smart Bidding optimizes toward the primary conversion it is shown, and it needs volume to learn, so a small account gets pushed toward the event it has most of, which is the raw first touch. Every campaign type then buys more of that event. Performance Max, AI Max and Demand Gen are each, in their own way, machines for finding more of whatever the primary action counts; pointed at an ungraded form event they find more ungraded form events, and the platform's cost per conversion falls while nobody learns whether a single one became a customer. Adding an automated campaign while the counted event is wrong raises the volume of the wrong event, at a lower unit price, with a better-looking dashboard.

So the ladder comes first, and the journey becomes a template applied inside a rung: the shape of the value and the lag once a graded outcome exists to attach them to. That reorders the work. Most of what this playbook asks of an account below the third rung is pipeline: a ledger, someone grading, a daily upload, a match rate worth checking. None of it is a campaign change, and most of it isn't Google Ads work at all. "Unmeasured" is a fact about our instruments. Fifteen accounts without grading is 15 pipelines we hadn't built, and that is the part we could have fixed without asking Google for anything.

3. The five gates

A gate is a test an account passes or fails from its own data, and an account holds a rung until it clears the gate above it. Five gates cover the plumbing, the counting, the ledger, the bridge between platform and ledger, and the phone. Each one below carries the query we used to check it, the interface path for readers without API access, what 19 accounts looked like against it on September 22, and the repair. The queries were run read-only against Google Ads API v25 on September 23, 2026, through a manager account. If you run them through your own manager account, send its id as login-customer-id; if a field fails, check it against that version's field reference rather than guessing a replacement.

19 of 19

G1 Attribution intact

Auto-tagging on across the fleet; 5 accounts still lacked an account-level tracking template.

13 of 19

G2 One trustworthy lead event

Had at least one contact action counting every conversion per click; on 2 of them it was the only counted signal.

4 of 19

G3 System of record with grades

15 accounts had leads landing in a ledger; 4 had a disposition path; 4 had no pipeline at all.

once

G4 Reconciliation

Platform count against ledger count against the client's own record, done fleet-wide once, for August, on September 14.

2 of 6

G5 Calls attributable

Of six call-dominant accounts, two could tie a call to a click.

G1. Attribution intact

What it tests. Every ad click carries a Google click ID, and the ID survives the trip to wherever the lead is recorded. Auto-tagging supplies the ID; a tracking template carries it through any collector that sits between the ad and the site, such as a call tracker or a lead router. Without both, nothing downstream can be matched, and every later gate fails quietly.

How to check. Two queries. The first reads the account; the second reads each campaign, and a campaign with its own template overrides the account's.

SELECT customer.id, customer.auto_tagging_enabled,
  customer.tracking_url_template, customer.final_url_suffix
FROM customer

SELECT campaign.id, campaign.name, campaign.status,
  campaign.advertising_channel_type, campaign.bidding_strategy_type,
  campaign.tracking_url_template
FROM campaign
WHERE campaign.status != 'REMOVED'

In the interface: Admin, Account settings, Auto-tagging; then each campaign's settings under Campaign URL options.

What we saw. ◇ Auto-tagging was on in 19 of 19 accounts. Five had no account-level template, which matters only where a collector exists; wherever one does sit in the path, the click ID stops at it.

The fix. Turn auto-tagging on. Put the template at account level and pass the click ID as a parameter the collector stores. Then prove it: click an ad, find the lead, and read the ID off the row. Until somebody has traced it end to end, the template is only a setting.

G2. One trustworthy lead event

What it tests. One primary conversion action per real lead channel, counted once per click, firing on submission rather than on a page view, gated against spam, with no GA4 twin of the same form also set to primary and no engagement action inside the lead total. The counting setting is the one that fails most: an action set to count every conversion turns a nervous caller who rings three times into three leads, and a bid strategy learns from all three.

How to check. The first query lists every enabled action with the four settings that decide whether it can be trusted. The second shows which goals each campaign actually bids on, because a campaign-level goal can differ from the account default on purpose.

SELECT conversion_action.id, conversion_action.name, conversion_action.type,
  conversion_action.category, conversion_action.status,
  conversion_action.primary_for_goal, conversion_action.counting_type,
  conversion_action.include_in_conversions_metric,
  conversion_action.value_settings.default_value,
  conversion_action.value_settings.always_use_default_value,
  conversion_action.click_through_lookback_window_days
FROM conversion_action
WHERE conversion_action.status = 'ENABLED'

SELECT campaign.id, campaign.name, campaign_conversion_goal.category,
  campaign_conversion_goal.origin, campaign_conversion_goal.biddable
FROM campaign_conversion_goal
WHERE campaign.status = 'ENABLED'

In the interface: Goals, Conversions, Summary, then each action's settings for Count and for Primary or Secondary.

