“AI” in Google Ads is a marketing wrapper around a stack of distinct features. Smart Bidding, Broad Match with Smart Bidding, PMax asset generation, Auto-Applied Recommendations, Responsive Search Ads, Advantage+ Audience, Data-Driven Attribution, Optimised Targeting — each is a different tool with different economics. Grouping them under one AI banner is convenient for Google, but it means the honest answer to “should I use Google's AI?” is: which specific feature, on which specific account, with which specific conversion signal.
The pattern in 2026: features that consume account data to make bidding decisions (Smart Bidding, DDA) are typically excellent when the signal feeding them is clean. Features that generate creative or reallocate budgets on their own (asset generation, auto-applied recommendations) are usually cost-negative because they optimise for Google's metrics, not yours. Features that expand match or targeting silently (broad match with Smart Bidding, Optimised Targeting) split cleanly by account size and tracking quality.
The Honest Scorecard
| Feature | Default | Verdict |
|---|---|---|
| Smart Bidding (tCPA, tROAS, Max Conv, Max Conv Value) | On | Keep on — with clean conversion signal |
| Data-Driven Attribution | On | Keep on — the only reasonable choice in 2026 |
| Broad Match with Smart Bidding | Prompted | Depends — on for > 50 conv/mo, off otherwise |
| Auto-Applied Recommendations | On | Turn off — quietly reallocates budget and adds keywords |
| PMax Asset Generation (Gemini) | On | Turn off — generic outputs that hurt brand consistency |
| Responsive Search Ads (asset pinning) | Auto-mix | Pin — pin headline 1 & description 1 for consistency |
| Advantage+ Audience (audience signals in PMax) | On | Keep on — but feed it real audience data, not defaults |
| Optimised Targeting (Display / DemandGen) | On | Depends — on with tight negatives, off without |
| Auto-created Assets | On | Turn off — same problems as PMax asset generation |
| Search Term Insights (grouped queries) | On | Ignore — read but don't act; grouping is too aggressive |
What's striking about this list in 2026 is how much has become non-negotiable. Manual bidding is no longer a serious option for accounts > 30 conversions/month; DDA is the only attribution model that gets real product investment from Google; broad match paired with tCPA is now the default in every fresh account. The debate has moved from “use AI or not” to “which AI features earn their default status.”
Turn Off: Auto-Applied Recommendations
Google's Recommendations tab is a genuinely useful diagnostic view. Auto-Applied Recommendations — the feature that lets Google implement those recommendations on its own — is not. What Google's AI “applies” on the average account:
- Adds new keywords to campaigns (usually broad match, aggressive expansion)
- Adjusts budgets across campaigns (moves budget toward Google-defined winners)
- Pauses low-performing ads (based on internal quality metrics, not your CVR)
- Adds asset variations to RSAs and PMax (from Google's templates)
- Bid-strategy switches (moves campaigns from Max Clicks to Max Conv without conversion data)
Every one of those actions is a reasonable thing for a human strategist to consider — and a bad idea for Google to do unattended. Auto-Applied Recommendations has caused more account regressions in our audits than any other single feature. It reallocates spend during learning phases, introduces broad match without your review, and quietly writes to the account in ways that make change-history reads confusing weeks later.
Turn it off account-wide: Recommendations → Auto-apply → uncheck all boxes. Read the Recommendations tab manually as a diagnostic. Apply what makes sense; ignore what doesn't.
Turn Off: Gemini-Generated Asset Text and Images
PMax and RSA now offer to generate headlines, descriptions, and even images using Gemini. The pitch is a fully-populated asset set from a landing URL and a prompt. The output is usually competent, occasionally embarrassing, and always generic.
The problem isn't the quality of individual assets — it's the consistency and specificity you lose. A hand-written headline set has voice, specificity, and testable variation. Generated headlines regress to statistically-safe language: “Discover quality solutions,” “Trusted by professionals,” “Get started today.” They produce lower CTR than pinned hand-written alternatives on every A/B we've run this year, and they dilute what remains of your ad copy's brand signal.
Turn off: campaign or asset group settings → Automatically created assets → disable. Write your own headlines. Test them properly.
