Why Performance Max brings junk leads, and how to fix it

Performance Max is not broken. It is doing exactly what you asked, and you asked for form fills.

Rows of desks with computer workstations in a very large call centre
Photo: Petiatil (talk) on Wikipedia-en., Public domain, via Wikimedia Commons

Performance Max lead quality drops because the campaign's bidder optimises toward whatever you mark as a primary conversion, and a raw form submission is the cheapest thing on the internet to produce: bots, job seekers, students and competitors all fill forms. Fix it by changing what the bidder learns from: import CRM outcomes (qualified lead, opportunity, closed deal) through offline conversion import or enhanced conversions for leads, make those the primary goal, give each stage a value, and demote the raw form fill to secondary.

Key takeaways
  • PMax bids on the primary conversion you give it; a raw form fill is the cheapest thing to produce, so that is what it finds.
  • Goodhart's law in four forms (regressional, extremal, causal, adversarial) explains why, and which fix applies.
  • Import CRM stages with values and switch to Maximise conversion value; demote the form fill to secondary.
  • Google reports a median 10% more conversions when first-party data is added to GCLID imports (vendor figure).
  • From 15 June 2026, offline imports run through the Data Manager API, not the Google Ads API.

Most B2B and edtech marketers who switch on Performance Max (PMax) for leads see the same pattern: cost per lead falls, the dashboard looks wonderful, and sales says the leads are rubbish.

Both are true at once. Here is why, from Google's documentation and research on optimisers that chase a proxy.

Why does Performance Max lead quality get worse over time?

Performance Max lead quality gets worse because the campaign learns from the conversions you count, and the cheapest counted conversions come from the people least likely to buy.

Google's help centre says PMax uses Smart Bidding to set bids in real time based on the predicted probability of a conversion, across Search, YouTube, Display, Discover, Gmail and Maps from one campaign. Nothing in that loop knows what happens after the form is submitted unless you tell it.

Think about who fills a "Request a demo" form on a mobile Display placement at 1 a.m. Some are real buyers. Many are not:

  • Bots and form spam. Automated scripts submit forms at near zero cost, and they never stop at one.
  • Job seekers. A company's ad is often the easiest way to reach its "Contact us" form.
  • Students and researchers. Common on edtech, SaaS and anything with a free download.
  • Accidental taps. Display and in-app placements produce misclicks that sometimes finish a short form.
  • Vendors and competitors. People selling to you, or checking your pricing.

Each of these converts more cheaply than a CFO in Pune comparing three vendors. Once the bidder finds a pocket of inventory that produces form fills far more cheaply than real buyers do, it moves budget there. That is not a bug. It is the job you gave it.

The story

In April 1902, the French colonial government in Hanoi fought a rat problem in its new sewers by paying a bounty of one cent per rat, collected on proof of a severed tail. The numbers looked superb: 7,985 rats in the last week of April, more than 4,000 a day in early May, and 20,112 on 21 June alone. Then officials began to see live rats running through the city without tails. Hunters were cutting off tails and releasing the rats to breed, and health inspectors found farms on the city's outskirts raising rats for their tails. The bounty was cancelled because it had made the problem worse, and plague broke out in 1903. Historian Michael Vann rebuilt the episode from the colonial archives. The city had paid for a proof of the result, not the result.

Wikipedia: Great Hanoi Rat Massacre (drawing on Michael G. Vann, French Colonial History, 2003)
Aerial view of the main Google campus in Mountain View, California
Google's campus in Mountain View. The bidding models built here optimise toward the signal you send them, nothing more.Photo: Austin McKinley, CC BY 3.0, via Wikimedia Commons

How does the bidder decide which conversions matter?

The bidder only counts conversion actions that are set as "primary" and that sit inside a goal the campaign is optimising toward. Google's conversion-goals documentation is precise about this:

  • Conversion actions are grouped into goals such as "Contacts" or "Submit lead form".
  • Each action is either primary or secondary.
  • Primary actions in the campaign's goal feed the "Conversions" column and bidding.
  • Secondary actions only show in "All conv." and do not steer bids.

This is the lever most accounts never touch. The default setup, created when someone installed a tag on the thank-you page, makes the raw form submission primary at account level. Every new campaign inherits it. So PMax is told, in effect, that a bot's submission and a CFO's submission are the same event, worth the same amount.

What the auto-bidding research says

The research on auto-bidding explains why that matters. The 2024 survey by Aggarwal and 25 co-authors, most of them from Google Research, describes target CPA as a value-maximising problem with a return-on-spend constraint: get as many conversions as possible while average cost per conversion stays at or below the target.