What we saw. ◇ Thirteen of 19 accounts had at least one contact action set to count every conversion per click. On two of them that action was the only counted signal, so the total was not one per lead and the account sat on T0; on the others a one-per-click action carried the count, which is why the ladder can show 16 accounts above T0 while 13 carry the flag. One account had a GA4 lead event as its primary action with an enabled lead-form twin beside it, so one enquiry could count twice. Three accounts had engagement actions enabled, one of them a YouTube campaign whose subscriptions and follow-on views made up most of the account's counted total.

The fix. Set every contact action to one per click. Demote the GA4 twin to secondary, or remove it. Keep engagement actions as campaign-level goals where the campaign is meant to buy engagement, and out of the account's lead total everywhere else. Put a spam gate on the form: a challenge such as Turnstile, or a qualifying field the server checks before the event fires. Then annotate the change, because the conversion trend breaks the day the counting does.

G3. System of record with grades

What it tests. Every lead lands in one ledger with its source, its click ID where one exists, and a path for somebody to grade it. Grading means a person or a system records what happened after the form or the call: spam or duplicate, contacted, qualified, booked or quoted, won, with a date on each transition and a consistent reason for rejection. Without a ledger the account can pass the two gates above and still teach bidding nothing, because a raw form event with no outcome behind it is the most the strategy will ever see.

How to check. There is no query for this one; the ledger is the client's. Ask three questions instead. Where does a lead land the moment it arrives? Who marks it qualified, and by when? Can you show me last month's rejected leads with the reason on each? A client with no CRM can run the ledger in a spreadsheet; ● Google accepts a spreadsheet as the source for offline imports. The columns that matter:

What we saw. ◇ Fifteen of 19 accounts had leads landing in a ledger. Four had a disposition path: someone was grading. Four had no pipeline at all, so their platform counts had nothing to reconcile against.

The fix. Build the ledger before touching a bid. Register who grades and how often. Export it nightly, because the upload in the next gate depends on it and a grade that cannot be queried cannot be uploaded.

G4. Reconciliation

What it tests. For each primary action, the platform's Conversions column, its All conversions column, the ledger count, and the client's own record agree for the same period, the same currency, and the same campaign goals, or the difference is explained. The explanations are finite: attribution windows, modelled or fractional conversions, repeat calls, spam, legacy actions still receiving hits, and uploads that never matched. Until that bridge exists, a Google Ads total is a platform metric, and labelling it a lead count is the error most reports are built on.

How to check. The first query puts the two conversion columns side by side for every action over the last 30 days. The second lists who changed what, which is how a swing in the count gets tied to a change in the account rather than to demand. The third reads the lag buckets, which tell you whether the counted event is a first touch. The fourth reads each upload action's own diagnostics: how many uploaded rows Google accepted, when the last upload arrived, and the alerts it raised, which is where an upload that matches nothing announces itself.

SELECT segments.conversion_action_name, segments.conversion_action_category,
  metrics.conversions, metrics.all_conversions, metrics.conversions_value
FROM customer
WHERE segments.date DURING LAST_30_DAYS

SELECT change_event.change_date_time, change_event.change_resource_type,
  change_event.changed_fields, change_event.client_type, change_event.user_email
FROM change_event
WHERE change_event.change_date_time >= '2026-08-25'
  AND change_event.change_date_time <= '2026-09-23'
ORDER BY change_event.change_date_time DESC
LIMIT 200

SELECT segments.conversion_lag_bucket, metrics.all_conversions
FROM campaign
WHERE segments.date DURING LAST_30_DAYS

SELECT offline_conversion_upload_conversion_action_summary.conversion_action_name,
  offline_conversion_upload_conversion_action_summary.status,
  offline_conversion_upload_conversion_action_summary.successful_event_count,
  offline_conversion_upload_conversion_action_summary.total_event_count,
  offline_conversion_upload_conversion_action_summary.last_upload_date_time,
  offline_conversion_upload_conversion_action_summary.alerts
FROM offline_conversion_upload_conversion_action_summary

Replace the two dates with your own window. Three traps in the change query, all hit live: it rejects relative ranges such as DURING LAST_30_DAYS, it requires a LIMIT, and it refuses a start date more than 30 days back, so the audit has to run inside the month it covers. In the interface: Goals, Conversions, Summary, with both conversion columns showing; Change history; and the Time lag report in the Attribution section.

