by SeanGoogle Ads has quietly rewritten how phone leads count. The old system was crude but honest. If a caller stayed on the line past your threshold, it counted. That was it. A chatty misdial, a lengthy complaint, a persistent spammer, all identical to the bidding algorithm.
The new system records the call, runs it through Google’s own large language model, and decides whether the conversation actually sounded commercial. It tags each call #HighIntent or #ConsultationScheduled and feeds that judgement straight into Smart Bidding. In the US and Canada this is already live. Call recording is on by default. Healthcare and finance are out of the auto-opt-in, but every other advertiser woke up in April with Google’s AI listening to their sales line.
Bigger shift than it sounds. Google hasn’t just added a feature. It has moved the definition of a “qualified lead” from the advertiser’s spreadsheet into its own algorithm.
What actually changed
Google now runs qualification through three tiers. First choice is the AI reading the recording. If recording can’t happen, it falls back to the old duration proxy. If there’s no forwarding number, it falls back to ad-interaction data. Search and Demand Gen are covered: call ads, call assets, calls from website visits. One gap: location asset calls. The map-pack ones Performance Max and Local rely on aren’t supported. If your map pack drives your phone volume, you get none of this.
Why this matters at the account level: Smart Bidding is only as good as the conversion data it eats. Optimise toward a duration proxy and the algorithm bids aggressively on long conversations regardless of whether anyone has a wallet. Narrow the conversion definition to documented intent, in theory, and it starts spending money on buyers rather than talkers. That’s the theoretical win. The risk is volume. If the filter is too aggressive, bidding models lose enough conversion signal to wobble. Google has explicitly warned against sharp bid or budget moves in the first few weeks.
What to expect from your numbers
Your call conversions will drop. That’s the point. The AI is filtering out the noise your duration threshold was counting as wins. The reporting interface adds an AI-generated summary of each call alongside the hashtags, sitting inside the Call Details report. Useful for triage, but not a substitute for pulling the recording when a number looks off.
The sales team disconnect
Lead qualification used to be the sales team’s job. BDRs, intake coordinators, call centres, they existed precisely to sort noise from opportunity. Google has now moved a piece of that work upstream, without asking sales. The awkward consequence: Smart Bidding now optimises for conversational grammar, not commercial qualification. Your BDRs may start fielding more calls that sound like buyers but fail basic BANT (Budget, Authority, Need, Timeline). A caller can trip every semantic flag the LLM is watching for and still have no money, no decision-making power, and no real intent to buy.
Marketing directors now have a translation job they didn’t have last month. The hashtags coming out of Google need reconciling against the notes coming out of the CRM. When they disagree, and they will, sales will notice first, and blame the media plan.
The historical pattern
Familiar pattern. Broad match was going to save money through better query understanding. Smart Bidding was going to automate manual drudgery. Performance Max was going to uncover incremental demand. Each one, in practice, consolidated control inside the platform, traded transparency for scale, and produced results that looked fine in aggregate and uneven at the account level. AI-qualified call leads fit the shape exactly. A transparent, auditable metric, duration, gets replaced with an opaque, proprietary one. LLM judgement. The intent is arguably better. The mechanism demands blind trust.
It feels like one of the clearest examples yet of Google moving from just tracking behaviour to actually interpreting intent. In theory, that should sharpen up lead quality signals, but the flip side is you’re now one more step removed from how ‘quality’ is actually being defined.
The structural problem
Here’s where marketing needs to be honest with finance. Google owns the inventory. Google runs the auction. Google records the call. Google’s AI decides if the call was good. Google’s bidding algorithm then spends your money based on that self-generated grade.
That’s a closed loop. If Google’s model loosens its definition of “qualified” next quarter, your conversion rate rises, Smart Bidding bids harder, CPCs climb, and Google’s revenue grows. You only notice if you have independent pipeline data telling you the new “qualified” leads are worse than the old ones. There’s a reason this isn’t obvious from the outside. The party that owns the measure owns the outcome. The incentive isn’t to grade harshly, it’s to grade in a way that keeps the auction churning. The phrase “This caller sounds like a buyer” is one Google has every incentive to find more often, not less.
The counter-view is that this is how AI in paid media actually works. You teach the machine. You audit its output. You override when it’s wrong. Blanket rejection isn’t a strategy. Blind trust isn’t either.
Where the AI will break
Commercial phone calls don’t happen in recording studios. They happen on construction sites with wind in the background. From Glasgow on bad 4G. From bilingual callers who switch languages mid-sentence because it’s faster. An LLM trained largely on US corporate English will have opinions about a heavy Geordie dialect, a thick Valleys accent, or Multicultural London English on a noisy road. Some of those opinions will be wrong.
Now imagine. You’re a roofer in Bradford and a builder calls about a £40k job. Thick accent, noisy yard, half the call is him shouting “hang on” at someone off the phone. He hits every commercial trigger, “how soon can you start,” “what’s the price,” “can you come Monday,” but the recording is half-unintelligible. Branch A: the LLM tags this #HighIntent and Smart Bidding bids harder for that postcode. Branch B: the LLM can’t parse the audio and grades it as low-intent noise. Smart Bidding pulls back from Bradford.
Now flip that on its head. The polite caller from a quiet office, asking three rounds of clarifying questions about your insurance certifications, sounds exactly like the LLM’s training data of a qualified buyer. They get tagged #HighIntent with confidence. They never close. They were doing competitive research for a procurement role.
The question isn’t whether the new system is smarter. It’s whether surrendering the definition of “a good lead” to the company selling you the ads is a trade you want to make, without your own measurement running underneath it.
