AI Max isn’t a campaign you set and forget. It’s a wild horse you wrangle.
Hand Google’s AI more control — AI Max, broad match, Performance Max — and it will generate thousands of search terms a week. For B2B SaaS, most of them are junk: multiple stakeholders, a precise buyer, tiny bottom-funnel volume, zero impulse purchases. You cannot hand-block your way out of that. The fix is a search-term evaluation matrix — a rubric that scores every term on intent, ICP fit, funnel stage, and match type, then routes it to a verdict: keep, negate, or mine. Clicks never count. Used after your account has real data flowing, that matrix let us pull opportunities worth $30,000–$70,000 at $1,000–$2,000 cost-per-opportunity — at or below our core exact-match keywords — while killing junk terms by the thousands. It will never hit zero junk. Nothing will. You manage it, you don’t cure it.
Here’s the whole system, given away.
Why does AI Max flood B2B SaaS accounts with junk search terms?
Because B2B SaaS is the worst-case scenario for uncapped matching. You’re selling software that needs a champion, an economic buyer, and usually IT to all nod along. Nobody impulse-buys a $40k/year platform the way they impulse-buy sneakers. So when you let the machine chase “conversions” across the open web, it finds the cheapest thing that looks like a conversion — and cheap almost never means qualified.
Want proof it’s happening in your account? Sort your AI Max campaign’s search terms by impressions and look at the top 100. In nearly every B2B account I audit, those 100 terms are ~90% of the volume — and they’re almost all one-word searches. Single words are broad, navigational, and rarely attached to a real buying intent.
I audited an e-commerce-automation SaaS that was showing up for the one-word query “OpenAI.” Not because they’re an AI company — they’re not — but because the machine saw “automation” and “AI tooling” as adjacent and went fishing. That’s not a glitch. That’s the predictable failure mode. Even if the person searching “OpenAI” were a perfect-ICP buyer (they’re almost never — they’re navigating to a website), they will not convert on that term, and the click still costs you.
It gets more expensive than a wasted click. One account was spending $10,000 a month on AI Max and getting leads — cheap ones, at a cost-per-lead that looked incredible on the dashboard. Then you open the CRM. Every one was a Gmail address. Nobody replied. Nobody showed up to the meetings they’d booked themselves. Nobody started a trial. Leads, sure. Pipeline, no. We pulled them off AI Max entirely and waited until the account had enough real data to try again. (This is the core disease behind most accounts we audit — your CPL can look fine while pipeline quietly stalls.)
So why not just run exact match and skip the whole mess? Fair question, and for years my answer was “yeah, basically.” But bottom-funnel volume in B2B SaaS is tiny, and your buyers don’t search the way your keyword list predicts. There are people in a buying window right now phrasing it in ways you’d never think to bid on. AI Max is where Google is putting its reach — it’s already auto-upgrading Dynamic Search Ads and campaign-level broad match into it — so this is coming whether you invited it or not. The matrix is how you get AI’s reach and exact-match discipline at the same time.
This is also why the matrix earns its keep even if you never touch AI Max: it’s the same rubric you should be running on your broad and phrase terms in normal campaigns. Search-term discipline is one pillar of a B2B SaaS PPC program that’s built for pipeline, not vanity volume — and this is where campaign structure is heading. Learn the model now.
Why doesn’t hand-adding negative keywords work anymore?
Because you’re refereeing something like 6,000 unique queries, and a block-list you babysit by hand doesn’t scale to that. Worse, you can’t even see all of them. Google only shows you a fraction of the terms you actually paid for — in our accounts it ranges from about a third to 70%, never the full picture. That’s not a bug; Google withholds low-volume and privacy-sensitive terms by design. The stingy long-tail — the exact stuff you’d most want to see — is the stuff most likely to be hidden.
So the job changes. You stop trying to block every bad term and start scoring every term against criteria the machine’s output has to pass through. Same instinct a great media buyer has always had — “is this the right person?” — just turned into a repeatable rubric instead of a gut call you make 6,000 times and get tired halfway through.
