Turning AI Max on takes one click. Cleaning up after turning it on across a whole account takes about three weeks. I know the second number because I earned it: a home services account, nine Search campaigns, all switched over on a Tuesday afternoon because the projected lift in the recommendation panel looked free. Cost per lead sat around $90 going in. Eleven days later it was closer to $140, and the search terms report was full of queries for a service the client didn't sell in a state the client didn't operate in. Two of those nine campaigns had been exact-match-only for a reason nobody had written down. AI Max didn't break anything. It did precisely what it says it does, to campaigns that were built on the assumption nothing like it would ever be switched on.
So before the walkthrough, the thing Google's own interface soft-pedals: AI Max is not a feature. It's four features sharing a single toggle inside a standard Search campaign. Search term matching lets the campaign serve on queries your keyword list never contained, using the same keywordless matching that powers Performance Max, running alongside your keywords rather than replacing them. Text customization rewrites and generates headlines and descriptions tuned to the specific query. Final URL expansion sends the click to whichever page on your site Google judges a better match than the one you specified. And a set of new controls arrived with it, brand inclusions and exclusions, locations of interest, URL parameters for reporting, which exist because the first three features are wide enough to need fencing. Google's launch numbers for the beta were an average of roughly 14% more conversions at similar CPA, rising to around 27% for advertisers running mostly exact and phrase match. Both figures are averages across accounts that opted in. Yours is not the average.
Here's the part that matters more than any individual setting, and it's why this guide is ordered the way it is. AI Max changes the pool of traffic your campaign competes for. Your bid strategy decides what that traffic is worth. Those two systems learn at different speeds, and if you change the first without adjusting for the second, you get exactly what I got on that home services account: a bidding algorithm confidently paying full price for queries it has no history on. Setup order is the whole game. Controls first, expansion second, bidding untouched until you have data worth reacting to.
The pre-flight matters more than the setup, because AI Max inherits your account's existing sloppiness and then buys traffic with it. First, your negative keyword lists. If your account-level negatives are a graveyard list from 2022 with fourteen entries, expanded matching will find every gap in it within days. Pull your last 90 days of search terms, mine the categories you never wanted (jobs, free, DIY, competitor names you don't want to conquest, geographies you don't serve), and get them into a shared list applied at account level before you expand anything. Second, conversion tracking. AI Max is only as good as the signal it optimizes toward, and if you're counting form fills, phone clicks, brochure downloads and newsletter signups all as primary conversions, the expansion will find you a lot more of the cheapest one. Pick the conversion action that reflects revenue, make it the only primary, demote the rest to secondary. Third, look at your site. Final URL expansion sends clicks to whichever page Google thinks fits the query better than your chosen landing page, so go look at what it will find. If your site has twelve thin category pages and a blog full of 2019 posts, that feature is a liability until the pages are worth landing on.
Google moved this setting around during rollout, so the current path is worth stating plainly. In the Google Ads UI, open the Search campaign, go to Settings, and look under the campaign-level settings for the AI Max section. Turn it on there, then open the sub-settings before you save, because the defaults are set for maximum expansion and not for your comfort:
Then add the reporting plumbing, which almost nobody does and which is the difference between measuring AI Max and guessing about it. In the campaign's tracking template or final URL suffix, append a parameter that flags AI Max traffic, something like utm_source=google&utm_content=aimax, so your analytics can separate expanded traffic from keyword-matched traffic. Google's search terms report will show you the new queries and label the ones matched without a keyword, but it won't tell you what those sessions did on your site. Your own parameter will. Two weeks in, you want to answer one question: is the incremental traffic converting at a cost you'd have accepted if you'd bid on it deliberately? Without the parameter you're reading a blended campaign average and calling it evidence.
Smart Bidding prices every auction on a prediction: how likely is this specific impression to convert, and what is that conversion worth. The prediction is built from history. Your keywords, your landing pages, your conversion patterns by device, hour, location, audience. Switch on search term matching and the campaign starts entering auctions for queries with no history at all, so the model falls back on coarser signals: the account average, the vertical, whatever it can infer from semantic similarity to queries it does know. Coarse predictions produce mispriced clicks in both directions. Some are bargains. Some cost you $22 for a search that was never going to buy. On the accounts I've watched through this, expanded matching takes 10% to 25% of spend in the first week, and the CPA on that slice runs above the campaign average until the model has enough conversions to sharpen the forecast. That's not a defect. That's what learning on new inventory costs.