Depends: Broad Match with Smart Bidding
Google's default in 2026 is broad match paired with Smart Bidding. The pitch is that Smart Bidding will use CVR signal to bid low on irrelevant queries. In practice, this works well on some accounts and spectacularly badly on others. The dividing line is conversion volume.
| Account volume | Broad + Smart Bidding | Why |
|---|---|---|
| < 30 conv/mo per campaign | Bad idea | Smart Bidding lacks signal to filter noise; broad chases every query |
| 30–100 conv/mo per campaign | Cautious yes with negatives | Signal borderline; needs weekly negative-keyword hygiene |
| > 100 conv/mo per campaign | Strong yes | Signal dense enough for Smart Bidding to filter effectively |
For low-volume accounts, keep phrase match as default and use broad only on specific, high-intent themes. For high-volume accounts, broad + Smart Bidding does open real incremental audiences — but pairing it with weekly search-term reviews and quick negative additions is non-negotiable. Without hygiene, broad match becomes a slow leak within a quarter.
Keep On: Smart Bidding (with Preconditions)
Smart Bidding in 2026 is a mature product. tCPA, tROAS, Max Conv, Max Conv Value all work well — for accounts where the conversion signal feeding them reflects business truth. That's the qualifier that matters. On any account where conversions are noisy (form fills without CRM validation, soft actions marked as primary, cross-channel attribution broken), Smart Bidding will confidently optimise against the wrong signal.
The three preconditions we treat as non-negotiable before switching a campaign to Smart Bidding:
- Primary conversion is a real business outcome (booked call, sale, contract), not a soft signal
- Enhanced Conversions or Enhanced Offline Conversions are wired — see the EOC piece
- Consent Mode v2 firing updates so attribution isn't under-reported by 30–50% — see the Consent Mode piece
With those three in place, Smart Bidding earns its default status. Without them, we regularly find accounts where manual CPC beats tCPA — because tCPA at least commits to a bid; noisy tCPA thrashes and wastes budget.
Keep On: Data-Driven Attribution (Reluctantly)
DDA became the only attribution model in Google Ads in 2023 for accounts with sufficient data, and the alternative (last-click) is being phased out. In 2026, DDA is effectively mandatory. Two things to know:
- DDA weights are opaque — Google gives you no meaningful transparency into how credit is assigned per touchpoint
- DDA is more accurate than last-click on average for multi-touch funnels but occasionally over-weights top-of-funnel display touches in ways that don't match observed behaviour
Keep DDA on; it's the only choice. Do not treat DDA credit as gospel — validate against your CRM or GA4 Explore reports for multi-touch attribution when the numbers seem off.
Depends: Optimised Targeting (Display / DemandGen)
Optimised Targeting expands your Display or DemandGen audience beyond the audiences you selected, based on conversion signal. On accounts with tight negatives and strong conversion signal, it's a small but real multiplier. On accounts without either, it's a fast way to burn budget on irrelevant placements.
Account-level negatives with 30+ excluded categories (mobile games, kids'-content, MFA sites, dating). Without this, Optimised Targeting spends heavily on Google's network junk.
If your primary conversion is a soft signal, Optimised Targeting will happily expand to placements that hit the soft signal cheaply.
Placement report every 2 weeks; add junk domains as negatives; repeat. Without this, expanded audiences drift over time.
Ignore: Search Term Insights (Grouped)
Google now groups search terms into “themes” in the Search Terms report. The grouping is aggressive — themes lump together terms that behave very differently. You'll see a theme like “wooden furniture” that includes both “buy wooden chair” (buying intent) and “how to build wooden furniture” (DIY, zero intent), and the grouped CPA looks fine because the good terms subsidise the bad.
Ignore the grouped view. Read the raw search terms report at term level, paginate through with the search-term filter, and manage negatives at term level. The grouping is a marketing feature; the raw list is where the work happens.
Audit Checklist
What's Coming Next
Google's next AI push, based on beta features rolling out in 2026, is deeper into asset generation (video assets from a landing page, image variants from a product feed) and into audience expansion (Advantage+ Audience becoming the default on Search, not just PMax). Our expectation is that both will follow the same pattern the current features have followed: useful in narrow cases, actively bad on the wrong account, and marketed as universally beneficial.
The correct posture isn't to reject Google's AI features on principle — Smart Bidding and DDA earn their keep — but to evaluate each feature separately against your actual account state. Signal quality matters more than any single toggle. An account with clean tracking, real conversion signal, and disciplined negatives will benefit from Google's AI. An account without them will be quietly harmed by it. See how we approach these audits inside Google Ads management and conversion tracking audits.
