Target CPA for an AI to work from: one hard number or a range?

Pasting one target CPA into an assistant, say 40, and telling it to hold every campaign there is the wrong brief, I think. A single figure reads as a hard line, so the AI cuts bids hard on anything that drifts above it, even when the campaign is still learning or the volume is too thin to judge. A band like 35 to 50, plus a note on which campaigns matter most, lets it weigh a noisy week against an actual trend. The catch is that a range makes the reasoning fuzzier, so ask for the raw numbers behind each suggestion. Do you set a hard number or a band, and what do you do with campaigns under 30 conversions a month?

Honestly, it depends on whether your 40 is a break-even number or a hope. If it's the most you can pay per lead before you lose money, the AI should treat it as a ceiling and cut bids when a campaign stays above it for a full learning window. If it's a goal you'd like to hit, a band makes more sense, because then the AI can judge drift against a range instead of reacting to every noisy week.

The account only accepts a single tCPA per campaign, so the band lives in the instructions and the hard number lives in the settings. Those can disagree, and that's where the trouble starts.

On campaigns under 30 conversions a month, I wouldn't let the AI touch the bid at all. At that volume one expensive week can swing the cost per conversion by half, so any judgment is mostly noise. Pool a few similar campaigns, or stretch the window to 60 or 90 days, and only act on what survives that.

Over that 60 to 90 day window, make sure the CPA the AI reads is built from the same conversion actions the tCPA is bidding on. If a form fill and a call both count as conversions but only one is set as primary, the report CPA and the bidding CPA can drift apart, and the AI ends up judging the wrong number.

Agreed on the mismatch, but for local services the bigger problem is what actually counts as a call. If your call tracking logs every ring as a conversion, wrong numbers and hang-ups pad the volume and make the CPA look better than it is. How much that matters comes down to your minimum call duration. A plumber might be fine at 30 seconds, while a clinic booking appointments could reasonably go to 60 or more. Shorter thresholds feed the AI more data but with more noise, and longer ones give cleaner numbers but fewer conversions. So set that threshold first, before you let the AI judge anything.

I wouldn't treat the call threshold as settled before the AI starts judging. For a plumber, a 30-second call is often a real lead, so the cutoff matters less than which calls bidding actually optimizes on. Before you pick a number, look at how many calls in each duration bucket end up as booked jobs.

The bucket check beats a rule of thumb, but only if you can match calls to jobs in the first place. If your CRM or booking tool records which calls turned into jobs, you can compare call-to-job rates per duration bucket and set the cutoff from that. If a lot of your work comes in by text, a web form, or a walk-in with no call record, the buckets only show part of the picture. A 30-second cutoff can look like noise when it's really leads that never got logged. In that case I'd keep the threshold loose and watch the trend instead of trusting any single bucket.

I disagree with keeping the threshold loose and watching the trend. If a big share of your jobs never gets logged as a call, the trend only shows the logged part moving. The AI will still optimize toward whatever does get recorded, so a loose 30-second cutoff just teaches it to chase padded calls, which is Marco's problem again, only harder to spot.

I'd pick a rough threshold now and fix the gap at the same time. Add a required "how did you hear about us" field to the booking form, have staff tag jobs that came in by text or walk-in, and check each month what share of your work is still untracked. If that share is large, fix the conversion data before touching bids. Once most jobs are logged, the duration buckets start telling you something real.

Agree with the core of this. Bidding can only learn from what the platform counts as a conversion, so if booked jobs never make it back, the AI is optimizing blind. What I'd add is uploading those booked jobs as offline conversions. In Google Ads you match them to the click ID (gclid) you already capture on web forms, and you can make the booked job the primary conversion. Then tCPA works from real jobs, and the duration debate matters much less, because the bids stop rewarding a 30-second ring that never became work.

Two caveats. Walk-ins and texts won't carry a click ID, so that share stays untracked no matter what. And "how did you hear about us" answers tend to say "Google" for everything, so ask for the specific ad or keyword, not just the channel.