When we first wrote this article, machine learning in Google Ads was a set of optional features you switched on if you were feeling brave. In 2026 it is the platform. Responsive search ads are the only search ad format. Smart Bidding runs most budgets. Data-driven attribution is the default. Performance Max sits in nearly every ecommerce account. The question has changed from “should I use the automation” to “where do I hand over control, and where do I keep guardrails”.
This is a practical tour of the tools, in the order we introduce them on a client account. For each: what it does, when it works, and what to watch.
- Smart Bidding
- Responsive search ads
- Broad match with Smart Bidding
- Performance Max
- Demand Gen and lookalike segments
- AI Max for Search campaigns
- Dynamic Search Ads
- Data-driven attribution and conversion measurement
Smart Bidding
Manual CPC asks you to set a bid per keyword. Smart Bidding asks you for a goal and sets a different bid in every auction using signals you cannot see or act on: device, location, time, browser, audience membership, the query itself and how it combines with everything else. No human can do that at auction speed, which is why this is where machine learning earns its place.
The strategies map to the funnel much as they did before:
- Target impression share for awareness and brand defence.
- Maximise clicks for consideration, or for new accounts with no conversion history.
- Maximise conversions, with an optional target CPA, for lead generation.
- Maximise conversion value, with an optional target ROAS, for ecommerce and any account that passes values.
Target CPA and target ROAS are no longer separate strategies. They are optional targets inside the two Maximise strategies. The practical difference: without a target, the system spends the whole budget. With a target, it holds back when the target cannot be met. Start with a target close to your recent actual, then tighten in small steps and let each change finish learning before judging it.
Three guardrails we use on every account:
- Experiments before switching. Google Ads Experiments splits a campaign into a control and a test arm. Run the new bid strategy in the test arm for several weeks and compare on cost per conversion, not clicks. Our experiments guide and experiments case study show how.
- Seasonality adjustments and data exclusions. Tell the bidder about a sale before it starts, and exclude the fortnight your tracking broke, or it will learn the wrong lesson.
- Conversion volume. Smart Bidding needs steady conversion data. Thin campaigns sit in learning indefinitely. Consolidate campaigns or use a portfolio strategy so the model sees enough signal.
A note on ad rotation, which had its own section in the original. With Smart Bidding on, rotation is optimised automatically. If you still run manual CPC, choose Optimise.
Responsive search ads
You supply up to fifteen headlines and four descriptions. Google assembles combinations per auction and leans into the ones that perform. Manual A/B testing of two fixed ads is gone, and for good reason: it was slow, it only tested one variable, and someone had to remember to act on the result.
What has improved since the early days:
- Asset ratings. Each headline and description gets a Low, Good or Best rating once it has enough impressions. Swap out the Low assets, keep the Best, and write new ones against the themes that are winning.
- Combination reporting. Still thin. You see the combinations served and their impressions, not conversions per combination. Judge the ad as a whole against other RSAs in the ad group.
- Ad strength. Useful as a hygiene check, not as a target. An Excellent rating with unpinned, generic headlines can lose to a Good rating with disciplined pinning.
Pinning is the control lever. For lawyers, financial services and anything with compliance wording, pin the mandatory line to a fixed position and let the rest rotate. You give up some optimisation and keep your regulator happy. Review the automatically created assets setting too. It lets Google generate headlines from your landing page, which is fine for a retailer and a risk for a regulated firm. Our note on Quality Score for legal ads covers the trade-off.
Broad match with Smart Bidding
Broad match used to be the keyword type you added by mistake. Modern broad match uses the bid strategy’s signals, the landing page content and the other keywords in the ad group to decide what a query means, and it only makes sense paired with Smart Bidding and clean conversion tracking. Google suggests it by default on new campaigns.
When it works, it finds converting queries no keyword list would have included. When it fails, it fails quietly, through lead quality rather than cost. Rules:
- Only with Maximise conversions or Maximise conversion value, never with manual bidding.
- Use the broad match experiment feature rather than flipping the whole campaign.
- Set brand inclusions or exclusions so brand and generic budgets stay separate.
- Keep reviewing the search terms report and adding negatives. Automation widens the net; you still decide what to throw back.
- For lead generation, import offline conversions or lead quality signals. Otherwise the bidder optimises for form fills, and the cheapest form fills are the worst leads.
Performance Max
One campaign that serves across Search, Shopping, YouTube, Display, Discover, Gmail and Maps, driven by a product feed, your assets and your conversion goals. For ecommerce accounts with a healthy feed it has replaced Smart Shopping entirely and usually sits alongside a Standard Shopping campaign for products you want to control by hand. Our Google Shopping management page explains how we structure that.
The early complaint was that PMax was a black box. The controls have caught up:
- Brand exclusions, so PMax does not cannibalise your brand search campaign.
- Campaign-level negative keywords.
- Search themes to steer which queries the campaign chases.
