The advice to pair broad match with Smart Bidding is everywhere, and it is correct — but it quietly skips the precondition. Broad match plus Smart Bidding only suppresses weak queries once the campaign has enough conversions for the bidding model to steer. Below that floor the combination does not self-clean; it spends on the junk while it learns, and on the widest match type the junk is a large share of the spend. Google's own guidance strongly recommends using broad match only with a Smart Bidding strategy but publishes no conversion threshold for when that recommendation starts paying off. This post is about that missing number — why there is a floor, roughly where practitioners put it, how to tell which side of it your campaign is on, and what to do until you clear it.
The stakes are concrete. If you widen to broad match expecting Smart Bidding to handle query quality, and the campaign is below the floor, you have not built a discovery engine — you have removed the keyword boundary without installing the thing that was supposed to replace it. The fix is not to avoid broad match forever. It is to know the precondition, check it, and run manual search-term negation as the stopgap while the signal matures.
The short answer: no single number, but a real floor
There is no Google-confirmed conversion threshold for broad match, and anyone who quotes one as official is guessing. Google's documentation recommends broad match with Smart Bidding and claims advertisers who moved exact and phrase keywords to broad match in Target CPA campaigns saw about 25% more conversions, a figure from its own 2020 internal data — but it stops short of saying "you need N conversions first." That omission is not an oversight. The floor is not a fixed count; it is the point where your conversion signal becomes dense enough for the model to distinguish intent, and that depends on your conversion rate, sales cycle, and tracking quality.
So the useful framing is not "what is the magic number" but "does this campaign have enough recent conversions for Smart Bidding to steer a wide net." Practitioners have converged on rough bands precisely because the exact number is unknowable in advance: enough that the model sees patterns, not so few that it is bidding on noise. The rest of this post uses those bands as a gate, while being honest that they are rules of thumb calibrated by experience, not settings Google exposes or endorses.
Why the floor exists: the model needs signal to sort
Broad match works by handing query selection to Smart Bidding. Instead of you deciding which searches qualify, the match type invites a wide range of them and the bidding model decides, in each auction, whether a given query is worth bidding on. That decision is only as good as the conversion data behind it. With a trained signal, the model can look at a query it has never seen and estimate its conversion probability from patterns in what has converted before. With a thin signal, it cannot — so it bids broadly and learns from the results, which means paying for the queries that turn out to be worthless.
This is why the waste is front-loaded and concentrated on new or low-volume campaigns. The wider the match type, the more unfamiliar queries arrive, and the more the model is guessing rather than steering. A tighter match type limits the damage not because it bids smarter but because it lets fewer unknown queries in to begin with. The floor, then, is the amount of conversion history at which the model's guessing turns into steering — below it, broad match is a cost of training the algorithm; above it, broad match is the algorithm doing useful work. The interaction between negatives and Smart Bidding's own suppression only becomes favourable once you are on the right side of this line.
What the numbers actually are (and why they disagree)
The published rules of thumb disagree, and the disagreement is informative. Some agencies put the floor around 15 conversions per month per campaign, below which they say the algorithm cannot bid accurately on broad terms. Others treat 30 per month as the working minimum and 50 or more as the level where broad match reliably beats tighter types. Performance Max guidance often cites higher figures still. These are not contradictions so much as different risk tolerances applied to the same mechanism: the more conversions, the more confidently the model steers, and each practitioner draws the "confident enough" line in a different place.
Use the spread, not any single point. If a campaign is comfortably above the top of the range — say 50-plus conversions in the last 30 days — the signal is dense and broad match is on solid ground. If it is below the bottom — under roughly 15 in 30 days — no practitioner band supports running broad on Smart Bidding yet, and you should not. The grey zone between 15 and 50 is where judgement lives, and the tiebreaker is not the conversion count at all but the search terms report: a campaign in that band whose terms read as mostly relevant is steering; one whose terms read as mostly junk is not, whatever the number says.
How to tell which side of the floor you are on
Check it per campaign, in the UI, in two minutes. Open the campaign, set the date range to the last 30 days, and read the Conversions column — then compare that figure to the bands above. Do this per campaign rather than per account, because a healthy brand campaign and a cold prospecting campaign can sit on opposite sides of the floor inside the same account, and the account-level total hides that. A campaign carried by one high-volume ad group can look fine in aggregate while the broad-match ad group you actually care about is starving.
Then cross-check the count against reality by opening the search terms report for that campaign. The conversion count tells you whether the model has enough signal; the search terms tell you whether it is using that signal well. If the count clears the floor but the terms are full of off-intent queries, the signal is noisy — maybe conversions are firing on low-value actions, or tracking is counting things that do not matter to the business — and the model is steering toward the wrong target. In that case the fix is upstream, in your conversion setup, not in the match type. The two reads together are what separate "ready for broad match" from "a number that looks ready."
What to do below the floor: negate manually, stay tight
Below the floor, do two things: keep the match types tighter and run manual search-term negation as the active control. Default to exact and phrase match so the early search terms stay readable and the spend stays on-intent while you build conversion history. This is the same logic the match-type decision turns on — choose the width your account can absorb, and a thin-signal campaign cannot absorb broad's width yet. If you do run broad match below the floor for discovery reasons, treat it as a managed experiment with a tight budget guardrail, not as the campaign's main engine.
Manual negation is what substitutes for the steering the model cannot yet do. Open the search terms report at least weekly, sort by cost, and read the top terms the way you would if there were no algorithm at all — because below the floor there effectively is not one you can trust. Negate the irrelevant queries, group the recurring irrelevant patterns into shared negative lists so they carry forward, and watch the terms converge toward intent over the following weeks. This is exactly the discipline the unpaired broad match failure mode warns about: broad match without an active negative list is a leak, and below the floor the negative list is not a backstop, it is the primary mechanism.
As you clear the floor: widen deliberately, keep the cadence
Clearing the floor is a reason to widen, not a reason to stop reviewing. Once a campaign has a dense conversion signal, Smart Bidding can steer broad match toward converting queries you would never have added by hand, and that is the upside the whole combination exists for. Widen one campaign at a time, keep broad keywords in their own ad groups so you can still read which match type produced which query, and treat the first few weeks after widening as a period that needs a heavier negative pass, not a lighter one — the net just got wider, so more unfamiliar queries will arrive before the model settles.
And keep the weekly cadence permanently. Clearing the floor shifts negation from primary control to ongoing hygiene, but it never removes it, because Smart Bidding optimises for conversions rather than for relevance or margin. It will keep buying the occasional off-topic query that converts by luck, and only a human reading the report catches the queries that convert but should not. If you want a repeatable structure for that review rather than an ad-hoc scan, the broad match test protocol lays out the guardrails to run a widening as a controlled test. The campaign that wins on broad match is not the one that hit a magic conversion number — it is the one that checked the floor, matched its negation effort to which side it was on, and never mistook "Smart Bidding is on" for "the account is being managed."