The optimal strategy under that constraint is to bid roughly in proportion to each impression's predicted value. If every conversion has the same value, the cheapest-to-convert impressions win, every time.

Auto-bidding and Auctions in Online Advertising: A SurveyGagan Aggarwal et al. (26 authors) · arXiv · 2024

A review of the theory behind automated bidding, written largely by Google researchers. It formalises target CPA and target ROAS as maximising total conversion value subject to a return-on-spend constraint, shows that the optimal bid is essentially proportional to predicted value (uniform bidding), and summarises results such as a price of anarchy of 2 for value-maximisers in second-price auctions. The limit: it models what the bidder does with the values it is given; it says nothing about whether those values reflect real revenue. That part is the advertiser's job.

Read the paper (free)

What is the Goodhart problem in lead generation?

Goodhart's law says that a measure stops tracking the goal once you optimise hard for it, and PMax with a form-fill goal is a textbook case. David Manheim and Scott Garrabrant's 2018 paper splits the law into four mechanisms, and three of them map straight onto lead generation.

Goodhart variant (Manheim and Garrabrant)What it meansHow it shows up in PMax lead gen
RegressionalSelecting hard for a noisy proxy also selects for the noiseThe cheapest form fills are cheap partly because they are low-intent; pushing CPL down pulls in more of the noise
ExtremalThe proxy-goal link that held in normal conditions breaks at extremesForm fills tracked sales well on branded Search; at 5x the volume on Display and Discover the link collapses
CausalThe proxy correlates with the goal but does not cause itA form fill is a symptom of intent, not intent itself; buying more form fills does not buy more intent
AdversarialAnother agent games the measure for its own endsSpam bots and lead-farm fraud exist because form fills are paid for; they are built to look like conversions
Categorizing Variants of Goodhart's LawDavid Manheim, Scott Garrabrant · arXiv · 2018 (revised 2019)

A short theoretical paper that separates four ways a metric can fail under optimisation: regressional (selecting on a noisy proxy also selects the noise, the "tails come apart" effect), extremal (relationships break at extreme values), causal (intervening on a correlated proxy does not move the goal) and adversarial (other agents exploit the metric). It uses simple statistical models rather than field data, so it gives no effect sizes. Its value here is the vocabulary: it tells you which fix addresses which failure.

Read the paper (free)
Form fills per month (the proxy)Revenue (the goal)Branded Search:fills track revenueScaled on Display, apps:more fills, flat revenue
A sketch of extremal Goodhart: the proxy tracks the goal in normal conditions and comes apart where the optimiser pushes hardest. Illustrative, not data.

Each variant needs a different fix:

  • Regressional and causal problems need a better signal, closer to revenue.
  • Extremal problems need fences on where the campaign can go while it learns.
  • Adversarial problems need friction against bots.

What should you feed Performance Max instead of form fills?

Feed Performance Max the outcome you are paid for, or the closest stage to it that happens often enough to learn from. In practice that means sending CRM stages back to Google Ads and making a qualified stage the primary conversion.

What the bidder sees, and what it should see Ad click (GCLID stored) Form fill: cheap, noisy, gameable Marketing qualified lead Sales qualified lead Opportunity / deal CRM stage + value + GCLID + email Offline import: primary goal = qualified stages, with values Form fill becomes a secondary action: still reported, no longer steering bids.
The fix is a feedback loop. CRM stages, matched by click ID and hashed email, go back to Google Ads and become the goal the bidder chases.

Google offers two routes, and they now sit on the same pipe:

  • Offline conversion import (OCI). You store the Google Click ID (GCLID) with each lead in your CRM, then upload it back with the stage and the time it happened. Google's help page says advertisers who added first-party data to their GCLID imports saw a median 10% more conversions than with standard imports.
  • Enhanced conversions for leads. Google calls this an upgraded version of offline import. Your site's tag captures hashed customer data (SHA-256 email or phone) when the form is submitted; later you upload the qualified lead with the same hashed identifier, and Google matches it inside what it describes as a confidential computing environment. It works even when the GCLID was lost, which happens with cross-device journeys and some browsers.

Timing rules for the import

Timing matters more than most teams realise. Google's GCLID import page says the click-to-conversion cycle must be under 14 or 90 days depending on the data source:

Data sourceLook-back each import
Files: Google Sheets, Cloud Storage, SFTP90 days
Warehouses: BigQuery, Snowflake14 days

A new conversion action needs 4 to 6 hours before the first upload.