I’ll be honest about my own arc here: I used to be flatly anti-AI-Max and anti-broad. “That’s foolish, don’t do it.” Then Google kept shoving reach into it, clients kept asking, and I got tired of leaving volume on the table. The matrix is my answer. AI Max isn’t a solution for a campaign — it’s a layer you wrangle after data is flowing, to build better, more precise campaigns underneath it.
What is a search-term matrix — and why aren’t clicks in it?
It’s a rubric that scores every term before any verdict gets rendered, then outputs one of three calls: KEEP, NEGATE, or MINE (opportunity). The inputs are deliberately few:
- Intent (1–10) — how close is this term to actually buying?
- ICP fit (1–10) — does this look like your ideal customer, not just any human?
- Theme (1–10) — for standard campaigns, how relevant is it to that ad group’s core intent?
- Alignment (direct / adjacent / off) — for AI Max, how much category overlap is there?
- Conversions — the one performance signal allowed to override a verdict. More on that below.
Notice what’s missing: clicks. Clicks and click-through rate are not inputs, on purpose. I’ll defend that hard in a minute, because it’s the rule people fight me on most.
Who actually does the scoring?
Not a person eyeballing 6,000 rows. A cheap AI model judges each term one axis at a time — one keyword, one column — against a written rubric. That produces the scores. A deterministic set of tables (the ones below) turns those scores into a verdict. And then a human takes the ambiguous middle. In practice the system sorts every term into three piles: auto-negate, auto-allow, and a human-gated middle it hands to me with a category and a reason, so I can dig into the genuinely ambiguous ones instead of all of them.
That division of labor is the whole point. The AI handles volume. The tables handle consistency. The human handles judgment. Miss any one of those and you’re back to junk in, junk out.
How do you score a term? The standard matrix
Two tables, layered. First decides eligibility, second decides what happens to the misses.
The intent floor (funnel × match type). A term’s match type sets which column it lands in. The keep-threshold rises with buying intent and relaxes as match loosens — because looser match casts a wider net, so you forgive it a slightly lower bar:
| Funnel stage | Exact | Phrase | Broad |
|---|---|---|---|
| Bottom-funnel | 7 | 6 | 5 |
| Mid-funnel | 5 | 4 | 3 |
| Top / non-brand | 2 | 1 | 1 |
Read it plainly: a bottom-funnel exact-match term needs an intent score of 7+ to stay. The same theme running on broad only needs a 5. A “high-intent” campaign adds +1 to the whole column. That’s it — no mystery.
The ICP gate (what happens to a term that misses the floor). Intent decides whether a term is good enough; ICP decides how gently or harshly you handle the ones that aren’t:
- ICP 1–4 → auto-negate. No human needed. (On broad-match groups, even these get a human glance — bigger blast radius.)
- ICP 5–6 → human review. Surfaced to you with its score and the reason.
- ICP 7–10 → keep-eligible. Still has to clear the intent floor, but the door’s open.
A term can look off-intent and still be worth a human’s two seconds if the ICP fit is high. That nuance is what a plain negative list can’t give you.
How is AI Max different? You mine, you don’t just negate
Here’s where the goal flips. In a normal campaign you’re negating the off-theme stuff. In AI Max, the match-type table goes out the window and the goal becomes mine the valuable terms. It runs on pure ICP × Funnel, with how high a term can score capped by its alignment (direct-category terms uncapped, adjacent ones limited, off-category ones capped low):
| ICP ↓ \ Funnel → | Low (1–3) | Mid (4–6) | High (7+) |
|---|---|---|---|
| 7–10 | KEEP | KEEP | KEEP + MINE |
| 4–6 | NEGATE | REVIEW | KEEP + MINE |
| 1–3 | NEGATE | NEGATE | NEGATE |
Then one guard rail on top: any AI Max term with even one conversion is handed to a human, never auto-killed. Under expansive matching, converting terms are rare and precious. You don’t let a table murder one on a technicality.