What we saw. ◇ The bridge had been built once, fleet-wide, for August, on September 14. The pull also turned up two conversion actions that had been removed and were still receiving hits, and campaigns paused during the window whose spend still sat inside the window's totals. Both are ordinary, and both would have gone into a monthly report as performance.

The fix. Reconcile at least quarterly, and immediately after any tracking, website, or primary-goal change; a tracking break nobody notices costs a month of optimization before the reconciliation catches it. Write the bridge down: platform count, ledger count, client count, and a line per difference with its cause. Annotate every change to a primary action, because it moves the learning and the trend at once.

G5. Calls attributable

What it tests. A call can be tied to the click that produced it, and somebody classifies what happened on it. Google's own call actions count without any of that: a call from an ad becomes a conversion once it passes a duration threshold, 60 seconds by default, and a click-to-call action counts the tap itself, and either can happen with no click ID ever reaching the ledger. The account then has a call total it cannot audit and a bid strategy learning from every wrong number.

How to check. The conversion-action query under G2 already lists each action's type; the call kinds are calls from ads, click-to-call, and website calls. Then ask the ledger the only question that matters: of last month's call rows, how many carry a click ID? In the interface: the same Summary view, filtered to call actions.

What we saw. ◇ Two of the six call-dominant accounts could tie a call to a click. Three clinics logged 56, 245 and 144 tracked calls in the month, 445 between them, with no click ID on any row. The two that passed used pool-based dynamic number insertion on the tracked site; one of them attributed 67 calls to clicks in 30 days and classified all 67. ● Google's own answer is AI-qualified call conversions, documented on April 21, 2026, in the United States and Canada, with healthcare and finance excluded from default call recording, so the clinics that most need it are the ones that must opt in deliberately after a consent review.

The fix. Put a number pool on the site so each session gets a number tied to its click. Classify calls as answered, relevant, and booked, by a person or by a model whose agreement with a person has been measured. Report answered relevant calls by location, service, and hour, because an ad that rings a phone nobody answers at 7 p.m. has bought a conversion and lost a customer. Confirm consent, recording settings, and category rules before any automated classification, and then ask the business whether it can answer at the times the ads run before buying more calls.

4. Allowable CAC and maximum CPL

The bid ceiling comes from the client's economics, and the formula most guides print for it runs backwards. Two numbers set the ceiling. Allowable acquisition cost is the gross profit a new customer brings in over the client's payback window, multiplied by the share of that profit the client is willing to spend acquiring them. Maximum cost per lead is that allowable cost multiplied by the lead-to-sale rate, because only that fraction of leads ever becomes a customer. A business that closes one lead in five can afford, per lead, one fifth of what it can afford per customer. Where calls, forms, appointments and quotes close at different rates, each gets its own ceiling.

The values in the table are illustrative and belong to no client; the arithmetic is the point.

Dividing by the rate instead of multiplying is the common error, and it is not a small one. The two operations differ by the square of the rate's reciprocal: at the 25 percent close rate the research behind this playbook used, the wrong ceiling is 16 times the right one, and in the table above, at 20 percent, it is $3,000 against $120. Nothing in the platform flags it. An account bidding toward a $3,000 lead ceiling when the business can afford $120 reports a cost per lead comfortably under target for as long as the budget lasts, the report is arithmetically true, and the shortfall appears on the client's income statement.

Put your own numbers in. The third output is the wrong formula, kept visible so the size of the error is never abstract.

Three cautions on the inputs. They come from the client's books, not from the ad platform: the gross profit and the payback policy are decisions the owner makes, and an agency that fills them in on the client's behalf has invented the ceiling it will then hit. A flat conversion value typed into Google Ads is a default somebody chose, and section 11 counts how many accounts were bidding on one. And the shorthand some guides use, deal value times margin times close rate, is the special case where the client is willing to spend all of the first period's gross profit on acquisition, a policy no client of ours has stated. Two widely shared guides carry a version of the error: Silverback Strategies' 2026 offline-conversions guide divides by the close rate, and Connective Web Design's high-ticket ROAS playbook uses the all-of-first-period-profit shorthand and then applies one break-even ROAS target to two streams valued on different bases. The sources list links both, for the formula and not for the advice.