The wrong reaction is the instinctive one, and I've made it myself: CPA jumps, so you tighten the target. Drop your target CPA from $90 to $70 in week one and you don't punish the new traffic specifically, because the target applies to the whole campaign as an average. You throttle every auction the campaign was already winning profitably, volume drops, the model gets fewer conversions to learn from, and the expansion stays dumb for longer. You've made the problem slower to solve and paid for the privilege in lost leads. Leave the bid target where it is for a full 14 days. If the numbers genuinely can't absorb the turbulence, cut the exposure rather than the price: add negatives, tighten brand exclusions, switch off final URL expansion, restrict locations of interest. Those act on the new traffic only. The bid target acts on everything.
One more sequencing rule, because it's the cheapest mistake to avoid. Don't enable AI Max in the same week you change your budget, your bid strategy, or your conversion actions. Each of those disturbs Smart Bidding on its own, and stacked together you get a campaign in flux with no way to attribute the result to any single change. I give it a two-week gap either side. If someone tells you that's overly cautious, ask them how they plan to tell whether the AI Max test worked.
Here's the arithmetic nobody runs before flipping the toggle. Your campaign had 40 keywords pointing at four landing pages. After a month of expanded matching it's serving on 400 distinct query themes, still pointing at those same four pages. Conversion rate is a function of how closely the page answers the query, and you just multiplied query variety by ten while page coverage stayed flat. So the incremental traffic converts worse than your keyword traffic, you look at the blended CPA, and you conclude AI Max doesn't work. It worked. Your pages didn't. Final URL expansion is Google's attempt to patch this by picking a better existing page, which helps when a better page exists and does nothing when it doesn't.
This is also where the workload lands, and it's the part that gets quietly dropped after week three. Somebody has to read the search terms report every week, sort the new queries into keep and kill, write the negatives, and notice which winning query themes have no page built for them. On a single account that's two to four hours a week of genuinely mechanical work, and it never stops, because expansion keeps finding new queries. I spent years doing it manually at 1am and I don't miss it. Building a page per query theme by hand is worse: at $600 a page and a two-week dev queue, nobody does it, which is exactly why the four-page account stays a four-page account. The reason I ended up working on groas is that both halves of this are automatable. The query mining and negative work runs continuously instead of weekly, and the platform generates a landing page variant matched to each search intent, so expanded traffic lands on a page written for what the person typed rather than on the catch-all. That doesn't make AI Max risk-free. It removes the specific failure I described above, which is the most common one.
Before you enable anything, write down four numbers from the last 30 days of the test campaign: non-brand conversions, non-brand CPA, impression share lost to rank, and the share of conversions coming from queries that literally appear in your keyword list. Then run the campaign for 30 days with the controls set as described and the bid target untouched. At the end, you're not comparing blended CPA to blended CPA. You're comparing the expanded slice, isolated by the URL parameter you added, against the keyword slice. Three outcomes, three responses. If the expanded slice converts within about 20% of your keyword CPA, keep AI Max on and start building pages for the query themes it found. If it converts at double your keyword CPA but the search terms look right, you have a landing page and offer problem, not a matching problem, and turning AI Max off just hides it. If the search terms themselves are wrong, the fix is controls: locations of interest, brand exclusions, a harder negative list. Only after all three of those have failed does turning it off become the answer.
A few things people keep asking that the help docs answer badly. AI Max does not replace your match types; your exact-match keywords still match exactly and still get priority when they're the best fit, so you can run it on an exact-only campaign without converting anything to broad. You can run it while Performance Max is live in the same account, but check your brand exclusions on both sides first, because two systems with permission to expand into the same query space will find your brand terms and bid against each other with your money. And if you open campaign settings and the AI Max section isn't there, it hasn't reached your account yet. Don't email your rep about it. It'll appear.
That home services account did eventually get AI Max back, on two campaigns instead of nine, with locations of interest set to the counties the client actually serves and a negative list twice as long as the original. Cost per lead settled back near where it started and lead volume came up. The feature was never the problem. Nine campaigns, no controls, and a Tuesday afternoon were the problem. If you take one thing from this before you touch anything: open your test campaign, find the impression share lost to budget column, and look at it. If that number is high, close the tab and go fix your budget allocation instead. AI Max cannot buy you traffic you already can't afford.