So what is AI Max actually for, if most of it is junk? Mostly it re-finds terms you already want — and buys them at their cheapest available click, for the best-fit audience, at the right moment. (To be clear: that’s cost-efficiency of the buy. It is not clicks-as-a-quality-signal — a term can be cheap and still wrong.) Occasionally it surfaces something you could never have targeted: I’ve seen a term like “AI e-commerce image iteration software for DTC business” — one impression, one click, one conversion. Try to add that keyword yourself and Google tells you it has no volume. AI Max found it anyway. That’s the upside you’re paying the junk tax for.
Why are clicks and CTR worthless signals here?
Because people click your ad regardless of whether they’ll ever buy. You evaluate a keyword on fit for your product, not on click-through rate. Full stop.
The “OpenAI” term I mentioned? In the (incomplete) search data it had about 15 clicks and a 10–11% CTR. That’s a great CTR. It’s also a completely worthless term — 999 out of 1,000 people searching “OpenAI” are navigating to a website, not evaluating e-commerce automation software. If you judged that keyword by CTR you’d scale it. (Footnote to the story: that account became a client, five-figure-plus ACV. The buyer was real. The search term still wasn’t.)
This is why your ad setup matters more than people think. Start AI Max with structured, controlled ads — not auto-generated AI headlines or keyword insertion. If you let the machine write flattering headlines that juice CTR on junk queries, you’re literally training it to chase more junk. Give it tight, honest creative so a good click-through rate can’t lie to it.
And the single highest-ROI negative you’ll ever add: your own brand. AI Max will happily run for your brand name, convert on it, and report itself a genius. Negate every spelling, misspelling, product, and sub-product across all campaigns — unless it’s a dedicated brand campaign. Otherwise brand traffic hides inside AI Max and inflates the whole thing.
What actually improves — and what never will?
The metric that moves is cost per opportunity, not cost per lead. Cheap leads that never become pipeline are the entire trap; if you optimize to CPL you optimize to the trap.
On that account we brought back: once we killed brand and let the matrix run, cost-per-opportunity volume went up — that was the goal, more real shots on goal. We’d braced for a higher cost-per-opp as the price of reach. Instead it came in as low as, and sometimes lower than, our core bottom-funnel exact-match keywords — opportunities worth $30,000–$70,000 acquired at $1,000–$2,000 each. One account isn’t a law of physics, so here’s the honest generalization: it worked well enough that I now run AI Max across several clients and multiple campaigns. That’s the proof it travels, not the single dollar figure.
Now the part most agencies won’t tell you: the junk rate may never drop to zero. Even with all of this negation running weekly, you will keep surfacing genuinely garbage terms — maybe one in a thousand slips through gorgeous and useless. You don’t eliminate the junk. You shorten the amplitude — you reduce the swing between the good stretches and the bad ones. If you can’t stomach the occasional garbage term, don’t run AI Max. That’s a real answer, not a dodge.
Two non-negotiables make the whole thing work:
- Feed ICP signal back from your CRM. Not “any lead” — ICP-qualified. If you’ve got an ICP rating and offline / server-side conversion tracking, send a custom conversion back to Google that says “this one was actually our buyer.” The machine can only learn quality if you teach it what quality is.
- Don’t expect it to invent demand. AI Max is a volume play — “we’ve got extra budget and want more of the right people.” It will not conjure category traffic that doesn’t exist. If nobody searches for what you do, no amount of AI fixes that. (Google’s own pitch is ~14% more conversions at a similar CPA; in B2B, whether those conversions are pipeline is entirely on you and your feedback loop.)
Start this week — without hiring anyone
You have the whole model now. Three moves you can make today:
- Negate every brand variant, everywhere (except a real brand campaign). Instant clarity.
- Sort your AI Max/broad search terms by impressions and eyeball the one-word junk at the top. You’ll feel the problem in about ninety seconds.
- Set your intent floor by funnel × match type using the table above, and decide your ICP auto-negate cutoff. Now every term has a home.
That’s the mental model. Running it as an automated workflow across thousands of terms every week — cheap AI scoring each axis, exact negatives generated then carefully expanded to broad negatives that won’t nuke your good terms, a human on the ambiguous middle — is the part that eats your Thursdays. If you’d rather it ate mine, that’s what we do. Either way: don’t hand the machine the keys without a matrix in the passenger seat.