5. What each rung permits

A rung grants permission to run a bidding experiment. It describes what the bid strategy can see and nothing about the quality of the business. The table says what each rung requires, what bidding it permits, what it has earned, and the test that moves an account up. Two accounts on the same rung can differ in every commercial respect and still need the same next move, because the next move is about what the bid strategy can see.

The rung above T1 is decided by matched uploads, never by how many leads the CRM says are good, and the gap between the two is large enough to change the rung. ◇ One auto service account shows the whole climb in a single set of numbers. Its client graded 31 leads qualified in a shared sheet during the window; a nightly upload sent them to Google, which matched 16.94; and the upload action was secondary, so the bidding column read zero. That account has met the T2 promotion test in the table above, and the move that would put it on the third rung is one setting: make the matched action primary under target CPA, after reconciling the 16.94 against the sheet and against the account's Smart-campaign call total, and only with the owner's approval of that specific change. It was the only account on the fleet in that position. The loss runs the other way too: another account sent 604 hashed uploads that matched nothing at all, and nothing in the campaign reports said so; the upload diagnostics query in section 3 is where it shows. Google matches an upload to a click by the click ID or by hashed contact data, inside a retention window, and every lead the match loses is a lead the auction never learns from. The number that counts is the All conversions figure on the upload action, read from the platform, because bidding learns from matched rows and from nothing else.

Three kinds of threshold appear in this table and the one in section 6, and they carry different weight. One is a hard platform gate: ● Demand Gen campaigns need a daily budget of at least $5 through the API since April 2026, and nothing else in this playbook is enforced by Google. The second is platform guidance, published but unenforced: 15 conversions per 30 days before target ROAS; AI Max at $50 a day with no budget cap over eight weeks; a Demand Gen daily budget of 10 to 15 times target CPA and about 50 conversions before judging; and roughly 60 conversions per 30 days for Performance Max, the last of those a correlation in a vendor study rather than a Google figure. The third kind is ours: the 15 raw events before target CPA in the T1 row, borrowed from Google's target ROAS floor rather than documented for target CPA, and the capacity bands, which are hypotheses rather than measured cutoffs: under $1,500 a month, one campaign and one live test at a time, with no AI Max, no Demand Gen, and Performance Max only where it already runs and is being measured against Search; $1,500 to $5,000, Search plus one measured secondary campaign; $5,000 to $10,000, two concurrent tests; above $30,000, geo or campaign experiments become readable. We have set no band between $10,000 and $30,000 yet, and a reader in that range should take the $5,000 rule until a test says otherwise. ◇ Eleven of the 19 accounts spent under $1,500 in the 30-day window, the first band. Revise the third kind as tests read out, and treat the first kind as the only threshold that will stop a campaign from serving.

6. Campaign types by what can be measured

No channel receives a fixed share of spend. Budget funds the next observable test and the service areas the client can fulfil, and for a small account that usually means consolidating a few high-intent themes rather than dividing the same money among Search, Performance Max, Demand Gen and AI Max, where none of the four will produce a readable result. Each campaign type below has a job, an entry rule, a decision metric and a stop condition. The entry rule is the rung, and the decision metric is never the campaign's own conversion column.

Defaults that hold on every rung

Four settings are the same on every rung, because each one leaks spend the bid strategy cannot see. Final URL expansion stays off in Performance Max and AI Max unless an experiment is running, so a person chooses the landing page. Brand terms and out-of-scope services sit in shared negative lists attached to every non-brand campaign. ● Location targeting is set to presence, meaning people in or regularly in the service area; Google's default and recommended option is presence or interest, which also serves people who have only shown interest in the area, and a local business cannot serve a lead from a city it does not work in. And Search partners are off on brand campaigns, where the brand query is already won on Google; ◇ one account's brand campaign was serving on partners in September.

Search

Search is the baseline and the comparison arm, and it sits on every rung because it is the one campaign type whose queries can be read. Build it around service, geography and the high-intent query clusters. Use exact and phrase first while quality feedback is weak: ◆ across lead-generation accounts in a 30,000-account match-type study, phrase held the largest share of spend and conversions and broad lost ground wherever value data was absent. Review search terms and negatives on a cadence, and split brand from non-brand when their economics differ; ◇ 13 accounts were serving search terms with no negatives against them, their own brand names among them. Maximize Conversions or target CPA once 15 or more reliable raw events arrive per 30 days; below that, Maximize Clicks with a tight query set, which is the honest setting for an account that cannot yet feed the strategy. The decision metric is qualified leads, appointments or quotes, and won customers per dollar spent. There is no stop condition, because Search is what every other type is compared against.

