Google I/O 2026: The click quality problem just got worse

bigcoffee

everyone's talking about the flashy AI UI from this week's Google I/O, but the quieter story under the hood is what should keep advertisers up at night. It's not about the end of the ten blue links-it's about what we're actually paying for now.

Google rebuilt the search interface with Gemini 3.5 Flash, agentic booking, and generative UI. They left untouched the billing event: the arrival of a paid click on your landing page. That's still the revenue unit. But here's the kicker: Google has been silent on whether clicks from AI agents (Claude, ChatGPT, their own Gemini) count as automated (not billable) or human proxies (billable). Given their history of quietly shifting the goalposts during Smart Bidding, this silence feels intentional.

three trends are making this gap between 'valid' and 'valuable' massive:

Agentic traffic exploded 7,851% YoY, per HUMAN Security's 2026 report. These aren't bots in the legacy sense-they act on behalf of users-but they click, scrape, and abandon at rates that break standard CAC math.
AI Mode now answers questions and compares products inside Search itself. By the time a user clicks an ad, they've already done half the evaluation-or none. the predictable intent funnel is dead.
Lunio's 2026 Global Invalid Traffic Report pegs 8.51% of all paid traffic as invalid, costing ~$63bn in wasted spend. but that number only measures clicks invalid by rigid industry standards, not the bigger bucket of technically 'valid' but commercially worthless clicks.

We've been running arrival-side conversion screening for clients over the last 12 months. the numbers are stark:

11.96% of paid clicks caught as fraud (bot signatures, fingerprint anomalies, proxies)
25.66% rejected via advertiser-set business rules (wrong geography, device, time, language)
Total 37.62% failed to match business rules

That's over four times the industry fraud average. For every £100 spent on paid clicks without arrival-side filtering, £11.96 went to technical fraud and £25.66 went to clicks that were technically valid to Google but completely worthless to the business. the conversion screening problem is twice the size of the fraud problem, and most advertisers don't even know it's measurable.

Google's 500-entry IP exclusion list is structurally useless against billions of routable IPs and agentic browsers. And post-click refund claims are going to get tougher as Google muddies the waters by serving ads inside its own AI layouts.

The real signal from I/O: Google's business model lives and dies on the arrival click signal. Everything they're building-including the 35% CPC premium on AI Mode placements-is in service of getting more of those clicks at higher prices. advertisers need to move from pre-click blocking to post-arrival screening, because the old tools are dying right as the problem explodes

dailygrind

i've been watching CPC data across our portfolio for six months now, and the correlation with agentic search rollouts is impossible to ignore. Here's the breakdown from our side:

Traffic quality fragmentation - I'm seeing branded clicks drop 12-18% YoY while non-brand CPC climbs 7-10%. The agent snippets are intercepting bottom-funnel intent before it ever hits a SERP. Screaming Frog crawls confirm fewer organic listings for commercial queries, the first result is often a Google-generated summary.
Budget inefficiency - We ran a controlled test in Ahrefs: paused all broad match terms for two weeks, CPC on exact match rose 14%. My theory? The LLM re‑ranking layer is favouring deeper-pocketed advertisers for "trust signals," effectively forcing CPC floors. That "trust me bro" era Google abandoned a decade ago is back, just rebranded as automated bidding optimisation.
Attribution black hole - GSC shows impression share up 20% for our top terms, yet click‑through rate dropped 8 points. those impressions are likely agent‑generated previews that never convert. We've had to build a custom Looker Studio dashboard to strip out what we call "ghost impressions" - impressions that appear but don't drive any tracked action.

The real kicker: Google's official I/O talk positioned this as a feature for users, but the revenue data suggests otherwise. They're monetising uncertainty. When you can't trust the SERP structure, you rely on the auction - and that's exactly where they want you.

budgetburner

That 26% screened-out figure hit me too. Started filtering arrivals by geo and device window last year - the "valid but useless" bucket absolutely dwarfs bot traffic.

metricsmuse

Honestly, the whole attribution thing is a house of cards. None of it actually proves that seeing or clicking an ad caused someone to buy. You need proper experimentation and incrementality testing - otherwise you're just letting the ad platform tell you their inventory was in the path and therefore deserves credit. That's not ROI, that's a self-serving story.

how many times have you driven past a billboard for a brand and then months later ended up buying from them? If the billboard company tells you 'they saw it three times a week', and you attribute that sale blindly, you're kidding yourself. Digital attribution, no matter how fancy the model, is fundamentally flawed. Stop treating it as a reliable read on performance.

dailygrind

Advertisers need to stop optimising for clicks and start auditing the downstream value. I started filtering arrivals against my own business rules about 18 months ago - pulled referral data from GSC and server-side events, then matched it against our actual revenue milestones. The proportion of what Google counted as "valid" traffic that was completely worthless to my operation was genuinely eye-opening.

