AI Max for Search – when it works, and when it doesn’t

AI Max should be used, and it’s now far better than it was six months ago. That is where our search team has landed on Google’s keywordless layer in 2026.

Google account reps have been pushing AI Max on ecommerce search operators since the product opened to beta in spring 2025. For most of that time, the right answer was “not yet”. Through 2025, the matching logic was too loose, the operator controls were too thin, and our own test set across fashion and equestrian accounts showed AI Max underperforming broad match on most accounts. The product has moved. The clearest change is the Gemini-powered guidelines layer, which gives operators plain-language control over what AI Max says, what it matches against, and how it speaks to the target customer.

The Nest stance for 2026 is to switch the full AI Max feature set on for Generic Search and leave it on. Brand search stays on manual CPC.

The Generic Search rollout

After the A/B test set that ran through summer 2025, we rolled AI Max out across Generic Search on accounts where the structural conditions held: clean conversion tracking, first-party audiences imported, brand traffic split off into separate manual-CPC campaigns, and existing broad-match coverage that was not already exhausting the efficient demand available. Full feature set enabled.

The last of those conditions is the one the summer 2025 test set did most to clarify. The win/loss pattern split on search-category saturation. On smaller and less-covered categories, AI Max won, sometimes by a wide margin, because keywordless expansion found valuable traffic the loaded keywords were missing. Where broad match was already capturing the available efficient demand, the same enablement produced losses, because the algorithm widened reach into lower-quality versions of traffic the campaign had already been converting. The product has improved since, and the Gemini guidelines layer narrows the operator-control gap. AI Max still needs somewhere useful to expand into, though, and an account where broad match is already efficient on the available demand is not that place.

The 82-day pre-vs-post comparison:

The numbers read as scale, not efficiency. Spend went up by three quarters and revenue tracked it roughly one-to-one. ROAS gave back a little under 6%, comfortably above the target the account works to. The campaign was still budget-limited at the end of the window, which means there is more efficient demand AI Max can absorb before efficiency starts to bend. Performance Max and Shopping activity over the same period showed no cannibalisation; the additional spend was net new demand.

The 82-day comparison window is not arbitrary. Google’s smart bidding runs on a 30.4-day cycle, and a longer-term view is what allows the algorithm to smooth out daily and weekly noise. Run any AI Max change through at least one full cycle, ideally more, before drawing conclusions either way. Accounts that kill AI Max after a bad fortnight are usually reading noise.

What makes AI Max work: the inputs that matter

The read on Performance Max applies to AI Max in exactly the same way. It works when given the right data. It fails if you think Google can deliver without strategic guidance. The machine is not powerful enough on its own; you have to give it the tools. The best builder in the world cannot build anything without hammer and nails.

The algorithm is reading four input streams to decide which queries to bid on, what to bid, what to say in the ad, and where to send the click. Weak inputs cap AI Max performance regardless of how good the algorithm gets.

Conversion tracking. A single primary purchase conversion action is the only conversion the campaign should be bidding to. On-site funnel events like page view, add-to-cart, and form submission belong in remarketing audiences, not in the primary conversion column. If multiple primary conversions are firing, AI Max is bidding into a confused signal, and the keywordless layer will expand aggressively against the loudest event rather than the one that matters commercially.

First-party audiences. The customer segments imported from your CRM are the highest-quality signal you can hand the algorithm. Newsletter subscribers, loyalty members, purchasers split by value, by frequency, by category. The minimum viable list size for Search is 1,000 users. The richer the segmentation, the more precisely AI Max can route its expanded matching toward consumers who look like your converting base.

Product feed. AI Max can match against queries that have no direct keyword overlap with what you have loaded, and the feed is one of the places it draws context from. Front-weight the most important words in product titles. List colour as the main colour name (Blue, not Saxe). Use the product description field, which is not user-facing, as a place to load relevant search terms you want the campaign to be eligible for. A feed built only to user-facing copy standards is leaving matching surface on the table.

Creative assets. Asset generation works from what is already in the account, so the account needs assets at the ratios and lengths the platform actually serves. The headlines, descriptions, and images you upload set the ceiling on what AI Max can generate from. Thin asset libraries produce thin AI-generated ads.

Success on Google is determined now by the work that happens outside the platform. The options inside have narrowed enough that operators converge on the same defaults. Where accounts differ is in the four inputs above.

The Gemini-powered guidelines layer: what to actually write

The Gemini-powered interface is the operator lever most accounts running AI Max are not using to anywhere near its potential. It lives inside the AI Max settings on the campaign, takes plain-English input across three boxes, and shapes what the AI does with the matching, asset, and routing decisions you have just handed it.

Messaging guidelines. What ads should and should not say. Google ships the product with examples like “always mention free delivery” and “never say cheap”. Useful starting points, and most operators should be writing more. The promises that matter commercially go here: a specific delivery threshold, a returns position, a price-stance hard-no, a phrase you have spent budget building into the brand vocabulary. This box is operating across every ad the campaign generates, which makes it the most efficient single place to enforce tone.

Matching guidelines. The query boundaries the campaign should respect. This is where you tell AI Max what kind of search it should be matching against. For Generic Search, the simplest guardrail is one that keeps the campaign out of brand-search territory, because brand sits in a separate manual-CPC campaign and you do not want keywordless expansion pulling brand-name traffic back into the Generic Search budget. A second guardrail worth writing is one that keeps the campaign out of pure content or research queries that are not commercial in intent. The Matching guideline is what stops keywordless expansion drifting into the long tail of low-quality traffic.

A caution on how tight to write these. The same trap that catches operators on negative keywords applies inside this box: most are too quick to exclude relevant searches that haven’t directly converted. A relevant non-converting search can still be doing work earlier in the consideration cycle, or it can be flagging a relevancy gap on the landing page rather than a bad query. The matching guideline should keep AI Max out of structurally wrong traffic (brand, content, pure research) without stripping out every search that hasn’t converted in the last 30 days. Over-restrictive matching is the failure mode that wastes the keywordless layer’s reason for being there.

Audience guidelines. Plain-language description of the customer you are trying to reach. Google’s described use is tone and voice of generated copy. Treat it as the brief you would give a copywriter on a new account: who is this person, what are they shopping for, what are they trading off, what makes them choose this brand over the next one. The first-party audiences you have already imported do the targeting; the audience guideline is what tells AI Max how to write to them.

The preview step is not optional. Before pushing changes live, read the sample ads and the matched searches AI Max returns for your guideline set, and if either reads wrong, rewrite the guideline and preview again before going live.

Where AI Max does not belong: brand search

AI Max does not belong on brand search. None of the brand-side tests across our portfolio have shown AI Max outperforming a tight manual-CPC brand structure.

The mechanism is straightforward. Brand search captures demand that already exists; the searcher is typing your brand name. Target ROAS bidding bids aggressively against conversions the algorithm expects to win. On brand search, the algorithm expects to win most of them, because the user had intent before they searched. That confidence translates into higher CPCs than the demand needs. Manual CPC captures the same brand volume at materially lower cost.

There is a related structural point. Most advertisers should hold brand spend at no more than 15% of total search budget, or 25-30% for clients with multiple resellers, because brand traffic is largely non-incremental: the user would have found the site through organic or direct anyway. Letting AI Max expand inside brand makes that problem worse, because keywordless matching widens the brand-adjacent surface the algorithm can bid into. The structural answer goes the other way: split brand off, run it on manual CPC with tight match types, and put AI Max on the Generic Search budget where keywordless expansion has somewhere efficient to look.

FAQ