AI Max for Search

AI Max is off by default and is tested, never adopted. The entry rule is the third rung: a graded signal, a stable Search campaign with $50 a day or more on it, and a native experiment rather than a toggle, read on cost per graded lead split by match source so the expansion traffic is judged on its own. The evidence for caution is directional and small, and it points one way. ◆ One agency's controlled SaaS test found blended cost per lead improved while the expansion layer's signups cost about twice as much as the control's to become customers. ◆ Across 23 tests at another agency, 54 percent of the queries AI Max reported as new were already being captured elsewhere in the account, and the net uplift was about 3 percent. ◆ The expansion launches mostly from exact match, 80.11 percent exact, 19.52 phrase and 0.38 broad in a dataset of more than 250 retail campaigns, so the campaigns most likely to be widened are the ones built tightest. ◆ Two single-account reports saw cost per lead rise while cost per click fell, and one found the exit had a cost of its own: after AI Max was switched off, CPC kept falling and cost per acquisition stayed high. Stop condition: the added traffic misses the client's acquisition ceiling on cost per graded lead. The mechanics, the seven readiness gates we score a campaign on, and the September auto-migration are in the AI Max field guide, the verified dossier on the migration, and our September audit of the same 19 accounts.

Performance Max

Performance Max is permitted on the third rung, with offline feedback steering it. On the first two rungs it exists only as a test whose readout is cost per graded lead against the Search arm, with final URL expansion off, campaign negatives in place (they reach Search and Shopping inventory only), and the first-party audience exclusions Google added for 2026. ◇ Eight accounts were running Performance Max on the bottom two rungs. None had a readout, and three had no Search campaign running, so no comparison was possible there without building one. ◆ On overlapping terms Search converts better than Performance Max more often than not, but the margin is thin: a 503-account study found no significant difference in 75 to 88 percent of cases. ◆ Below about 60 conversions per 30 days, Performance Max underperformed on every metric except click-through rate in a 9,199-account cross-section, and brand exclusions, audience signals and search themes, the settings every agency recommends, came out flat to negative in the same data. ◆ Two agency cases show that uploading a downstream milestone changes what the campaign buys: one trade-school account took a 17 percent worse enrolment rate in exchange for a 65 percent lower cost per lead and a 59 percent lower cost per enrolment, which is the trade a graded signal makes visible and an ungraded one hides. No incrementality figure for Performance Max in lead generation has been published anywhere we could find. ● Google offers an uplift experiment; whether an account is eligible and whether the test has enough power is decided account by account. Stop condition: total-account graded outcomes do not improve once Search cannibalization is netted out.

Demand Gen and YouTube

Demand Gen and YouTube create or reinforce demand, and below roughly $10,000 a month against a lead target they have not earned a place. ● Google's own performance guide asks for a daily budget of 10 to 15 times the target CPA and about 50 conversions before judging a changed campaign. ○ One agency reported very low conversion rates at $50 to $100 a day, and another practitioner's wins came from campaigns optimized toward cheap, shallow events. The readout has to be branded-search lift or pipeline against a control, never the campaign's own engagement count. ◇ One account ran a $25-a-day playlists campaign optimized to engagement; it booked 244 of the account's 303.96 counted conversions as YouTube subscriptions (81) and follow-on views (163), every one at $0 value. That is a legitimate campaign-level goal, and it belongs in an engagement report rather than a lead total; it was also the single largest reason the account's Conversions column could not be read as leads. Stop condition: no lift in branded search or pipeline against a control.

Smart and local-style campaigns

Smart campaigns migrate at the first opportunity. Their tracked-call conversions carry no value and no click tie, so an account whose largest counted signal is a Smart-campaign call total holds a number that can be neither audited nor taught. ◇ Three of the 19 were in that position. Plan the migration for when the replacement's tracking and service coverage are ready, and judge it on verified calls or bookings, not on preserving the old total.

What the published record supports, claim by claim:

7. Modifiers inside a rung

The journey, the lead channel, the category and the budget all matter, and each of them applies inside a rung. Four modifiers change what the work looks like without changing what the bid strategy can see.