Average uplift Google reports from better first-party conversion data
See the data as a table
Reported uplift (%)
Data Manager: offline and app data connected (incremental ROAS)26
Enhanced conversions vs standard imports (Search conversions)11
OCI with first-party data vs standard OCI (conversions, median)10
Google tag gateway (conversions)14

Source: Google Ads & Commerce Blog, Sep 2026; Google Ads Help (offline conversion import) · Google's own figures, different metrics and periods (Apr 2025 to Jun 2026). Treat as vendor claims, not independent tests.

Which conversion goal should be primary at each stage of maturity?

Make the deepest stage primary that still produces enough conversions for the bidder to learn, and add values so the bidder can tell stages apart. The table below is how I would set it up, depending on how much CRM volume an account has.

Primary goalWhat the bidder learnsTypical effect on lead qualityWhen to use it
Any form submission (default)Find the cheapest people who will type into a formWorst; spam and low-intent leads rise as the campaign scalesNever as the only primary goal for B2B or high-ticket B2C
Form fill with validation (work email, OTP-verified phone)Find cheap people who pass a filterBetter; removes most bots and junk numbersDay one, while CRM import is being built
Marketing qualified lead (imported)Find people who match your ICP fieldsGood; tracks firmographics but not buying intentWhen you get dozens of MQLs a month per campaign
Sales qualified lead or opportunity (imported, with value)Find people sales wants to talk toBest realistic option for most B2B accountsWhen stage volume is steady and the CRM is updated within days
Closed revenue (imported value)Find people who pay, weighted by deal sizeClosest to the goal, but slow and sparseHigh-volume, short-cycle businesses; otherwise as a value layer, not the only goal

Most mid-sized Indian B2B accounts end up in the middle: MQL or SQL primary, with values. A simple scheme works:

  1. Raw form fill: value 0, set as secondary.
  2. MQL: a small value.
  3. SQL: a larger one.
  4. Opportunity: the largest, scaled to how often each stage turns into revenue.
  5. Switch the bid strategy from Maximise conversions to Maximise conversion value.

The bidder now has a reason to prefer one SQL over ten junk fills, which is the whole point.

This is the same shift I argue for in why edtech should optimise to cost per enrolment, not cost per lead: the metric on the dashboard should be the one the business is paid on.

How much does a junk lead really cost you?

A junk lead costs far more than its CPL, because the real unit is cost per qualified lead, and that is CPL divided by the qualification rate.

A ₹400 lead that qualifies 10% of the time is a ₹4,000 qualified lead. A ₹900 lead that qualifies 45% of the time is a ₹2,000 qualified lead. The cheaper-looking campaign is twice as expensive where it counts, before you even count the sales time spent calling dead numbers.

True cost per qualified lead and per deal

Use this to check what you just read. Run the calculator twice: once with your current PMax numbers and once with your Search campaign's numbers. If PMax wins on CPL but loses on cost per qualified lead, you have your answer, and you also have the business case for building the CRM import. For how channels compare on lead-to-deal conversion, see B2B lead sources ranked by conversion.

Women operators working at a long Bell System international telephone switchboard
Every junk lead lands on somebody's call list. Sales time spent on bad numbers is the hidden half of cost per lead.Photo: The U.S. National Archives, No restrictions, via Wikimedia Commons

How do you stop spam and bot leads in Performance Max?

You stop spam leads by adding friction that bots cannot pass and humans barely notice, and by making sure spam never counts as a primary conversion. Bidding changes alone will not do it, because adversarial traffic is designed to look like a conversion.

  1. Fire the conversion after validation, not on page load. Count the lead only when the server accepts it: valid work email domain, phone that passes an OTP or format check, hidden honeypot field empty. A thank-you page that anyone can reach by URL is a conversion bots can fake.
  2. Add a bot check on the form. An invisible CAPTCHA score or a time-to-submit check catches most scripted fills. Keep it invisible; a visible puzzle costs real leads.
  3. Use the form to qualify. One or two qualifying fields (company size, budget band, role) cut casual fills and give you MQL data for the import. Longer forms lower volume and raise quality; test the trade-off rather than assuming.
  4. Upload disqualifications fast. When sales marks a lead as junk, it should never be uploaded as a qualified conversion. If you use conversion adjustments, retract or restate the original value.
  5. Fence the inventory while the new goal learns. PMax supports campaign-level negative keywords and brand exclusions. Use account-level placement exclusions for apps and sites that produce junk. Google's October 2026 update adds channel priority and budget caps so low-priority channels get only a share of spend.
  6. Feed audience signals from your best customers. Customer Match lists of closed-won accounts tell PMax where to start looking. They are hints, not limits.