- Channel performance and asset group reporting, so you can see whether spend went to Shopping or to Display.
- Search terms insights and placement reports for where ads actually showed.
- Final URL expansion settings and URL exclusions.
Use it when you have a feed and conversion volume. Be cautious for lead generation without lead quality feedback, where it will happily fill a form with anyone. Give it complete asset groups: all image sizes, a video (it will make one for you if you do not, and you will not like it), and headlines written for the audience, not the product code.
Demand Gen and lookalike segments
Demand Gen replaced Discovery campaigns and absorbed Video Action campaigns. It runs on YouTube including Shorts, Discover, Gmail and the Display Network, with image and video creative. It is the home of Google’s prospecting and the only place lookalike segments still exist since similar audiences were retired from Search and Display.
The lookalike method from the original article still holds. Build a seed audience of your most valuable users, let the model find people who behave like them, and show that group a dedicated creative. Good seeds:
- Customer Match lists of actual buyers or qualified leads, uploaded from your CRM.
- A GA4 audience of highly engaged visitors. We typically define that as people who spent several minutes on site, visited multiple service pages, or downloaded a guide.
- YouTube viewers of your own videos.
Start narrow, prove the conversion rate, then widen. Pair with Maximise conversions or a target CPA, and read view-through as well as click-through results, because the format is closer to Meta than to Search.
AI Max for Search campaigns
AI Max is the newest of the set. It is a bundle of settings on a standard Search campaign that lets the system go beyond your keywords in three ways: search term matching that uses your landing pages and ads to find relevant queries, text customisation that generates headlines and descriptions from your landing page content, and final URL expansion that sends people to the most relevant page on your site rather than the one you chose.
It comes with controls: brand inclusions and exclusions, locations of interest, URL exclusions, and reporting that shows the search terms and the generated headlines it served. Our view:
- Treat it as an experiment, not a default. The Google Ads Experiments tool supports an AI Max test arm.
- Review the generated headlines weekly for the first month. For regulated clients, consider leaving text customisation off and using only the matching component.
- It overlaps with Dynamic Search Ads. If you turn it on, your DSA campaign’s role shrinks.
Dynamic Search Ads
DSA crawls your site, matches a query to the most relevant page and writes the headline from that page. The teaching from the original article is unchanged: use it to cover the long tail on large sites, point it at product and service pages only, exclude blog and category pages, and keep a strict negative list because it will otherwise write a competitor’s name into your headline.
In 2026 DSA sits between two newer tools. Performance Max covers product discovery for retailers with a feed, and AI Max covers keywordless matching on search campaigns. DSA still makes sense for large service catalogues without a feed, and for accounts where you want a keywordless layer with more predictable behaviour than AI Max. If you run DSA and AI Max together, expect them to compete for the same queries.
Data-driven attribution and conversion measurement
Data-driven attribution is the default model in Google Ads and GA4. It compares the paths of people who converted with the paths of people who did not, and gives more credit to the touchpoints that actually changed the outcome. Last click is still available. The rules-based models in between were removed, so the choice is binary.
Use data-driven. It feeds Smart Bidding a truer picture of which clicks matter, which is the end-to-end loop the original article described: the model learns from the conversion path and bids accordingly. It is still opaque, and you will not see why a particular keyword got a fraction of a conversion. Accept that and audit the inputs instead:
- Enhanced conversions for web, so conversions still match after cookie loss.
- Consent mode where a cookie banner exists, so modelled conversions fill the gaps.
- Offline conversion import for lead generation, so the bidder learns from closed deals rather than form fills.
- One source per conversion action, with primary and secondary status set correctly. Double counting from a Google Ads tag and a GA4 import at the same time is the most common fault we find in audits.
If you want to go a level deeper and test your own models against Google’s, our article on regression analysis for Google Ads shows what is possible with an account export.
What to leave off
Not every automated setting deserves a yes. Four we turn off or review on every new account:
- Auto-apply recommendations. Google will add keywords, raise budgets and switch bid strategies on your behalf. Review recommendations weekly instead.
- Display expansion on Search campaigns. Almost always wasted spend.
- Optimised targeting on remarketing. If a campaign exists to reach a specific list, do not let it expand beyond that list.
- Automatically created assets for any client where wording is regulated.
Where to start
In order, for an account that has not touched most of this:
- Fix measurement: key events, enhanced conversions, consent mode, offline import if you sell by phone or sales team.
- Rebuild responsive search ads using asset ratings. This costs nothing and improves everything downstream.
- Move to Maximise conversions or conversion value through an experiment, then add a target.
- Run a broad match experiment on your best-performing campaign.
- Launch Performance Max if you have a feed, Demand Gen if you need prospecting.
- Test AI Max last, once the account has clean data and you have time to read the generated copy.
Each step gives the next one better data. Skip the first and the rest optimise towards noise. This is the sequence our Google Ads management team runs on every new account.