Lead channel: calls against forms

Call-dominant accounts get the call-attribution build before anything else: a number pool on the tracked site, classification of what happened on the call, and AI-qualified calls where policy allows. Form-dominant accounts get the spam gate and the server-side qualifier first, because a form event that fires on a bot or a job-seeker is a conversion the auction will buy again. The 445 unattributed calls a month across three clinics in section 3 are the call-side version of the problem; the form-side version is quieter, since a spam submission counts exactly like a real one and nothing in the platform separates them. ◆ The one measured qualifier we could find is small and telling: one B2B account added its pricing to the headlines and lead qualification rose 27 percent, with bidding held constant, over one month each side. Filtering before the click is the cheapest filtering there is, because the unwanted lead is never bought.

Category constraints

● Google's personalized advertising policy names invasive medical procedures, cosmetic surgery and injections. For a surgical implant practice or a clinic selling injectables, the full restriction is supported by the policy text: no enhanced conversions, no Customer Match, no lookalikes, no your-data segments, uploads by click ID only. Routine care such as eye exams, physiotherapy and chiropractic is not named in that text, and ◇ on our own fleet the five routine-care accounts had never had a per-account policy read. The interim default we use for them is enhanced conversions off and no audience lists built, because the downside is asymmetric: a wrongly enabled setting on a health account is a policy breach, and a wrongly disabled one is some lost match rate. Seven of the 19 accounts were health-adjacent, so the question is not marginal. Where competitor-name conquest is common, legal services among them, wasted terms concentrate on the competitor names, and because brand-list exclusions inside Performance Max are unproven at scale, negatives carry the load. B2B accounts attract consumer enquiries that look like ordinary form conversions; the qualifier gate is the quality lever, and it must not hide the rejected enquiries from analysis, since the rejection reasons are how the targeting gets fixed.

Journey as a value-and-lag template

The three buying journeys survive as templates a third-rung account adopts once a graded outcome exists to attach a value and a lag to. Urgent local services (optometry, physiotherapy, auto repair, refrigeration repair, funeral care) take one flat value per booked or qualified event and count bidding; the first touch is the outcome and no lag handling is needed, but an answered relevant call and a missed one are not worth the same, and the template has to say so. Considered consumer purchases (cabinetry, paving, wraps, woodworking, aesthetics, implants, legal matters) take a two-stage value, contact and then booked or quoted, uploaded within seven days with an interim milestone standing in for the sale; ◇ on our fleet exactly one account had a calibrated pair of stage values, and it is the only one for which a two-stage template can be checked against outcomes. Long-cycle B2B (accounting, commercial building supply, commercial refrigeration) takes a proposal-sent style interim milestone before closed-won, per-stage values the client sets, and a bake-rate read of how much of a month's outcome has arrived by a given day. Read results by service or case type wherever the economics differ, because a flat form value hides the difference between a small enquiry and a viable project.

Budget as capacity

Budget is capacity for tests. The three threshold classes in section 5 (one hard platform gate, a handful of published guidance figures, and our own capacity bands) are the whole of what budget decides here. ◇ Eleven of the 19 accounts spent under $1,500 in the 30-day window, under $50 a day, which is below Google's own guidance for a single AI Max campaign before any other test is considered, and it is the reason the bottom band's rule is one campaign and one live test at a time. An account in that band that runs Search, Performance Max and an AI Max experiment at once has divided a readable budget into three unreadable ones.

8. Experiments that answer a business question

Write the success test before changing delivery. The record names the treatment, the control, the split (geography or traffic), the primary qualified outcome, the allowable cost, the expected conversion lag, and the date by which the outcome can be read. Run one major experiment per small account at a time. ● Google recommends sequential experiments because concurrent tests interfere, allows four to six weeks in many experiment settings, and treats the first seven days as ramp-up for certain tests; the actual duration also has to cover the client's qualified-lead delay, which the platform does not know.

Test the whole account, because a campaign's own conversion column cannot show what it added. AI Max has to show that its expansion traffic contributes qualified leads beyond what the existing keywords were already catching. Performance Max has to show that total-account qualified outcomes improved after Search cannibalization is netted out. Demand Gen has to show pipeline or a credible incremental-demand signal rather than inexpensive engagements. When qualified volume is too low for a reliable split, run a bounded observational pilot with an explicit inconclusive outcome written into the plan, and return the budget to the strongest verified demand source if the acquisition ceiling is missed. An inconclusive result recorded honestly is worth more than a win read off a column that cannot carry it.