How long does PMax take to relearn on a new goal?

Expect a few weeks of noisy results, because the bidder rebuilds its predictions on a smaller, slower signal. Google publishes no fixed learning period for imported lead goals, so judge by data:

  • Volume. If the new primary stage happens only a handful of times a month, move one stage up the funnel or consolidate campaigns.
  • Lag. If sales takes two weeks to mark an SQL, PMax learns two weeks late. Google's September 2026 "journey-aware bidding" pilot targets this lag.
  • Targets. Cost per conversion jumps by design when you move from form fills to SQLs. Reset target CPA or ROAS to the new unit.

Keep a Search campaign with exact-match keywords running alongside. Google's PMax documentation says exact-match Search keywords take priority over PMax when a query matches, so your highest-intent queries stay under your control while PMax relearns.

A single call centre workstation with a computer showing telephony software, waiting for a call
The CRM is where lead quality is decided. If nobody updates it, the bidder never hears the verdict.Photo: Petiatil, CC BY-SA 3.0 us, via Wikimedia Commons

Does this mean Performance Max is bad for B2B lead generation?

No. PMax is a strong lead engine once its target is right, and a liability when its target is a form fill. Its reach is exactly what makes it dangerous with a weak goal: it finds cheap fills in places a human buyer would never pay for.

Automation makes chasing a target nearly free, so choosing the target becomes the main job.Sapta, this post

If you are rebuilding that stack, start with how to build a measurement stack after cookies and how server-side tracking works on Meta, which follows the same logic of sending qualified events back to the platform. For a model that checks whether PMax is adding sales at all, see the primer on marketing mix modeling with Meridian and Robyn.

I judge paid acquisition by commercial outcomes, not a low CPL on a slide; my B2B and education work is on the work page.

Questions people ask

How do I stop spam leads in Performance Max?

Count a conversion only after server-side validation (honeypot field, bot score, work email or OTP-verified phone), so spam never becomes a primary conversion. Then import qualified CRM stages as the primary goal, so the bidder stops being rewarded for finding cheap fills. Placement exclusions and channel caps help at the edges.

Why did my cost per lead drop but sales say the leads are worse?

Because the bidder found cheaper inventory that produces form fills, not buyers. A falling CPL with a falling qualification rate is the classic sign that the campaign is optimising a proxy. Measure cost per qualified lead instead; it often rises when CPL falls.

Should I use offline conversion import or enhanced conversions for leads?

Use enhanced conversions for leads if you can; Google describes it as the upgraded version of offline import. It matches on hashed email or phone as well as the click ID, so it recovers leads whose GCLID was lost. Either way, the important part is uploading qualified stages, not just the form fill.

How many qualified conversions does PMax need to learn?

Google does not publish a hard minimum for imported lead goals. If a campaign sees only a handful of qualified conversions a month, use an earlier stage (MQL rather than opportunity) as primary, consolidate campaigns, or use stage values so earlier and later stages both teach the bidder.

Can I exclude placements and keywords in Performance Max?

Yes. Google's documentation lists campaign-level negative keywords and brand exclusions for PMax, and account-level placement exclusions apply across campaigns. Exact-match Search keywords also take priority over PMax for matching queries.

Sources

  1. Google Ads Help: About Performance Max campaigns (2026)
  2. Google Ads Help: About conversion goals (primary and secondary actions) (2026)
  3. Google Ads Help: About offline conversion imports (2026)
  4. Google Ads Help: Set up offline conversions using Google Click ID (GCLID) (2026)
  5. Google Ads Help: About enhanced conversions for leads (2026)
  6. Aggarwal et al.: Auto-bidding and Auctions in Online Advertising: A Survey, arXiv (2024)
  7. Manheim and Garrabrant: Categorizing Variants of Goodhart's Law, arXiv (2018)
  8. Wikipedia: Great Hanoi Rat Massacre (2026)
  9. Google Ads & Commerce Blog: Drive profitable growth with new data and measurement tools (Sep 2026)
  10. Google Ads & Commerce Blog: Build campaigns that drive high-converting, sales-ready leads (Sep 2026)
  11. Google Ads & Commerce Blog: Guide your campaign spending with new priority and allocation features (Oct 2026)
Performance Maxlead generationoffline conversion importenhanced conversions for leadsSmart BiddingGoogle AdsB2B
Sapta

Sapta (Saptarshi Sarkar) is a mechanical engineer turned marketer in Bengaluru and a solo touring rider. He writes the seven Saptaverse blogs. Work with him or read more of Marketblog.

Keep reading in Marketblog

From the other blogs