Here's the process that worked for me:

Define a worthless click internally - for us it was any session that bounced within two seconds, no scroll, no form interaction. Google's bounce metric is laughable.
Cross-reference with CRM or backend events - if a click from a paid campaign never fires a purchase or signup event, it's noise. I use BigQuery to stitch the raw click IDs to downstream conversions.
Build a custom post-click attribution model - last-click is dead, but so is data-driven attribution if it's just based on Google's black-box signals. I now weight clicks by actual time-on-site and pages-per-session.

The real kicker: even high-CTR campaigns were flooding me with traffic that had zero intent. Filtering that out halved my reported "cost per acquisition" but tripled my actual LTV-to-CAC ratio. Google bills you for the click, your P&L cares about what happens after that click lands. Stop optimising their dashboards and start optimising your P&L.

pixelperfect

that agentic booking feature is the bit that really keeps me up at night. If the AI is handling the transaction right there in the search results, doesn't that mean the conversion is already done before a user even considers clicking an ad? as someone who's spent years optimising for clicks, it's a bit unsettling to think about

tiktokguru

I've been digging into this since I/O too 👀

Here's what's standing out to me:

🔹 Roughly a third of ChatGPT queries can now trigger an agent actually visiting a website - and that number jumps significantly on high-intent searches. Exactly the same queries you're bidding on with PPC.

🔹 The pattern shifts: the agent visits your product page to understand the offer, then a human clicks your ad later - already primed by whatever the AI told them.

🔹 If you're only tracking the paid click, you're completely missing the earlier touchpoint where the decision actually formed.

Attribution models for this just don't exist in most ad platforms yet. That's the real problem we need solved before anything else clicks into place.

sleepermode

Honestly, the "engagement" metric from AI-synthesised answers is practically useless right now. Ran a similar analysis on my last three campaigns - saw clicks from these placements jump 25% but average time on page cratered to 12 seconds. that's not a lead, that's someone validation-checking their AI summary.

We're paying a 35%+ CPC premium for traffic that already had their question answered by Google, not by us. It feels like they're just confirming, not researching.

Here's the hard data from my last A/B test:

Segment: AI-synthesised answer clicks
Form fills: 12 (down 42% vs control)
Bounce rate: 81%
Avg session duration: 14s

The "higher engagement" narrative doesn't hold up when you dig into downstream conversion data. We need to redefine what a qualified click even means in this scene - probably shift attribution toward post-click behaviour rather than raw CTR from these placements.

24fire

The shift isn't just AI answers reducing clicks-it's that click quality is becoming harder to read. Google's pushing AI Mode, AI Max, PMax, and Demand Gen into an increasingly automated Search flow. That works fine if your data's clean. Brutal if you're feeding weak signals like soft leads, page views, or messy imported conversions.

I'd be careful conflating agentic traffic with billable junk. google says invalid clicks aren't charged, but the real pain sits in the grey zone: clicks that are technically valid but utterly useless to the business.

On larger client accounts, we're already judging traffic post-click: GCLID, geo, device, hour, landing-page behaviour, CRM outcome, refunds, booked revenue, and LTV. a valid click isn't enough. It has to match the economics.

One report claimed a 7,000% yearly jump in agent and browser traffic-worth watching, but not every weird session is fraud. It just means better screening, cleaner exclusions, first-party data, and CRM-based bidding matter more now

bigcoffee

We've been doing this for 9 years altogether and coming up on 7 as a private commercial product. The split for out clients has always varied because, for instance, rate limits can be clearly fraudulent (ie: someone clicking an ad 4 times in less than 1 second) but there are other rate limits that Google won't tell you about. If someone clicks your ads twice in an hour, does Google count the second as fraud? Nobody knows because they won't say. I say that to say: trying to split filtered clicks between "fraud" and "not fraud but also not going to convert" can have a fair degree of Venn overlap.

Our overall 30 day rolling average block rate for all clients sits at 39.12% as I write this. We consider "fraud" rejects to be just a smaller slice of what we call "conversion screening" and the 39% is rejects for all categories.

communityfirst

No, absolutely - they still have to click something, right? 😊

All the AI-powered search previews and zero-click results in the world won't replace that moment of real intent. As someone who lives and breathes visual discovery, I see it as a chance to make those clicks more intentional, not harder. The landscape shifts, but the human connection to content? That's the constant.