Keep a decision record. For every test, save the live settings at the start, the baseline period, the client's economics (gross profit, permitted share, lead-to-sale rate), the change event, an export of the conversion-goal configuration, the success rule, and the readout date, in whatever system holds the client's work. The platform's change log reaches back 30 days; the record has to outlive it. And no budget, bid, targeting, campaign or conversion-action change follows from a document, this one included, without the specific change being approved by the person accountable for the account.

9. The scorecard

One scorecard for every account, one explicit date window, compared with an equal prior window. Six fields; each has a required reading, and a field with no reading is a finding.

Two rules for reading it. The window has to equal the prior one, because an unequal comparison reads as a trend. And the platform's count and the client's count go on separate lines until the reconciliation in section 3 has joined them; a scorecard that prints one number for both has already decided which to believe.

10. Claims to reject

Some of what circulates as Google Ads lead-generation advice is wrong on its face, and more of it is unverifiable. The first list is the one to stop repeating.

• The maximum-CPL formula that divides by the close rate. The correct chain is in section 4; the division form overstates the ceiling by the square of the rate's reciprocal, 16-fold at a 25 percent close rate.
• "Search beats Performance Max 84 percent of the time" as a magnitude claim. The 84 percent measured direction with no significance threshold; the study that applied one found no significant difference in 75 to 88 percent of cases.
• Any platform Conversions figure quoted as leads. On our fleet the four largest counted totals were 934.40, 303.96, 94.99 and 63.99, and each was a count of whatever the account's actions happened to count: call extensions, YouTube subscriptions, Smart-campaign calls.
• "30 to 50 conversions a month" as a hard Smart Bidding gate. Google's documented floor is 15 per 30 days for target ROAS; the higher figure is agency comfort, and nobody sources it.
• "An incrementality experiment costs $5,000." Google's figure has no defined scope, and the competing claim of a $100,000 to $250,000 monthly spend floor is equally unsourced. Neither tells a $2,000-a-month account anything.
• Benchmark cost-per-lead and cost-per-acquisition figures from mixed business models used as lead-generation targets. The public Optmyzr benchmark mixes business models and gates the lead-gen split; the Pixis benchmark is 100 self-selected consumer brands. Neither is a target for a clinic or a paving company.
• The first-touch lag report as evidence that your journeys are short. It sees first touches only, and a first touch has no lag.

Quarantined, meaning unverifiable rather than wrong: Google's 14 percent and 7 percent AI Max uplifts (the baseline is AI Max matching only, the sample is undisclosed, and retail is excluded); the "47 percent junk-lead rate", "98 percent of PMax leads are spam" and "34 percent CPL drop with offline imports by month three" figures that circulate with no method behind them; and any median incremental ROAS or "platform overstates by 30 to 70 percent" figure, which comes from direct-to-consumer ecommerce revenue measured on a vendor's own platform.

11. What we found on our own fleet

◇ Thirteen of 19 counted a contact action many per click. Eight of 19 ran Maximize Conversion Value or target ROAS on flat default values in at least one enabled campaign, and the defaults we found across the fleet were 1, 10, 20, 25, 50, 75, 100, 150, 200, 400, 500, 2,000 and 50,000; with one value per action, value bidding degenerates into count bidding weighted by whoever typed the number. Six accounts had an offline upload action and one of them steered bidding. Eleven accounts met the 15-per-30-days floor on raw counted conversions and zero met it on a graded action. Three clinics logged 445 calls a month with no click ID on any row. One account booked 244 of its 303.96 conversions as YouTube subscriptions and follow-on views. Eight accounts ran Performance Max with no readout against Search, three of them with no Search campaign running to read against. Thirteen served search terms without negatives, their own brand names among them. And every account, all 19, booked 90 to 100 percent of counted conversions in the under-one-day lag bucket, because every counted conversion was a first touch.

Clearing Google's floor on a graded signal is structurally hard for a small account, and the reasons are not negligence. Grading is human work: somebody has to look at each lead and decide, within days, what it was. The upload is plumbing that has to run daily and carry the click ID or hashed contact data through to Google. And the match between grade and upload is lossy at both ends: ◇ one account had 31 graded qualified leads in its ledger and 16.94 matched in the platform, and another sent 604 hashed uploads that matched nothing. An account producing 30 graded leads a month reaches the floor of 15 matched only if every link in that chain holds, and on September 22 none of the 19 had every link in place.

