Weekly Search & AI Digest: Targets Stop Being Decorative

Four things happened this week that change what a marketing team should be measuring, and one of them changes what it is paid on.

Google's 17 August bidding change reached the accounts that had been quietly benefiting from it. OpenAI's advertising business crossed into 31 European markets. Cloudflare moved its AI payment scheme from charging per crawl to paying per answer. And a fresh set of crawl-to-referral figures put a number on an exchange that most publishers have been describing in adjectives.

Each of these is easy to summarise and easy to misread. What follows is the summary, and then the part that matters more: what each number can and cannot be used to claim.

1. Google Ads stopped over-delivering, and some CPAs doubled

From 17 August, campaigns that are limited by budget on Target CPA, Target ROAS, or Target CPC for Demand Gen optimise more literally toward the target that was typed in. Before the change, Smart Bidding on a budget-limited campaign cherry-picked the cheapest auctions it could find and quietly beat the target: a £35 target CPA landing at £22, a 4× target ROAS landing at 5.5×.

That over-delivery is now going away. The £22 drifts back toward the £35 that was asked for. Optmyzr and PPC Land both reported accounts where effective CPAs moved by a factor of two in the days after the change, and the scope is wide: Search, Shopping, Performance Max, Demand Gen and Travel, through the interface, Search Ads 360, DV360, Editor and the API.

The uncomfortable part is that nothing broke. The system is now doing exactly what it was told. What has been exposed is that a lot of targets were never really targets — they were ceilings nobody expected to hit, and the gap between the number typed and the number delivered was doing work that nobody had written down.

The action is not to lower the target until the old number comes back. That reinstates the accident. The action is to work out what CPA the business can actually afford, type that, and let the budget find the volume — which is what the setting always claimed to do.

2. ChatGPT ads reached 31 European markets

On 24 August, OpenAI began serving ads across 31 European markets, six months after the US launch. European buying starts through the major agency groups, with self-serve promised later.

Alongside it, Similarweb launched an ad-intelligence data set covering placements across ChatGPT, Google AI Mode and AI Overviews, and published the density figures that come with it: roughly 26% of ChatGPT responses on the Free and Go tiers carry a sponsored placement, close to 30% of ad-eligible Google AI Mode queries show ads, and more than 40% of Google searches now trigger an AI Overview.

For anyone measuring brand visibility in assistants, this is a change in what the surface is. A cited source and a paid placement now sit in the same answer, and a visibility score that counts one without distinguishing the other has quietly become two different numbers wearing one label.

3. Cloudflare moved from pay-per-crawl to pay-per-answer

Cloudflare stopped charging AI companies per crawl and started paying publishers per answer instead. The shift is small in wording and large in incentive: charging for access prices the input, while paying for use prices the output.

It is worth noticing what each scheme rewards. A per-crawl toll rewards being expensive to read. A per-answer payment rewards being worth quoting. Those are not the same behaviour, and the second is the one that also happens to be good for readers — which is rare enough in this industry to be worth saying out loud.

4. The crawl-to-referral ratio got a number

The figures reported this week put the exchange in plain arithmetic. Anthropic's crawler takes roughly 11,122 pages for every one visit it sends back. OpenAI is at 848 to 1, Perplexity at 186 to 1, and Google at 5 to 1.

The wider picture from Cloudflare's own network data: as of 3 June 2026, automated requests made up 57.5% of HTML traffic against 42.5% from humans — the first time machines held the majority. As of June, 52% of crawler requests were for training, up from 22% in spring 2025, while search-purpose crawling was under 10% in May.

Those ratios are real and they are worth knowing. They are also the number most likely to be misused this quarter, because they look like a target and are not one.

Nothing you publish changes your crawl-to-referral ratio on its own. It moves when a crawler changes its appetite or an assistant changes how often it links out — neither of which is a thing your content can do. Optimising it directly has exactly one available move, which is to be crawled less. That improves the ratio and costs you the citations.

5. The one number this week that supports a causal claim

A randomised study found that AI Overviews cut publisher clicks by 39.8%.

That sentence is doing something the other figures in this digest cannot do. It came from an experiment with a control group, which means it supports a claim about what would have happened otherwise. "Traffic fell 40% in the most-crawled categories" — also reported this week — is a description of what happened, and cannot separate AI answers from seasonality, from a redesign, or from a market that moved.

Both are useful. They are not interchangeable, and the difference decides which decisions they can carry.

Also this week

Google's August spam update continued rolling out from 18 August, with the usual pattern: movement on thin and derivative pages, and a fortnight of volatility before anything can be read from it.

Regulation kept drifting toward the answer layer. The European Commission's July decision against Google is now the frame for an open discussion about how the same principles apply to AI Overviews and AI Mode — a question with more consequences for search marketing than the fine itself.

Grok 4.6 shipped with a 500,000-token context window aimed at long-running agent work, which continues a direction worth watching: the assistants that matter for brand visibility are increasingly the ones that browse rather than recall.

Conclusions

The week's real story is that targets stopped being decorative. Google Ads now optimises toward the number typed into the box, which means a figure that was safely aspirational is now operative. Anywhere a target was set because it looked reasonable rather than because the business could afford it, that gap is about to become visible in the spend.

Two of this week's headline numbers describe traffic and one measures an effect. The crawl-to-referral ratios and the 57.5% bot share are descriptions: they tell you what is happening, not what would have happened otherwise. The 39.8% click loss came from a randomised study and is the only figure here that can carry a causal claim. Use each for what it can hold.

The ratio is the trap. It is vivid, it is quotable, and it is not under your control. The only reliable way to improve it is to be read less, which is the opposite of what anybody wants. Keep it as context and steer by the measures that move when you publish something worth citing.

And the answer layer is becoming a paid surface. With roughly a quarter of ChatGPT responses carrying a placement and ads now live across 31 European markets, "appearing in the answer" has split into two different achievements with two different costs. Measure them separately, starting now, while the two series are still easy to tell apart.