For 15 of the 19 accounts, qualified lead volume was unmeasured, and unmeasured is our failure before it is the client's: the pipeline that grades a lead and sends the grade back is agency infrastructure, and it existed on four. The fleet's next year is therefore pipeline work, ledger by ledger, and the campaign decisions this playbook spends its middle sections on wait behind it. △ We think the same is true of most small fleets: no public source we found reports what grading raw leads does to Smart Bidding on accounts under $5,000 a month, which suggests few have had the pipeline to find out. Two limits on the numbers. They are one pull on one day. And the matched figures depend on Google's matching, which we can read but not audit, so the loss between grade and match is measured from the platform's side only.

12. The appendix your agent can read

The block below encodes this page's operating facts in a form to paste into whatever tool helps run your account, together with the queries in section 3. It asserts nothing the sections above did not source, and it keeps Google's documented mechanics apart from our own framework, which your agent is free to disagree with.

# Signal-first Google Ads playbook, operating facts
# Source: https://choice.marketing/research/signal-first-google-ads-playbook-2026/
# Queries verified read-only against Google Ads API v25 on 2026-09-23.

# --- platform mechanics (Google documentation) ---
value_bidding:
  troas_floor_conversions_per_30d: 15          # help 15099424
  distinct_values_required: 2
  upload_cadence: daily, one to two conversion cycles before promotion
retention:
  gclid_upload_days: 90                        # help 10029210
  user_provided_data_upload_days: 63
  adjustment_days: 55
demand_gen:
  api_minimum_daily_budget_usd: 5              # hard gate
  guidance_daily_budget_x_tcpa: [10, 15]       # help 16797388
  guidance_conversions_before_judging: 50
ai_max:
  guidance_daily_budget_usd: 50
  guidance_window_weeks: 8
experiments:
  guidance_weeks: [4, 6]                       # help 13826584
  ramp_days: 7
ai_qualified_calls:
  since: 2026-04-21
  regions: [US, CA]
  default_recording_excluded: [healthcare, finance]

# --- Choice OMG framework ---
gates:
  G1: attribution intact (auto-tagging, tracking template)
  G2: one trustworthy lead event (one primary per lead channel, one per click, fires on submission, spam gated, no GA4 twin as primary)
  G3: system of record with grades (ledger, gclid where present, disposition path)
  G4: reconciliation (platform vs ledger vs client record, quarterly minimum and after any tracking change)
  G5: calls attributable (click tie, classification where policy allows)
rungs:
  T0: {fails: [G1, G2], bidding: max_clicks_or_max_conversions_capped, campaigns: [search]}
  T1: {passes: [G1, G2], bidding: max_conversions_then_tcpa_at_15_raw, campaigns: [search, pmax_as_measured_test]}
  T2: {adds: graded_outcomes_uploaded_secondary, bidding: raw_event, campaigns: [search, pmax_test_if_volume]}
  T3: {adds: 15_matched_graded_per_30d_primary_within_7d, bidding: tcpa_on_graded_then_value, campaigns: [search, pmax_with_offline_feedback, ai_max_experiment]}
  T4: {adds: two_distinct_values, bidding: troas_or_mcv_staged_values, campaigns: [plus_demand_gen_under_budget_rule]}
economics:
  allowable_cac: gross_profit_per_new_customer_over_payback_window * permitted_share
  max_cpl: allowable_cac * lead_to_sale_rate
  never: allowable_cac / lead_to_sale_rate
budget_bands_hypothesis_monthly_cad:
  under_1500: one campaign, one live test
  1500_to_5000: search plus one measured secondary
  5000_to_10000: two concurrent tests
  above_30000: geo or campaign experiments readable
stop_conditions:
  ai_max: expansion traffic misses the acquisition ceiling on cost per graded lead
  pmax: account-level graded outcomes do not improve net of search cannibalization
  demand_gen: no branded-search lift or pipeline against a control
queries:
  g1: [customer.auto_tagging_enabled, customer.tracking_url_template, campaign.tracking_url_template]
  g2: [conversion_action.counting_type, conversion_action.primary_for_goal, conversion_action.include_in_conversions_metric, campaign_conversion_goal.biddable]
  g4: [metrics.conversions vs metrics.all_conversions by segments.conversion_action_name, change_event with explicit range and LIMIT, segments.conversion_lag_bucket, offline_conversion_upload_conversion_action_summary]
  g5: [conversion_action.type in call kinds]

If a field name in the block stops resolving, the API version moved; check the date above against Google's release notes before trusting the rest.

If the gates are where your account is stuck, that is the work we do first: Google Ads management at Choice OMG starts with the signal, and the first conversation is about your ledger.

Sources